New AI model releases — November 2026
85 tracked releases from 20 labs so far, 67 we’d call major. Confirmed-only — no rumors; every event’s full entry links to an official source.
AI models released in November 2026 (17)
- 2026-11-13ReleaseMajor
DeepSeek opens built-in tools and the full application API suite on V4.1-Flash, with new per-use tool fees
Separate API-suite event on the model first logged 2026-09-10; that launch entry and its RSS guid stay unchanged. DeepSeek's news post and the changelog dated 2026-11-13 open built-in tools on the existing deepseek-flash endpoint, still served as DeepSeek-V4.1-Flash-1017. @deepseek_ai posted at 2026-11-13T14:24:44Z, 2026-11-13T14:24:44Z, 2026-11-13T14:24:45Z, and 2026-11-13T14:24:46Z. Those posts are cited as timestamps. Their bodies are not used as a source of extra claims.
Show more (2 more paragraphs)
Built-in Web Search, Deep Thinking Search, File Input, URL Reading, and Safety Checks run server-side across Messages, Responses, and Chat Completions. Web Search and Deep Thinking Search use deepseek_web_search. URL reading uses deepseek_read_url. Public-facing apps must enable safety checks. Standalone Search and Fetch, OCR, ASR, and TTS APIs are also available. The news post names a demo at cdn.deepseek.com/chat-demo and code at github.com/deepseek-ai/chat-demo.
The pricing page checked 2026-11-14 keeps peak cache-miss at $0.30/$1.20 and off-peak at $0.15/$0.60 per million tokens. Tool fees are billed on top: search $0.75 per 1,000 searches, web fetch $0.75 per 1,000 fetches, file parsing and OCR $0.225 per 1,000 pages, TTS $0.45 list per 10,000 characters with a limited-time 50% off price of $0.225, and ASR $0.116 list per hour with a limited-time 50% off price of $0.058. No promo end date is published. No score row is added. This is not a new model id and not a token price change. DeepSeek V4 Pro is not edited.
Official source →What we can and can’t confirm
The 2026-09-10 launch entry remains that day's snapshot, and its RSS guid is unchanged. This entry is the application-API day.
The news page breadcrumb reads DeepSeek Full-Suite App API Release 2026/11/13. The changelog entry of the same date is titled Releasing the full suite of application APIs. Both describe capabilities on V4.1-Flash, API id deepseek-flash, across Messages, Responses, and Chat Completions, with tool calls and context assembly handled server-side. Web Search, with thinking off, searches the web or images and answers with citations, enabled by deepseek_web_search. Deep Thinking Search uses the same tool with thinking on, and searches and reads pages over multi-step reasoning. File Input parses PDFs, Office files, images, and more into the conversation. URL Reading is enabled with deepseek_read_url. Safety Checks are required for public-facing apps and return refusal indicators. The changelog says a public-facing service must enable that check and still meet obligations it names: content safety review, large-model registration, and AI labeling. Search and Fetch, OCR, ASR, and TTS are also standalone APIs. The news post names a demo at cdn.deepseek.com/chat-demo and code at github.com/deepseek-ai/chat-demo.
The pricing page checked 2026-11-14 still lists model version DeepSeek-V4.1-Flash-1017. Peak cache-miss input is $0.30 per million tokens and output is $1.20. Off-peak is $0.15 and $0.60. Those are the token rates already stored as price_in, price_out, and price_alt. They are not edited. DeepSeek-V4-Pro on the same page stays at peak $1.32 input and $3.96 output. That record is not edited.
Tool fees are billed in addition to tokens. Search, text or image, is $0.75 per 1,000 searches. Web fetch is $0.75 per 1,000 fetches. File parsing and the OCR API are $0.225 per 1,000 pages. TTS lists $0.45 per 10,000 characters, with a limited-time price of $0.225, which the page labels 50% off. ASR lists $0.116 per hour of audio, with a limited-time price of $0.058, also labeled 50% off. The page publishes no end date for either discount.
This entry does not add a model id for OCR, ASR, TTS, or a DeepSeek App. It does not add a score. It does not mark any model superseded. It does not backfill a 2026-10-17 in-house Search API tracker row. The four @deepseek_ai posts are cited by timestamp only: 2026-11-13T14:24:44Z, 2026-11-13T14:24:44Z, 2026-11-13T14:24:45Z (the pricing-card photo), and 2026-11-13T14:24:46Z. Their bodies are not a source of extra claims.
- DeepSeek news — The full DeepSeek App experience, now via API; sidebar date 2026/11/13 (checked 2026-11-14)
- DeepSeek changelog — Date 2026-11-13, Releasing the full suite of application APIs (checked 2026-11-14)
- DeepSeek pricing — deepseek-flash DeepSeek-V4.1-Flash-1017 peak $0.30/$1.20, off-peak $0.15/$0.60, plus tool fees (checked 2026-11-14)
- @deepseek_ai — post, 2026-11-13T14:24:44Z (timestamp only)
- @deepseek_ai — post, 2026-11-13T14:24:44Z (timestamp only)
- @deepseek_ai — pricing-card photo, 2026-11-13T14:24:45Z (timestamp only)
- @deepseek_ai — post, 2026-11-13T14:24:46Z (timestamp only)
- Demo named in the 2026-11-13 news post
- Demo source named in the 2026-11-13 news post
- 2026-11-12ReleaseMajor
Thinking Machines publishes Inkling 2 Small open weights under Apache 2.0
Separate weights-day event on the line first logged 2026-11-05; the original launch entry and its RSS guid stay unchanged. A blog update dated November 12, 2026 says the weights are available for download on Hugging Face. @thinkymachines posted at 2026-11-12T20:50:38Z. The repository thinkingmachines/Inkling-2-Small carries license apache-2.0 and lastModified 2026-11-12T21:35:28Z. The catalog license field is open-weights. Apache 2.0 is the license name, not that enum. Tinker prices stay $0.30/$1.00, the served window stays 262,144, and the stored counts stay 594B total and 17B active. No new score is recorded. This entry does not say Inkling 2 Small replaces Inkling, Inkling-Small, or Fledge Alpha.
Official source →What we can and can’t confirm
The November 5 entry remains the launch snapshot, including its statement that the weights were still closed, and its RSS guid is unchanged. This entry is the weights day.
The blog update dated November 12, 2026 says Inkling 2 Small's weights are available for download on Hugging Face. It also says the model is integrated with Unsloth, SGLang, vLLM, and TokenSpeed, can be post-trained with Miles, SkyRL, and Snowflake Cortex Training, and is available via API at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare. No price is stated for those hosts, and none is stored. @thinkymachines posted at 2026-11-12T20:50:38Z. @miramurati posted at 2026-11-12T20:55:24Z. Both posts are cited as timestamps. Their bodies are not used as a source of extra claims.
The Hugging Face repository thinkingmachines/Inkling-2-Small lists license apache-2.0 and lastModified 2026-11-12T21:35:28Z. This catalog's license field is open-weights, because that enum is only proprietary or open-weights. Apache 2.0 is stated in the model notes, not stored as the enum value. A display parameter count on the repository card is not the figure stored here. The stored counts remain 594B total and 17B active.
Tinker prices stay $0.30 input and $1.00 output per million tokens. The served context stays 262,144. This entry does not claim Tinker now serves 1M tokens. No tracked score is added. The 2026-10-01 Fledge Alpha tracker entry is left as written. Nothing here says Inkling 2 Small replaces Inkling, Inkling-Small, or Fledge Alpha.
- Thinking Machines blog — Inkling 2 Small, update dated Nov 12, 2026 (weights on Hugging Face; local and API hosts named, no third-party prices)
- @thinkymachines — weights post, 2026-11-12T20:50:38Z
- Hugging Face — thinkingmachines/Inkling-2-Small, license apache-2.0, lastModified 2026-11-12T21:35:28Z
- @miramurati — post, 2026-11-12T20:55:24Z (timestamp only)
- 2026-11-12ReleaseMajor
Google launches Gemini 4 Hydrogen — API id gemini-4-hydrogen, $1.50/$7.50 per 1M tokens, 1,048,576-token context
Google's blog dated November 12, 2026, written by Koray Kavukcuoglu, launches Gemini 4 Hydrogen under the stable API id gemini-4-hydrogen. Announcement posts are cited by timestamp: @sundarpichai at 2026-11-12T18:02:12Z, @GoogleDeepMind at 2026-11-12T18:02:36Z, @GoogleAI at 2026-11-12T18:03:58Z, @GoogleAIStudio at 2026-11-12T18:04:32Z, and @Google at 2026-11-12T18:09:08Z. This entry does not add wording from those posts beyond the timestamps.
The AI Studio model page checked November 13, 2026 states a unified context window of 1,048,576 tokens. Paid Standard is a flat $1.50 input and $7.50 output per million tokens across that window. A Free tier exists and is not stored as the headline. No score row is added. Carbon, Argon, and Gemini 3.8 Flash are not marked superseded.
