October 2026 archive
Every AI model release in October 2026
- 2026-10-29Release
Mistral publishes Mistral Large 4 open weights under the Mistral Ecosystem Open-Weight license
Official source →Mistral published open weights for the improved Mistral Large 4 (Le Chonk) on 2026-10-29 under the Mistral Ecosystem Open-Weight license — a modified MIT license; model-as-a-service use needs a separate commercial agreement. The checkpoint is 1.05T total parameters, 49B active per token and 52B including embeddings/output. API pricing moved to $0.70/$2.10 per 1M tokens (was $0.68/$2.09), cached input $0.07. The 2026-10-06 preview entry stays unchanged. Mistral's ~41% Terminal-Bench 4.0 claim is vendor-reported and is not recorded as an independent score; the existing vals.ai 22.73 is the preview checkpoint, and no independent score exists for the weights release.
- 2026-10-28Release
Google announces Gemini 4 Carbon — cyber-focused Gemini 4 model in Fairwind early access, no public API id, $2/$10 intro then $4/$20
Official source →Google announced Gemini 4 Carbon on 2026-10-28, and as of 2026-10-29 it is not generally available: access starts with a subset of trusted cyber defenders in the Fairwind Program, with wider availability promised only "as soon as possible" and no GA date. No public API model identifier is published, and Google's API models page still lists gemini-4-argon as the Gemini 4 flagship.
Pricing is $2/$10 per million tokens introductory, rising to $4/$20 after the intro period. Google states no end date for Carbon's intro rate — unlike Argon's through 2026-11-30 — so the page carries $4/$20 as headline and $2/$10 as the alternate.
Artificial Analysis's model page independently shows an Intelligence Index v5.0 of 59 at high effort (checked 2026-10-29); Google calls that #1. SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0% are Google's own cybersecurity figures, not independent tracked scores, and no tracked score row is recorded here. Carbon does not replace Argon.
- 2026-10-27Release
InclusionAI publishes Ling-3.1-flash-1019 weights — MIT, 560B/25B, 1M context
Official source →New checkpoint and weights day on the Ling 3.1 Flash line first logged on 2026-09-29. InclusionAI published Ling-3.1-flash-1019 and Ling-3.1-flash-base on Hugging Face on 2026-10-27, with MIT open weights, 560B total / 25B active parameters and a 1M context window per the lab. Ant Ling's X thread claims AA Intelligence Index v4.3.2 41→44 (a version-stamped vendor claim; v4.3.2 absolute scores are not comparable to Index v5, shipped 2026-10-28), AA-Briefcase ~1400→1524 Elo, AutomationBench-AA 62%→68%, CritPt 18%→26% and SciCode 54%→57%. None of those is a benchmark this site tracks, and no independent tracked row exists, so no score is recorded and no model page is opened; the 2026-09-29 entry remains the launch snapshot.
- 2026-10-26Release
OpenAI launches GPT-6.1 Terra — $1/$5, 1,050,000-token context, distinct from GPT-6.1 Sol
Official source →OpenAI released GPT-6.1 Terra on 2026-10-26 at $1/$5 per million tokens, 50% below GPT-5.6 Terra and GPT-6.1 Sol per OpenAI, with cached input $0.05 and cache write $1.25 for 30 minutes. It has a 1,050,000-token context, June 1, 2026 knowledge cutoff, text and image input, text output and reasoning effort none to max; API, ChatGPT Work/Codex for Plus/Pro/Enterprise/Edu and Bedrock GA the same day. No independent tracked score exists. OpenAI's Terminal-Bench 4.0 claim is 50.3% at max versus 21.5% for GPT-5.6 Terra; it is recorded as self-reported. The Agents' Last Exam claim is a +15pp gain at medium, not an absolute score; OSWorld 2.1 and MMMU-Pro claims are not recorded as scores. This is a distinct OpenAI line, not a Sol tier.
- 2026-10-23Release
Moonshot launches Kimi K3.1 Pro — $2/$10 on the API, weights promised by November 4
Official source →API id kimi-k3.1-pro since 2026-10-23. Homepage price $2/$10, cache hit $0.20, cache write $2; Highspeed is $4/$20 and is not a second record. AA row has no effort suffix (checked 2026-10-24): HLE 52.46, Terminal-Bench 4.0 67.17, Analyst 38.75. K3's AA max row the same day is 46.90, 12.63, and 38.75, so the gaps are not matched-effort. Weights are promised by 2026-11-04 and are not public.
- 2026-10-23Release
MiniMax publishes M3.1 Flash weights under the MiniMax Community License
Official source →Weights at MiniMaxAI/MiniMax-M3.1-Flash are public as of 2026-10-23 (gated false; card lastModified 2026-10-23T00:59Z). License is MiniMax Community License, not MIT: non-commercial use is free; commercial use must display Built with MiniMax M3.1-Flash; yearly revenue above $20 million needs prior written authorization. The card states about 428B total and 23B active. The 2026-10-19 entry's no-weights line stays as that day's snapshot. OpenCode Go marked the free route deprecated on 2026-10-22 (models.dev 665a3e033f); OpenRouter on 2026-10-23 lists the paid id at $0.15/$0.60 and no :free id.
- 2026-10-21Release
Recursive ships Flint — 40B/2.5B open-weight coding MoE with a free OpenRouter route
Official source →Recorded 2026-10-25, four days after the fact: Recursive published Flint's weights and quantized builds under its RecursiveSI organization on 2026-10-21, the same day a free recursive/flint-v1.0:free route appeared on OpenRouter (created 2026-10-21T22:46Z). About 40B total with 2.5B active parameters per token, a 262,144-token context window, positioned for coding agents and described by the maker as built by its "autoresearch system Eureka Machine" — a claim about the training process this site cannot check from outside. An attributed release, not a stealth listing: the maker names itself on both the repository and the route. Artificial Analysis has no page for it (checked 2026-10-25, 404) and models.dev carries no Recursive provider entry (checked 2026-10-25), so no board this site tracks has scored it — tracker-only, no model page. Marked minor: the maker is known, the route is free, and the only verifiable facts are the architecture and the artifacts.
- 2026-10-19Release
MiniMax M3.1 Flash attributed — $0.15/$0.60 API; OpenCode reveals Space Bunny Free as the same model
Official source →MiniMax