Official source →What we can and can’t confirm
What is confirmed, from the November 12 blog and the AI Studio pages checked November 13, 2026. The API id is gemini-4-hydrogen. Inputs are text, image, video, audio, and PDF. Output is text. The model page states one unified context window of 1,048,576 tokens. The default synchronous output cap is 65,536 tokens, and long decode can raise output to 1,048,576. Caching is supported. Explicit caching is not. Batch, Flex, Priority, and Thinking are supported. Image generation, audio generation, and the Live API are not. Knowledge cutoff, parameter count, and architecture are unpublished. The license is recorded as proprietary.
Price. Paid Standard is $1.50 input and $7.50 output per million tokens across the full window, including thinking tokens, with implicit cache at $0.075. The Hydrogen pricing table has no introductory period and no December 1 step-up. Batch and Flex are $0.75/$3.75 with cache $0.0375. Priority is $2.70/$13.50 with cache $0.135. Those cache rates are implicit only. A Free tier lists input, output, and context caching as free of charge. That free figure is not the paid headline, and it is not stored as $0. Grounding is not available on Free, and free-tier content may be used to improve products. Paid content may not. Paid Google Search grounding is 5,000 requests a month free, shared across Gemini 3 and newer, then $14 per 1,000, with a daily cap of 1,500. Maps uses the same 5,000 free requests, then $14 per 1,000. Every token is billed, including thinking tokens and, when grounding is on, internal tool instructions and function declarations.
Limits. The public rate-limits page has no RPM, TPM, or RPD table for Hydrogen. Interactive limits are viewed in AI Studio. Published Batch enqueued-token caps are 3,000,000 on Tier 1, 400,000,000 on Tier 2, and 1,000,000,000 on Tier 3.
What is not confirmed. No independent score is stored. Blog figures, including an Artificial Analysis Index, a speed claim versus Gemini 3.8 Flash, and named evals, stay vendor claims and are not score rows. Thinking levels described on the blog are low, medium, high (the default), and xhigh. Those levels are a vendor description. The five X posts are cited as announcement timestamps. Their bodies were not used as a source of extra claims. Carbon and Argon remain the frontier pair. Hydrogen is a Gemini 4 model for speed and everyday coding and agentic work. Migration language about Gemini 3.8 Flash is not a deprecation, and this entry does not mark Carbon, Argon, or Gemini 3.8 Flash superseded.
- Google blog — Introducing Gemini 4 Hydrogen, Nov 12, 2026, Koray Kavukcuoglu
- @sundarpichai — announcement post, 2026-11-12T18:02:12Z
- @GoogleDeepMind — announcement post, 2026-11-12T18:02:36Z
- @GoogleAI — announcement post, 2026-11-12T18:03:58Z
- @GoogleAIStudio — announcement post, 2026-11-12T18:04:32Z
- @Google — announcement post, 2026-11-12T18:09:08Z
- AI Studio model docs — gemini-4-hydrogen: text/image/video/audio/PDF in, text out, unified context 1,048,576, output 65,536 default / 1,048,576 long decode, caching without explicit caching, Batch/Flex/Priority/Thinking supported (checked 2026-11-13)
- AI Studio pricing — paid Standard $1.50/$7.50 flat, cache $0.075, Batch and Flex $0.75/$3.75, Priority $2.70/$13.50, Free tier exists, no intro period and no Dec 1 step-up (checked 2026-11-13)
- AI Studio rate limits — no RPM/TPM/RPD table for Hydrogen; Batch enqueued tokens Tier 1 3,000,000, Tier 2 400,000,000, Tier 3 1,000,000,000 (checked 2026-11-13)
- 2026-11-11ReleaseMajor
Gemini Decision 1 launches — API id gemini-decision-1, $0.10 input / 0 output tokens, 262,144 context, Decisions API, Free tier exists
Google's blog dated November 11, 2026 introduces Gemini Decision 1, API id gemini-decision-1, as a Decisions API model in Google AI Studio. Announcement posts are cited by timestamp: @sundarpichai at 2026-11-11T18:15:20Z, @GoogleAIStudio at 2026-11-11T18:15:51Z, and @Google at 2026-11-11T18:26:30Z. This entry does not add wording from those posts beyond the timestamps.
Show more (2 more paragraphs)
The model page checked November 12, 2026 states a 262,144-token input limit. The blog rounds that figure as up to 256K. The stored window is 262,144, not 256,000. Paid Standard is $0.10 per million input tokens. Output is not applicable because the model emits 0 output tokens, stored here as 0. A Free tier exists and lists input as free of charge. That free input price is not the paid headline. Free-tier content may be used to improve products. Paid content may not.
The model does not generate text. It returns calibrated probabilities for yes/no, choice, and rubric questions. Inputs are text, image, video, audio, and PDF. Up to 1,000 questions can share one input, and that shared input is billed once. Caching and the Batch API are not supported. The documented API version is v1beta. The blog says Google Cloud support is coming soon. This entry does not call Vertex or Cloud generally available. No score row is added. The blog's suite, calibration, nearly-5x, and latency claims stay off the tracked boards. Argon and Carbon are not marked superseded. The family is Gemini Decision, not Gemini 4. This is not a GPT-6 Luna equivalent.
Official source →What we can and can’t confirm
What is confirmed, from the November 11 blog and the AI Studio pages checked November 12, 2026. The API id is gemini-decision-1. The model is a specialized Decisions API model. It does not generate text. Inputs are text, image, video, audio, and PDF. Output is calibrated probabilities, and the output token limit is 0. One request can carry up to 1,000 questions. The quickstart uses API version v1beta in Google AI Studio. The blog says support is coming soon to Gemini in Google Cloud. That is not recorded here as Vertex or Cloud general availability.
Price. Paid Standard is $0.10 per million input tokens for text, image, video, audio, and PDF. The pricing page marks output as not applicable because the model emits 0 output tokens, on both Free and Paid, so the stored output rate is 0 rather than a made-up generative price. The Free tier lists input as free of charge. That $0 input figure is not the paid headline. Gemini 4 Carbon's pricing row, checked November 11, 2026, had no Free tier. Free-tier content may be used to improve Google's products. Paid content may not. Shared input is billed once, and each question adds tokens for instructions, criteria, and formatting. Context caching is not available. The Batch API is not supported, so there is no second paid rate.
Window. The model page states an input token limit of 262,144. The blog's up to 256K is a rounded figure and is not stored. Free rate limits on the rate-limits page are 200 requests per minute, 50,000 tokens per minute, and 10,000 requests per day. Paid Tier 1 is 1,500 / 1,000,000 / 50,000. Tier 2 is 5,000 / 4,000,000 / unlimited. Tier 3 is 15,000 / 10,000,000 / unlimited. The model page marks priority inference as not supported, so the general priority quota section is not applied to this model.
What is not confirmed. Knowledge cutoff, parameter count, architecture, and a license statement are unpublished. The model page records the license as proprietary because the official page does not state one. No independent score is stored. The blog's claim that Decision 1 leads a suite spanning agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error, and that it is nearly 5x cheaper on that shared suite than Gemini 3.5 Flash-Lite with the Interactions API, stays a vendor claim. So does the latency line: 50 to 100 ms of model compute and 150 to 250 ms end to end, scoped to multimodal inputs and contexts up to 16K tokens, nearly flat from 1 to 100 questions. The three X posts are cited as announcement timestamps. Their bodies were not used as a source of extra claims. Argon and Carbon are not marked superseded. The family is Gemini Decision, so the previous-version rule does not attach this model to Gemini 4. GPT-6 Luna is a generative text model at $0.10 input and $0.50 output. Decision 1 is not that product.
- Google blog — Gemini Decisions API, Nov 11, 2026, Min Choi and Felix Yu (Google DeepMind)
- @sundarpichai — announcement post, 2026-11-11T18:15:20Z
- @GoogleAIStudio — announcement post, 2026-11-11T18:15:51Z
- @Google — announcement post, 2026-11-11T18:26:30Z
- AI Studio model docs — gemini-decision-1: id, input 262,144, output limit 0, up to 1,000 questions, text/image/video/audio/PDF in, calibrated probabilities out, text generation/caching/Batch unsupported (checked 2026-11-12)
- AI Studio pricing — paid $0.10/1M input, output not applicable (0 output tokens), caching not available, Free input free of charge, free-tier data used to improve products, paid not, shared input billed once, Batch unsupported (checked 2026-11-12)
- AI Studio rate limits — Free 200 RPM / 50,000 TPM / 10,000 RPD; Tier 1 1,500/1,000,000/50,000; Tier 2 5,000/4,000,000/unlimited; Tier 3 15,000/10,000,000/unlimited; Batch not supported (checked 2026-11-12)
- 2026-11-11Release
Qoduck attributed to Qwen4-Max Preview — Qoder says its stealth model was a four-day Qwen test build
Qoder's post at status 2120343731673108936 is the attribution. Qwen's own 2026-11-11 post does not use the name Qoduck. Qoder says Qoduck was a Qwen test build that ran anonymously for four days and was updated each day, and that it now upgrades to Qwen4-Max-Preview. Qoder also says the upgraded model is live first on Qoder for every product and plan, including Teams and Enterprise. That access line is wider than the 2026-11-07 entry, which recorded individual Qoder International plans only.
The limited 80% discount at 0.1x credits is recorded as a promotion only. The 2026-11-07 appearance stays the record of what was known that day. This is a new event, not a rewrite of that row. No model page and no score row are created from the attribution alone.
Official source →What we can and can’t confirm
The identity claim comes from Qoder, not from an independent weight comparison. Qoder's post says the stealth model customers saw as Qoduck was a Qwen test build for four anonymous days, with daily updates, and that the offering upgrades to Qwen4-Max-Preview. @Alibaba_Qwen's preview post the same morning does not mention Qoduck, so this site does not treat the name as something Qwen stated. Nothing here claims the four-day checkpoint was always called Qwen4-Max, or that its weights match the preview id byte for byte. That stronger claim is not in either post.
Access changed relative to the 2026-11-07 tracker row. That row, left as written, described individual Qoder International plans and said Teams and Enterprise did not see Qoduck. Qoder's new post says the upgrade is live first on Qoder across all products and plans, Teams and Enterprise included. The 80% off at 0.1x credits stays a promotion, not a token price. The 2026-11-07 Qoder promotion page remains the source for what the launch-day terms were. No models.json page is opened from this attribution.
- @qoder_ai_ide — Qoduck attribution, four anonymous days then an upgrade to Qwen4-Max-Preview, live first on Qoder including Teams and Enterprise
- Qoder — 2026-11-07 Qoduck promotion page (individual Qoder International plans; left as that day's record)
- @Alibaba_Qwen — Qwen4-Max preview post, which does not name Qoduck, 2026-11-11T02:01:09Z
- 2026-11-11ReleaseMajor
Alibaba previews Qwen4-Max — Token Plan only, 1M context, retired after the 2026-11-25 preview
The official @Alibaba_Qwen post is timestamped 2026-11-11T02:01:09Z. It previews the model id qwen4-max-preview, and the docs page calls the same id a preview. Pay-as-you-go is not supported. Access is Token Plan only. The preview window runs from 10:00 on 2026-11-11 to 24:00 on 2026-11-25, UTC+8, with no changes during that period, and the model is retired after the window ends.
Show more (2 more paragraphs)
The docs state a 1M-token context and a 393,216-token maximum output that thinking and the final answer share. Inputs are text, images, and video. Output is text. Thinking is on by default. The API is compatible with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. The Token Plan base URL is https://token-plan.maas.qwencloudapi.com.
Qoder's post the same day says its stealth model Qoduck was a Qwen test build served anonymously for four days, updated daily, and that the product now upgrades to Qwen4-Max-Preview, live first on Qoder across every product and plan, including Teams and Enterprise. Qoder's limited 80% off, billed at 0.1x credits, is a promotion and is not a list price. This entry does not record a per-token rate. No Qwen4-Flash, Qwen4-Plus, or 27B sibling is announced here, no weights, model card, or Hugging Face repo is linked, and no independent tracked score exists. The license stays proprietary until weights exist.
Official source →What we can and can’t confirm
This is a preview, not a general launch. @Alibaba_Qwen posted at 2026-11-11T02:01:09Z, and the Qwen Cloud getting-started page documents id qwen4-max-preview. That page says pay-as-you-go is unsupported and that the only route is Token Plan, from 10:00 on 2026-11-11 through 24:00 on 2026-11-25 UTC+8. The docs say the preview will not change during that span and that the model is retired when it ends. No per-token price is published on the docs or on the Token Plan page, so this record invents none.
The stated context is 1 million tokens. Maximum output is 393,216 tokens, shared by the thinking trace and the final answer. Input can be text, images, or video. Output is text. Thinking defaults to on. Compatibility covers OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, with base URL https://token-plan.maas.qwencloudapi.com.
Qoder's same-day post is the attribution for the earlier stealth name: four anonymous days, a daily update, then an upgrade to Qwen4-Max-Preview, offered first on Qoder for all products and plans including Teams and Enterprise. The 80% off / 0.1x credits line is stored only as that promotion. This entry does not assert that the four-day build was byte-identical to qwen4-max-preview. It does not claim Qwen4-Flash, Qwen4-Plus, or a 27B model shipped, and it adds no weights link and no independent score.
- @Alibaba_Qwen — Qwen4-Max preview, 2026-11-11T02:01:09Z
- Qwen Cloud docs — qwen4-max-preview: Token Plan only, no pay-as-you-go, preview 10:00 2026-11-11 to 24:00 2026-11-25 UTC+8, retired after, 1M context, 393,216 max output (checked 2026-11-11)
- Qwen Cloud Token Plan page — no per-token list price stored from this page
- @qoder_ai_ide — Qoduck was a four-day Qwen test build, now upgrading to Qwen4-Max-Preview (attribution, not an independent weight check)
- 2026-11-10ReleaseMajor
Gemini 4 Carbon is now publicly available — API id gemini-4-carbon, not on the Free tier, intro $2/$10 through 2026-11-30
Google's 2026-11-10 availability blog and the @Google post at 2026-11-10T16:05:52Z say Gemini 4 Carbon is available now, on a public channel rather than Fairwind alone. The API id published for that channel is gemini-4-carbon. Surfaces named in the blog are the Gemini app, AI Mode in Search, Antigravity, AI Studio, the Gemini API, Android Studio, and Gemini for business, with Ultra for consumers and Google Cloud or AI Studio for developers and enterprises. The pricing page checked 2026-11-11 does not put gemini-4-carbon on the Free tier.
Show more (2 more paragraphs)
Standard list rates for this public id are $2.00 input and $10.00 output per million tokens through 2026-11-30, then $4.00 and $20.00 from 2026-12-01. Model docs checked the same day state a vendor context of 1,048,576 unified tokens, a default output cap of 65,536, and long decode up to 1,048,576. Accepted inputs are text, image, video, audio, and PDF, and the output modality is text. Thinking is supported on gemini-4-carbon, and explicit caching is not. Fairwind stays the separate Carbon-for-cyber program, while this public id is the other availability path.
Argon remains for sale after this opening. Google still points most software-engineering and knowledge work at Argon, and the hardest long-horizon tasks at Carbon, and this entry does not mark Argon superseded. Vendor Carbon benchmark claims are not recorded as independent tracked scores, and no score row is added for the 2026-11-10 opening. The prose-only Artificial Analysis Index v5.0 reading of 59 at high stays the 2026-10-29 check. This scan adds no new Artificial Analysis measurement.
Official source →What we can and can’t confirm
Availability. On 2026-11-10 Google posted that Gemini 4 Carbon is publicly available. @Google's post is timestamped 2026-11-10T16:05:52Z, with a reply at status 2120186190905090406 and a Sundar Pichai post at status 2120185896469496180. The AI Studio model page, checked 2026-11-11, documents the id gemini-4-carbon. The blog lists the Gemini app, AI Mode in Search, Antigravity, AI Studio, the Gemini API, Android Studio, and Gemini for business. Consumers are on Ultra. Developers and enterprises go through Google Cloud and AI Studio. The pricing page checked that day has no Free-tier price for this id. Fairwind is still the separate route described as Carbon for cyber. That program has not ended, and it is not this public API.
Price. Standard on the pricing page is $2.00 per million input tokens and $10.00 per million output tokens through 2026-11-30, then $4.00 and $20.00 from 2026-12-01. Implicit cache is $0.10 during the intro window and $0.20 after it. Explicit caching is unsupported. Batch and Flex are half of Standard. Priority is $3.60 input and $18 output during the intro, then $7.20 and $36. Thinking tokens bill as output. A footnote on the 2026-11-10 blog says the post-intro $4/$20 rate applies to public-access Carbon. The intro rate is dated through 2026-11-30. It is not described as permanent.
Specs. The model page states a unified context of 1,048,576 tokens. Default maximum output is 65,536 tokens. Long decode allows up to 1,048,576 output tokens. Input modalities are text, image, video, audio, and PDF. Output modality is text. Thinking is supported. Image generation and the Live API are not listed. Batch, Flex, and Priority are listed. Knowledge cutoff, parameter count, and architecture are undisclosed.
Measurement. No HLE, GPQA, Terminal-Bench, DeepSWE, or other tracked score row is added. Vendor cybersecurity claims are not recorded as independent scores. Artificial Analysis's Index v5.0 of 59 at high stays the prose figure from the check on 2026-10-29. This scan did not re-fetch Artificial Analysis, so it does not claim a fresh AA measurement.
Lineup. Argon remains Google's recommendation for most software engineering and knowledge work. Carbon is positioned for the hardest and longest-horizon tasks. This record does not say Carbon replaces Argon, and it does not mark Argon superseded.
- Google — Gemini 4 Carbon now available, 2026-11-10 (public surfaces, Fairwind remains Carbon for cyber, post-intro $4/$20 footnote for public-access Carbon)
- @Google — Carbon is available, 2026-11-10T16:05:52Z
- @Google reply — status 2120186190905090406
- Sundar Pichai — status 2120185896469496180
- AI Studio model docs — gemini-4-carbon: id, unified context 1,048,576, output 65,536 default / up to 1,048,576 long decode, text/image/video/audio/PDF in, text out (checked 2026-11-11)
- AI Studio pricing — gemini-4-carbon: not Free tier; Standard $2/$10 through 2026-11-30, $4/$20 from 2026-12-01; cache $0.10 then $0.20 implicit only; Batch/Flex half; Priority $3.60/$18 then $7.20/$36; thinking billed as output (checked 2026-11-11)
- 2026-11-09ReleaseMajor
Meituan publishes LongCat-2.5 open weights under MIT — public, ungated Hugging Face repo
Separate weights-day event on the LongCat-2.5 line first logged on 2026-11-02; that GA/API entry stays unchanged. The @Meituan_LongCat post at 2026-11-09T14:05:36Z links the public Hugging Face repo meituan-longcat/LongCat-2.5. Checked 2026-11-10, the repo is public and ungated, license mit, lastModified 2026-11-09T11:38:03Z, with 208 numbered safetensors shards plus two multimodal safetensors files and an MIT LICENSE file copyright Meituan 2026. This adds no model page: models.json has no LongCat entry, so model_id stays null. No new independent score is recorded, and the November 2 API price is unchanged.
Official source →What we can and can’t confirm
The @Meituan_LongCat post is timestamped 2026-11-09T14:05:36Z and links the Hugging Face repo. The repo meituan-longcat/LongCat-2.5 is public and ungated, with license mit, lastModified 2026-11-09T11:38:03Z, 208 numbered safetensors shards plus two multimodal safetensors files, and an MIT LICENSE file copyright Meituan 2026 (checked 2026-11-10). This entry uses the announcement date, 2026-11-09, not the repo's lastModified timestamp.
This is a separate weights-day event. The 2026-11-02 LongCat-2.5 GA/API entry, which said weights were not public, and its RSS guid stay unchanged. models.json has no LongCat entry, so this adds no model page, no model_id and no score row. The existing $0.30/$1.20 API price on the November 2 entry is unaffected; the weights release publishes no new price.
- Meituan LongCat on X — LongCat-2.5 open weights, MIT, 2026-11-09T14:05:36Z (checked 2026-11-10)
- Hugging Face — meituan-longcat/LongCat-2.5: public ungated, license mit, lastModified 2026-11-09T11:38:03Z, 208 numbered shards plus two multimodal safetensors (checked 2026-11-10)
- Hugging Face — LongCat-2.5 MIT LICENSE, copyright Meituan 2026 (checked 2026-11-10)
- 2026-11-09ReleaseMajor
Meta launches Muse Ember — $0.10/$0.50 short-context, 1,048,576-token context, long-context tier above 200K
Meta released Muse Ember on 2026-11-09. The blog is dated November 9, 2026; @AIatMeta posted at 2026-11-09T18:04:15Z and followed with pricing in thread post 2119853597542568405. The Standard id is `muse-ember-1.0`: input is text, image, video, audio and PDF; output is text; context is 1,048,576 tokens; maximum output is 524,288 tokens; reasoning runs from minimal to max with max as the default. Pricing is $0.10 input / $0.004 cached / $0.50 output per million tokens up to 200K input tokens, and $0.30 / $0.012 / $1.50 above 200K, with the whole request billed at the long-context tier. The $0.10/$0.50 short-context rate is the stored headline; the long-context tier is higher, so it is not stored as price_alt. No tracked independent Muse Ember score existed at the 2026-11-10 scan, so no score row is added. Family is Muse Ember, so Muse Wonder is not marked a previous version.
Official source →What we can and can’t confirm
Meta's blog is dated November 9, 2026; @AIatMeta posted at 2026-11-09T18:04:15Z and followed with pricing in thread post 2119853597542568405. The Standard id is `muse-ember-1.0`; input is text, image, video, audio and PDF, output is text, context is 1,048,576 tokens, maximum output is 524,288 tokens, and reasoning runs from minimal to max with max as the default (dev.meta.ai/docs/muse-ember, checked 2026-11-10).
Pricing is length-tiered: up to 200K input tokens, $0.10 input / $0.004 cached / $0.50 output per million tokens; above 200K, $0.30 / $0.012 / $1.50, with the whole request billed at the long-context tier (dev.meta.ai/docs/pricing-rate-limits, checked 2026-11-10). The $0.10/$0.50 short-context rate is the stored headline; the long-context tier is higher, so it is not stored as price_alt. Meta publishes no parameter count or knowledge cutoff and links no weights, so those fields stay empty and the license is proprietary. No tracked independent Muse Ember score existed at the scan, so no score row is added. The launch table's Agents' Last Exam 61.6% is an internal Linux subset, Terminal-Bench 4.0 at 40.9% is a best-published-harness figure, and OSWorld 2.1 is an internal 69.5% partial and 33.8% binary. None of those is a score row. Family is Muse Ember, so Muse Wonder is not marked a previous version.
- Meta AI Research — Muse Ember, dated November 9, 2026 (checked 2026-11-10)
- Meta AI on X — Muse Ember launch, 2026-11-09T18:04:15Z (checked 2026-11-10)
- Meta AI on X — pricing thread, post 2119853597542568405 (checked 2026-11-10)
- Meta developer docs — Muse Ember: Standard id muse-ember-1.0, text/image/video/audio/PDF input, text output, 1,048,576 context, 524,288 max output, reasoning minimal to max, default max (checked 2026-11-10)
- Meta developer docs — pricing and rate limits: up to 200K $0.10/$0.004/$0.50; above 200K $0.30/$0.012/$1.50, whole request at long-context tier (checked 2026-11-10)
- 2026-11-09ReleaseMajor
Claude Fable 5.2 launches — $8/$40, down from $10/$50, 1M context, Mythos 5.2 verification-only
Anthropic released Claude Fable 5.2 on 2026-11-09. The blog is dated November 9, 2026; the @AnthropicAI post is timestamped 2026-11-09T19:00:21Z, and the quoted @claudeai launch post is 2026-11-09T18:57:59Z. The documented model id is `claude-fable-5-2`, with a 1,000,000-token context, 128,000 maximum output tokens, $8 input / $40 output per million tokens, and a June 2026 knowledge cutoff. Pricing falls from Fable 5.1's $10/$50; cache reads are $0.20 per million tokens and fast mode is $16/$80. Claude Mythos 5.2 (`claude-mythos-5-2`) has the same specs and pricing but is limited to the Life Sciences and Cyber Verification Programs, recorded in the Fable 5.2 notes rather than as a separate model page. No tracked independent Fable 5.2 score existed at the 2026-11-10 scan, so no score row is added. The catalog marks Fable 5.1 as the previous version because Fable 5.2 is newer in the same Claude Fable line and the same name-tier; that rule reads the name, not the dollar price. No lifecycle field is stored.
Official source →What we can and can’t confirm
Anthropic's blog is dated November 9, 2026; the @AnthropicAI post is timestamped 2026-11-09T19:00:21Z and quotes the @claudeai launch post at 2026-11-09T18:57:59Z. The model docs publish the id `claude-fable-5-2` with a 1,000,000-token context, 128,000 maximum output tokens, $8 input / $40 output per million tokens, and a June 2026 knowledge cutoff (checked 2026-11-10).
Anthropic's pricing table confirms $8/$40 for Fable and Mythos 5.2, down from $10/$50 for 5.1, with cache reads at $0.20 per million tokens and fast mode at $16/$80. Claude 4.6 and later models, except Haiku 5.5, include the full 1M window at the standard rate. `claude-mythos-5-2` has the same specs and pricing but is limited to the Life Sciences and Cyber Verification Programs, so it stays in the Fable 5.2 notes rather than becoming a separate model page. The catalog marks Fable 5.1 as the previous version because the lineage rule matches the Claude Fable line and the name, not the sticker price. No tracked benchmark board had a Fable 5.2 row at the 2026-11-10 scan, so no score row is added; the launch table's tracked-board cells stay claims in the model notes.
- Anthropic — Claude Fable 5.2 and Claude Mythos 5.2, dated November 9, 2026 (checked 2026-11-10)
- Anthropic on X — launch post, 2026-11-09T19:00:21Z (checked 2026-11-10)
- Claude on X — quoted launch post, 2026-11-09T18:57:59Z (checked 2026-11-10)
- Anthropic docs — Claude Fable 5.2 overview: id, 1M context, 128K max output, $8/$40, November 9 2026, June 2026 cutoff (checked 2026-11-10)
- Anthropic docs — Claude Mythos 5.2 overview: same specs and pricing, verification-program only (checked 2026-11-10)
- Anthropic pricing — Fable/Mythos 5.2 $8/$40, down from 5.1 $10/$50; cache read $0.20; fast mode $16/$80 (checked 2026-11-10)
- 2026-11-07ReleaseMajor
Anonymous "Qoduck" appears on Qoder — stealth model, origin undisclosed, free daily allowance for individual Qoder International users
Qoder, Alibaba's coding product, announced a stealth model called Qoduck on 2026-11-07; Qoder's own account posted at 2026-11-07T02:11:24Z and followed up at 2026-11-07T13:59:04Z, and Alibaba Cloud's account posted the same announcement at 2026-11-07T09:00:47Z. Qoder's promotion page, checked 2026-11-08, says Qoduck is a stealth model whose identity has yet to be disclosed, with more information to be shared later.
Show more (2 more paragraphs)
The page dates the promotion from 2026-11-07 (Singapore Time, UTC+8) and publishes no end date. It is limited to individual users on Qoder International — Free, Pro Trial, Lite, Pro, Pro+, and Ultra — and is not shown to Teams or Enterprise. Within a daily free allowance, the credit rate drops from 0.5× to 0.0× even at the highest thinking intensity; standard 0.5× credits apply after the allowance. The Qoder IDE, JetBrains plugin, and Qoder CLI do not support the promotion.
No context window, modalities, API id, parameter count, licence, weights or benchmark result is published in the material checked. This is recorded as an appearance, not an attributed lab release: model_id stays null and no model page or score row is added. Community identity guesses are not evidence and are not recorded.
Official source →What we can and can’t confirm
Qoduck is a stealth model offered inside Qoder, an Alibaba coding product, not a lab release this site can attribute. The launch record is official and unusually clear about access: Qoder's own X account announced it at 2026-11-07T02:11:24Z and followed up at 2026-11-07T13:59:04Z, and Alibaba Cloud's account posted the same announcement at 2026-11-07T09:00:47Z. Qoder's promotion page, checked 2026-11-08, is the primary source for the terms below. It calls Qoduck a stealth model whose identity has yet to be disclosed and says more information will be shared later.
The promotion starts 2026-11-07, with times in Singapore Time (UTC+8); no end date is published, and Qoder says one will be announced in advance. It is for individual users of Qoder International on Free, Pro Trial, Lite, Pro, Pro+, and Ultra. Teams and Enterprise do not see Qoduck. Within a daily free allowance, the credit rate falls from 0.5× to 0.0× even at the highest thinking intensity; after the allowance, standard 0.5× credits apply. The Qoder IDE, JetBrains plugin, and Qoder CLI do not support the promotion.
What is not published is most of what a model page here would need: no context window, input or output modalities, API id, parameter count, licence, weights, named maker, or benchmark result. Those absences are why this is a tracker appearance rather than a model page; the site's model pages require a lab and licence as facts and scores with a named runner. The daily allowance is not a published $0 price for the model and this entry does not record one.
The honest summary as of 2026-11-08 is an officially announced stealth model inside Qoder with an undisclosed origin and an unpublished specification sheet. The Qoder docs and official posts confirm the appearance and the access terms, not the model's identity. Two things would change this entry: Qoder naming the model and its maker, or a tracked benchmark board publishing a score for it.
- Qoder docs — Qoduck free promotion: stealth model, identity undisclosed, start 2026-11-07 SGT, daily free allowance, Qoder International individual users, no end date (checked 2026-11-08)
- Qoder on X — first Qoduck announcement: stealth model, free every day for individual users, limited time, 2026-11-07T02:11:24Z (checked 2026-11-08)
- Qoder on X — follow-up: new stealth model, origin undisclosed, free daily for a limited time, 2026-11-07T13:59:04Z (checked 2026-11-08)
- Alibaba Cloud on X — Qoduck announcement, 2026-11-07T09:00:47Z (checked 2026-11-08)
- 2026-11-05ReleaseMajor
Ant Ling launches Ling-3.2-flash — native multimodal 560B/25B, 1M context, free API for two weeks
Major release, separate from the 2026-09-29 Ling 3.1 Flash launch and the 2026-10-27 Ling-3.1-flash-1019 weights day. The changelog dated November 5, 2026 names Ling-3.2-flash: 560B total, 25B activated, the same footprint the model docs list for Ling-3.1-flash, with native image, text, and video input. The docs state a context window of up to 1M tokens. That figure is vendor-stated. This row does not store a served-window measurement, and it does not copy Ling 3.1's 262K caveat onto 3.2. Chat, plus an API the docs say is compatible with OpenAI and Anthropic protocols. Free for the first two weeks after launch. The price page does not publish a post-promo rate for 3.2, and no weights release is announced. No machine model id was on the pages checked 2026-11-07.
Official source →What we can and can’t confirm
Ant Ling's changelog entry is dated November 5, 2026. The @AntLingAGI post is timestamped 2026-11-05T19:40:27Z. Both name Ling-3.2-flash. This is a new generation, not an edit of the Ling 3.1 Flash row or of the Ling-3.1-flash-1019 weights row.
The model docs list 560B total parameters and 25B activated, and they list the same two figures for Ling-3.1-flash. The stated difference is the modality. Ling-3.2-flash is described as native multimodal, with image, text, and video input. Ling-3.1-flash is described there as a hybrid-reasoning text model. The changelog repeats the 560B/25B/1M specification and the image, text, and video input.
Context is vendor-stated: the changelog says a 1M context window, and the model docs say up to 1M, capped at 1M. No independent measurement of the served window is recorded here. Ling 3.1's note that a route served less than the spec sheet is not applied to 3.2, because that check was not repeated for this model.
Distribution, per the changelog, is a chat interface and an API compatible with OpenAI and Anthropic protocols. The pages checked on 2026-11-07 do not publish a request model string, so this row has no model_id and no model page. Calling the product Ling-3.2-flash in the docs is not the same as confirming an API id.
Price: the changelog and the price page both say Ling-3.2-flash is free for the first two weeks after launch. The price page says original prices resume after the promotional period and does not date that end. Its numeric tables list Ling-3.1-flash and other models, not Ling-3.2-flash. No post-promo dollar or yuan rate is recorded.
No open-weights release is announced for 3.2. The MIT weights published on 2026-10-27 remain Ling-3.1-flash-1019. The model docs' series history says Ling has been open-sourced in earlier generations; that sentence is not a 3.2 weights release. Nothing here marks Ling 3.1 Flash superseded.
The X post claims Ant Ling's own release evaluations outscore Opus 5 (max) on OmniDocBench 1.5 (92.21 versus 90.56), ZEROBench (Sub) (43.71 versus 39.52), and Vision2Web (85.86 versus 80.20). Those are vendor claims. None of those boards is tracked here, and none of those figures is stored as a score.
- Ant Ling changelog — November 5, 2026: Ling-3.2-flash, 560B/25B, 1M, image/text/video input, free for the first 2 weeks (checked 2026-11-07)
- Ant Ling model docs — Ling-3.2-flash native multimodal, same 560B/25B footprint as Ling-3.1-flash, context up to 1M (checked 2026-11-07)
- Ant Ling price docs — Ling-3.2-flash free for the first 2 weeks; numeric tables do not list a 3.2 rate (checked 2026-11-07)
- Ant Ling on X — Introducing Ling-3.2-flash, 2026-11-05T19:40:27Z, vendor eval claims versus Opus 5 (max)
- 2026-11-05ReleaseMajor
Thinking Machines releases Inkling 2 Small — 594B/17B MoE on Tinker, weights closed, Apache 2.0 promised
Major release: Thinking Machines Lab put Inkling 2 Small on Tinker the day of the announcement, with weights still closed. It is a 594B/17B mixture-of-experts model, text and image in, text out, no audio. Tinker serves 64K and 256K at $0.30 input and $1.00 output per million tokens. The model card's claim of up to 1M tokens is not a served window. Apache 2.0 is promised after external safety testing, with no date. No tracked score is stored. Artificial Analysis displays Intelligence Index 31 on v5.0.1. The launch table labels AA Index (v5.0) at 30.6. Neither figure is stored, and neither is compared with any pre-v5 absolute.
Official source →What we can and can’t confirm
The launch post is dated Nov 5, 2026. The @thinkymachines post is timestamped 2026-11-05T23:43:57Z, and Mira Murati's post the same evening says the model is on Tinker today with weights coming soon. Neither post states a calendar date for the weights.
The model card lists 594B total and 17B active parameters, a 48-layer decoder-only transformer, and a sparse mixture-of-experts in which each token is routed to 8 of 384 experts. Input is text and image. Output is text. The launch post says this model does not support audio, and it describes a 292M-parameter vision encoder. The card's training-data sentence includes video. That is training data, not an input. Tinker's catalog labels Inkling 2 Small Hybrid + Vision. Inkling and Inkling-Small remain listed as Hybrid + Audio + Vision.
The card says the context window is up to 1M tokens. Tinker does not serve 1M. models.json lists thinkingmachines/Inkling-2-Small at 64K and thinkingmachines/Inkling-2-Small:peft:262144 at 256K. Serverless inference is thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4, also 256K. The model page stores 262,144. The launch table's "1M-token maximum trajectory limit" is an evaluation cap, raised from 256K for Inkling Small, not a serving claim.
Prefill is $0.30 and sampling is $1.00 per million tokens on both served tiers, cached prefill $0.03, checked 2026-11-07. Training is $1.50 at 64K and $3.00 at 256K. The pricing page says extended context costs more for training only. No long-context inference multiplier is published. The page slug inkling-2-small is not any of those Tinker ids.
Weights are closed. The launch post says full weights under Apache 2.0 are intended next week, after external safety testing and red-teaming. The model card's license line is Tinker Terms of Use. Artificial Analysis lists licenseName Apache 2.0, isOpenWeights false, and an empty weights URL. This tracker follows the launch post: proprietary until a repository exists.
No score row is added. Artificial Analysis's model page, checked 2026-11-07, displays Intelligence Index 31. The embedded intelligenceIndex is 30.6286781498958, and the page describes Intelligence Index v5.0.1. The launch table labels AA Index (v5.0) at 30.6. The blog lede and the X post say 31. v5 absolutes are not compared here with any pre-v5 reading, and no predecessor index is stored. Other cells on the launch table, including Terminal-Bench 2.1 at 85.1% on a Pi harness and 81.3% under Terminus, stay vendor claims.
The launch post says early checkpoints were tested on OpenCode under the name Fledge-Alpha. The 2026-10-01 Fledge Alpha tracker entry is unchanged. Nothing in the launch post, the model card, or the Tinker catalog says Inkling 2 Small replaces Inkling or Inkling-Small.
- Thinking Machines — Introducing Inkling 2 Small, dated Nov 5, 2026 (checked 2026-11-07)
- Thinking Machines — Inkling 2 Small model card: Nov 5, 2026, Tinker Terms of Use, up to 1M claimed, 8 of 384 experts (checked 2026-11-07)
- Tinker — Models and pricing: extended context costs more for training only (checked 2026-11-07)
- Tinker models.json — Inkling 2 Small 64K $0.30/$1.00 train $1.50, and 256K id peft:262144 train $3.00 (checked 2026-11-07)
- Tinker serverless.json — Inkling 2 Small sampler at 256K, input $0.30, cached $0.03, output $1.00 (checked 2026-11-07)
- Artificial Analysis — Inkling 2 Small: displayed Intelligence Index 31, embedded 30.6286781498958, Index v5.0.1, contextWindowTokens 1024000, isOpenWeights false (checked 2026-11-07)
- Thinking Machines on X — Introducing Inkling 2 Small, 2026-11-05T23:43:57Z
- Mira Murati on X — available today on Tinker, weights coming soon, 2026-11-05T23:45:03Z
- 2026-11-05ReleaseMajor
OpenAI announces GPT-6 Cyber — Daybreak Red only, API id gpt-6-cyber, $20/$100 short-context
OpenAI announced GPT-6 Cyber on 2026-11-05. The @OpenAI post is timestamped 2026-11-05T17:09:13Z. OpenAI describes it as a Daybreak Red cybersecurity model and, in the system card, as a cyber-specialized variant of GPT-6 Astra rather than a model that replaces Astra. The documented API id is gpt-6-cyber. First access is the API and Codex for a select group of approved Daybreak Red customers, widening later to all Daybreak Red customers. This is not a general public launch.
The only published input and output rates are $20 and $100 per million tokens, which OpenAI labels short-context prices. Cached input is $2 per million. Cache writes are $25 per million for 30 minutes or $50 per million for 12 hours. Every long-context cell in the Cyber models table is blank, and the batch endpoint is not supported. No independent tracked score is recorded as of 2026-11-06. ExploitGym 48.5% versus GPT-5.6 Cyber 32.1%, and the honeypot's 99.93%/100% in-scope rates, are OpenAI's claims. The announcement does not retire Astra, Sol, Luna, or the GPT-6.1 lines.
Official source →What we can and can’t confirm
Availability. GPT-6 Cyber was announced on the OpenAI blog dated November 5, 2026, and on X at 2026-11-05T17:09:13Z. The API docs publish the model id gpt-6-cyber and say an organization needs Daybreak Red access and approval before it can send a request. The blog says the first route is the API and Codex, for a select group of approved Daybreak Red customers, with progressive expansion to all Daybreak Red customers. The same paragraph says GPT-5.6 Cyber is being made available to all Daybreak Red customers. The changelog of the same day points new callers at Daybreak setup. None of those pages describes a general public launch.
Price. The Cyber models section of OpenAI's pricing page, checked 2026-11-06, lists gpt-6-cyber at $20.00 input and $100.00 output per million tokens on the short-context row, cached input $2.00, and cache writes $25.00 for 30 minutes or $50.00 for 12 hours. OpenAI's legend defines short context as at most 272,000 input tokens and long context as above that. Every long-context cell on the gpt-6-cyber row is a dash. The model page calls the $20/$100 figures standard short-context prices and still states a 922,000-token maximum input, so the missing cells are an unpublished price, not a statement that the short-context rate holds at the ceiling. This site stores $20/$100 as the headline rate and does not invent a long-context multiplier. The model page lists the batch endpoint as not supported, so no batch discount is recorded.
Context and tools. The model page states a 1,050,000-token context window, maximum input 922,000, and maximum output 128,000. Input is text and image. Output is text. Reasoning tokens are supported. Listed features are streaming, structured outputs, function calling, web search, and prompt caching. No knowledge cutoff appears on that page.
Independent measurement. No tracked benchmark has a GPT-6 Cyber row as of 2026-11-06, so this site stores no score. ExploitGym, the alignment honeypot, and the other cybersecurity evals in the launch post and the system card are not benchmarks this site tracks.
Vendor claims, kept as claims. OpenAI's Preparedness Framework classification for GPT-6 Cyber is Critical in cybersecurity, the same designation it assigns GPT-6 Astra, and up from High for GPT-5.6 Cyber. The launch post reports ExploitGym intended-exploit success of 48.5% against 32.1% for GPT-5.6 Cyber. The system card reports 48.45% on ExploitGym against 42.43% for GPT-6 Astra. The two comparisons use different baselines. The launch post reports the alignment honeypot stayed inside scope in 99.93% of graded rollouts without a deployment prompt and in 100% with the prompt. The system card states the complementary exploitation rates, 0.07% and 0%. The system card also reports further vendor-run cyber evaluations against GPT-6 Astra. Those numbers are not copied into this entry and are not stored as scores.
No supersession. The system card calls GPT-6 Cyber a cyber-specialized variant of GPT-6 Astra for authorized security work, and says it does not outperform Astra on a general coding evaluation. The blog still points at mainline models such as GPT-6 Astra for code review. Nothing in the blog, the model page, the changelog, or the system card says GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol, or GPT-6.1 Terra is replaced. This site does not mark any of them a previous version on account of this release.
- OpenAI — Introducing GPT-6 Cyber, November 5, 2026: Daybreak Red access, Critical classification, ExploitGym 48.5% vs GPT-5.6 Cyber 32.1%, honeypot 99.93%/100% (checked 2026-11-06)
- OpenAI API — gpt-6-cyber model page: id, Daybreak Red approval, 1,050,000 context, short-context $20/$2/$25/$100, batch not supported (checked 2026-11-06)
- OpenAI API pricing — Cyber models table: short-context rates and blank long-context cells (checked 2026-11-06)
- OpenAI API changelog — November 5, 2026, Feature, Model: gpt-6-cyber (checked 2026-11-06)
- OpenAI — GPT-6 Cyber system card: variant of Astra, Critical cybersecurity, ExploitGym 48.45% vs Astra 42.43%, honeypot exploitation 0.07%/0% (checked 2026-11-06)
- OpenAI on X — Introducing GPT-6 Cyber, 2026-11-05T17:09:13Z
- 2026-11-04ReleaseMajor
Moonshot publishes Kimi K3.1 Pro open weights under the Kimi K3.1-Pro License
Separate weights-day event on the line first logged 2026-10-23; that API-launch entry and its RSS guid stay unchanged. Moonshot's X post at 2026-11-04T16:00:35Z links the public repo moonshotai/Kimi-K3.1-Pro: 96 safetensors shards, 1.56 TB, MXFP4 weights and MXFP8 activations, ungated. The license is the Kimi K3.1-Pro License, not MIT or Apache. Model-as-a-Service use needs a separate Moonshot agreement once the licensee's aggregate revenue exceeds $20 million in any 12 months; a commercial product above 100 million monthly users or $20 million monthly revenue must display "Kimi K3.1-Pro". Internal use is exempt, as is use through Moonshot's own products. API id kimi-k3.1-pro and the stored $2/$10 rate are unchanged. No new independent score is recorded.
Official source →What we can and can’t confirm
The October 23 entry remains the API-launch snapshot, including its statement that the weights were still promised, and its RSS guid is unchanged. This entry is the weights day.
The Hugging Face repository moonshotai/Kimi-K3.1-Pro has createdAt 2026-11-02T04:13:51Z. Repository creation is not the public-release date used here. The date is Moonshot's X post at 2026-11-04T16:00:35Z, whose first link unwinds to that repository. Checked 2026-11-05, the repo is public and ungated, with license tag other and license_name kimi-k3.1-pro. It holds 96 safetensors files totaling 1,561,293,160,568 bytes (1.56 TB). The card states MXFP4 weights and MXFP8 activations, 2.8T total parameters, 104B activated, 896 experts with 16 selected and 2 shared, a 1,048,576-token context, and text plus image input.
The license file is titled Kimi K3.1-Pro License. It grants use, modification, redistribution and commercial use, then adds two conditions. If the licensee or its affiliates operate Model-as-a-Service and their aggregate revenue exceeds $20 million over any consecutive 12 months, a separate agreement with Moonshot is required before commercial use of the weights. A commercial product or service above 100 million monthly active users, or above $20 million in monthly revenue, must display "Kimi K3.1-Pro" on the interface. Internal use that does not expose the model or its outputs to third parties is exempt, as is use through Moonshot's official products or certified inference partners.
No tracked independent score is added. The rows already on the model page stay the 2026-10-24 Artificial Analysis readings and the 2026-10-25 launch-chart harvest. The weights post does not announce an API price change, so the stored $2/$10 homepage rate from 2026-10-24 is left as recorded. The launch blog at kimi.com, rechecked 2026-11-05, still says the full weights will be released by Nov 4; that sentence is the promise, not the current status.
- Hugging Face — moonshotai/Kimi-K3.1-Pro, public ungated repo, 96 safetensors shards, createdAt 2026-11-02T04:13:51Z (checked 2026-11-05)
- Kimi.ai on X — "Releasing the model weights of Kimi K3.1 Pro", 2026-11-04T16:00:35Z, link unwinds to the Hugging Face repo
- Hugging Face — Kimi K3.1-Pro License (checked 2026-11-05)
- Kimi launch blog — still says full weights will be released by Nov 4 (rechecked 2026-11-05; the October 23 promise, not the weights-day status)
- 2026-11-03ReleaseMajor
Anonymous "Crouton Alpha" appears on OpenRouter — free, 1M-token context, maker undisclosed
An anonymous third-party listing on OpenRouter, id `stealth/crouton-alpha`, dated 2026-11-03. It is free, with a 1,048,576-token context; the listing gives text, image, video, file and audio input and text output, plus tool calling and structured output. OpenRouter states that prompts and completions may be retained by the provider but are not used for training. OpenRouter's X announcement is timestamped 2026-11-03T20:17:08Z; OpenCode's X account posted a free-week note at 2026-11-03T20:37:04Z.
Show more (2 more paragraphs)
No lab has claimed the model and no licence, weights or parameter count are published, so this is recorded as an appearance rather than a release from a known maker — the same treatment given to Ox Alpha, Space Bunny Alpha, Fledge Alpha and Cat Jam Free.
No model page or score row is added as of 2026-11-04: the id has no models.json entry and no tracked benchmark result. The specs above are OpenRouter's listed route facts, not measurements.
Official source →What we can and can’t confirm
Crouton Alpha is a stealth listing on OpenRouter, not a lab release this site can attribute. OpenRouter's page, checked 2026-11-04, dates it 2026-11-03, lists input and output at $0 per million tokens, and gives a 1,048,576-token context window. It accepts text, image, video, file and audio input and returns text; tool calling and structured output are listed. Those are route facts supplied by the operator, not third-party measurements.
The only identity statement is OpenRouter's: the model is offered by an anonymous third-party provider, OpenRouter is the router rather than the maker, and prompts and completions may be retained by that provider though not used for training. No lab, licence, parameter count or weights are published. The public timestamps are tight — OpenRouter announced the route on X at 2026-11-03T20:17:08Z, and OpenCode's X account posted a free-week note at 2026-11-03T20:37:04Z, twenty minutes later. Neither post names a maker.
That is why this is a tracker event rather than a model page. A model page here states a lab and a licence as facts and carries scores with a named runner behind each one; Crouton Alpha offers none of the three. This entry therefore leaves model_id null, adds no lab field, and adds no score row. It also does not turn the free-week post into a published end date: no end date appears in the material checked for this entry.
The honest summary as of 2026-11-04 is a free, multimodal OpenRouter route with a 1M-token context and an undisclosed maker. Two things would change this entry: a lab claiming the model, or a tracked board publishing a score for it.
- OpenRouter — Crouton Alpha listing: released 2026-11-03, free, 1,048,576-token context, multimodal input, tools and structured output, anonymous-provider statement (checked 2026-11-04)
- OpenRouter on X — Crouton Alpha announcement, 2026-11-03T20:17:08Z (checked 2026-11-04)
- OpenCode on X — free-week post, 2026-11-03T20:37:04Z (checked 2026-11-04)
- OpenRouter — stealth provider page (checked 2026-11-04)
- 2026-11-02ReleaseMajor
Meituan releases LongCat-2.5 GA — 1.6T MoE, 1M context, API live, weights not public
Meituan released LongCat-2.5 on 2026-11-02, the non-Preview follow-up to the model this tracker recorded on 2026-09-25. The API is live under the model id `LongCat-2.5`, with OpenAI- and Anthropic-compatible endpoints; Meituan states 1.6T total parameters (~48B active), a native multimodal agentic model, a 1M-token context and up to 128K output tokens.
Pricing is $0.30 per million input tokens and $1.20 per million output tokens, with cached input at $0.006, described by Meituan as a limited-time pay-as-you-go rate. No independent tracked benchmark score exists as of 2026-11-03: the launch table's numbers are vendor self-reported and are not recorded as scores here. Weights are not published; Meituan says they are "coming soon". This is a separate GA/API event; the 2026-09-25 Preview entry remains that day's snapshot and is not rewritten. No LongCat-2.5 model page is added.
Official source →What we can and can’t confirm
The launch was announced on Meituan's LongCat X account at 2026-11-02T14:35:23Z (22:35 HKT); the blog is dated 2026-11-01 and the API changelog records LongCat-2.5 on 2026-11-02. This entry uses the X/changelog date, 2026-11-02, rather than the blog date.
The model is a 1.6T-parameter MoE with about 48B active parameters, a 1M-token context, up to 128K output tokens, native multimodal input and an agentic focus, per Meituan. The API id is `LongCat-2.5`; Meituan documents both OpenAI- and Anthropic-compatible endpoints. Pricing is $0.30/$1.20 per million input/output tokens, cached input $0.006 (also ¥2/¥8/¥0.04), and is labelled a limited-time pay-as-you-go rate.
Meituan's launch table reports vendor-run benchmark figures, including Terminal-Bench v2.1 88.4, AutomationBench v1.0.6 51.8 and BrowseComp 91.9; these are self-reported claims, not independent measurements, and no score row is added. No LongCat-2.5 weights are published as of 2026-11-03; the vendor says they are coming soon. The 2026-09-25 LongCat-2.5-Preview entry is left unchanged.
- Meituan LongCat on X — LongCat-2.5 launch post, 2026-11-02T14:35:23Z (checked 2026-11-03)
- LongCat — LongCat-2.5 blog, dated 2026-11-01; specs, pricing and vendor benchmark table; numbers are self-reported (checked 2026-11-03)
- LongCat API platform — changelog, LongCat-2.5 entry dated 2026-11-02 (checked 2026-11-03)
- LongCat API platform — LongCat-2.5 pricing, limited-time pay-as-you-go (checked 2026-11-03)
- Meituan — LongCat-2.5 news release, dated 2026-11-02 (checked 2026-11-03)
AI models released in October 2026 (20)
20 events (16 major) — full month →
- 10-29ReleaseMajorMistral publishes Mistral Large 4 open weights under the Mistral Ecosystem Open-Weight license
- 10-28ReleaseMajorGoogle announces Gemini 4 Carbon — cyber-focused Gemini 4 model in Fairwind early access, no public API id, $2/$10 intro then $4/$20
- 10-27ReleaseInclusionAI publishes Ling-3.1-flash-1019 weights — MIT, 560B/25B, 1M context
- 10-26ReleaseMajorOpenAI launches GPT-6.1 Terra — $1/$5, 1,050,000-token context, distinct from GPT-6.1 Sol
- 10-23ReleaseMajorMoonshot launches Kimi K3.1 Pro — $2/$10 on the API, weights promised by November 4
- 10-23ReleaseMajorMiniMax publishes M3.1 Flash weights under the MiniMax Community License
- 10-21ReleaseRecursive ships Flint — 40B/2.5B open-weight coding MoE with a free OpenRouter route
- 10-19ReleaseMajorMiniMax M3.1 Flash attributed — $0.15/$0.60 API; OpenCode reveals Space Bunny Free as the same model
- 10-16ReleaseMajorReflection launches Beam 1.0 — 501B/23B Apache 2.0 open weights
- 10-15ReleaseNVIDIA ships Nemotron 3.5 Bolt 74B-A9B — open agent model aimed at 24GB desktops
- 10-15ReleaseMajorNVIDIA ships Nemotron 3.5 Super VL — 120B/12B open multimodal under OpenMDW-1.1
- 10-15ReleaseMajorMeta releases Muse Spark Open weights — 816B/42B, Muse Model License, hosted at $1.25/$4.25
- 10-15ReleaseStepFun publishes Step 5 Preview weights — StepFun Model License 1.0, not MIT
- 10-13ReleaseMajorGemini 4 Argon API is now available — endpoint gemini-4-argon, not on the Free tier, intro $2/$10 through 2026-11-30
- 10-12ReleaseMajorAnonymous "Cat Jam Free" appears on OpenCode Zen — free, 262K listed context, maker undisclosed
- 10-12ReleaseMajorMeta launches Muse Wonder — $4/$20 per million tokens, 1M context, on the Meta Model API
- 10-07ReleaseMajorAnthropic launches Claude Haiku 5.5 — cheapest and fastest small model to date, $0.10/$0.50 up to a 100K prompt
- 10-06ReleaseMajorAnonymous "Exo Free" appears on OpenCode Zen — free, 1M-token listed context, maker undisclosed
- 10-06ReleaseMajorMistral Large 4 enters public preview — 1.05T multimodal MoE, open weights promised for late October
- 10-01ReleaseMajorAnonymous "Fledge Alpha" appears on OpenCode Zen — free, 1M-token listed context, maker undisclosed
AI models released in September 2026 (22)
22 events (19 major) — full month →
- 09-30ReleaseMajorGoogle announces Gemini 4 Argon — frontier model, not yet generally available, and priced at $2/$10 only for an introductory period
- 09-29ReleaseInclusionAI (Ant Group) ships Ling 3.1 Flash — 560B/25B hybrid reasoning, free on two gateways through October 13
- 09-29ReleaseMajorOpenAI ships GPT-6.1 Sol — same $2/$10 as GPT-6 Sol, one real independent gain, ties Astra on HLE
- 09-28ReleaseMajorClaude Sonnet 5.5 launches — Sonnet 5's $2/$10 unchanged, "not at the capability frontier" by Anthropic's own card
- 09-25ReleaseMajorMeituan releases LongCat-2.5-Preview — 1.6T parameters, 1M context, no benchmarks published
- 09-23ReleaseMajorAnonymous "Space Bunny Alpha" appears on OpenRouter — free, 1M-token context, maker undisclosed
- 09-22ReleaseMajorGPT-6 Luna launches alongside Sol — cheapest GPT-6 at $0.10/$0.50, ties GPT-5.6 Luna on every clean comparison
- 09-22ReleaseMajorOpenAI launches GPT-6 Sol — half GPT-5.6 Sol's promotional price, no measurable gain over it yet
- 09-22ReleaseMajorClaude Opus 5.5 launches — Anthropic's "new leading model" at 20% below Opus 5's per-token price
- 09-22ReleaseMajorXiaomi launches MiMo-V2.6-Flash — 309B open-weight at $0.14/$0.28 per 1M tokens
- 09-22ReleaseMajorXiaomi launches MiMo-V2.6-Pro — MIT-licensed 1.02T open-weight at $0.435/$0.87 per 1M tokens
- 09-21ReleaseMajorGrok 4.7 launches — same price as Grok 4.6, thinner independent record so far
- 09-20ReleaseMajorStepFun launches Step 5 Preview — $1.00/$2.70 per 1M tokens, one independent score, open weights promised for October 15
- 09-17ReleaseUnion Alpha revealed as Unbiased's Pareto 26.9 — a multi-model blend, now paid at $2.50/$7.50 per 1M tokens
- 09-16ReleaseMajorAnonymous "Union Alpha" appears on OpenRouter — free public API, maker undisclosed
- 09-10ReleaseMajorDeepSeek V4.1 Flash launches — replaces V4 Flash; V4 Pro API remains available
- 09-09Price changeGLM-5.3-Flash launch promo ends — back to $0.15/$0.50 per 1M tokens
- 09-04ReleaseMajorAnonymous "Omen Alpha" appeared on OpenCode Go — no longer listed as of September 17
- 09-03ReleaseMajorGPT-6 Astra launches — OpenAI's first model to cross its own "Critical" cybersecurity threshold
- 09-02ReleaseMajorMuse Spark 1.3 launches — same price as 1.2, weights decision now explicitly undecided
- 09-02ReleaseMajorGemini 3.8 Flash launches — same price as 3.7 Flash, built on its checkpoint, not a new pretraining run
- 09-01ReleaseMajorClaude Fable 5.1 launches — successor to Fable 5, same price, cheaper cache reads
AI models released in August 2026 (14)
14 events (10 major) — full month →
- 08-28ReleaseMajorTencent Hy4 preview — 770B open-weight at $0.834/$2.501 per 1M tokens, added to tracker retroactively
- 08-26ReleaseMajorQwen3.8-Flash-Next ships open-weight — a Qwen4-architecture preview under a Qwen3.8 label
- 08-26ReleaseMajorOx Alpha revealed as GLM-5.3-Flash — GA, MIT license, $0.15/$0.50 per 1M tokens
- 08-26Price changeGPT-5.6 Sol price drops to $4.00/$20.00 per 1M tokens (promotional)
- 08-25ReleaseGLM-5.3 weights published on Hugging Face — Z.ai keeps its two-week promise
- 08-20ReleaseMajorAnonymous "Ox Alpha" stealth model appears on OpenRouter
- 08-19Price changeMajorGLM-5.3 API goes live — $1.40/$4.40 per 1M tokens, same price as GLM-5.2
- 08-14ReleaseMajorGLM-5.3 released — subscription-only for now
- 08-13ReleaseGemini 3.7 Flash reaches GA
- 08-13ReleaseMajorDeepSeek V4 Pro (0813) exits preview
- 08-12ReleaseMajorGrok 4.6 released
- 08-05ReleaseMajorMuse Spark 1.2 launches — a coding-focused point release in Meta's closed model line
- 08-05DeprecationClaude Opus 4.1 retired from the API
- 08-03ReleaseMajorQwen3.8-Max launches
AI models released in July 2026 (6)
6 events (4 major) — full month →
- 07-31ReleaseMajorDeepSeek V4 Flash 0731 ships with a changelog line, no launch post
- 07-24ReleaseMajorClaude Opus 5 released
- 07-21ReleaseGemini 3.6 Flash reaches GA
- 07-16ReleaseMajorKimi K3 launches (API); open weights follow July 26-27
- 07-09ReleaseMajorGPT-5.6 Luna launches — cheapest tier of the GPT-5.6 family
- 07-09ReleaseGPT-5.6 Sol reaches general availability
AI models released in June 2026 (5)
5 events (2 major) — full month →
- 06-30ReleaseMajorClaude Sonnet 5 released
- 06-16ReleaseGLM-5.2 reaches general release
- 06-15DeprecationClaude Sonnet 4 and Opus 4 retired from the API
- 06-09ReleaseMajorClaude Fable 5 — first Mythos-class model goes GA
- 06-01ReleaseMiniMax M3 — open-weight value play, added to tracker retroactively
AI models released in May 2026 (1)
1 event — full month →
- 05-28ReleaseClaude Opus 4.8 launches
AI model release tracker FAQ
How many AI models were released in November 2026?
17 tracked events in November 2026, out of 85 since this tracker started. Each one is a release, price change or deprecation we could confirm against an official source — announcements we could not confirm are not listed at all.
Which AI model releases here actually matter?
67 of the 85 tracked events are marked major. That label is a judgment, not a vendor claim: it means the release changed something measurable — a price, a capability tier, or a benchmark position — rather than that it was heavily promoted.
Where do these AI model release dates come from?
Official sources only — vendor blogs, API changelogs, model cards. Every event's full entry links to one. Where a model shipped with no announcement, the date is reconstructed from version strings and repository timestamps and the row says so, because a date we inferred and a date a vendor published are not the same fact.
What counts as an event on this AI model tracker?
Release, price change, deprecation — across 20 labs so far. Deprecations and price changes are tracked alongside launches because for anyone already running a model in production they matter more than a new name does.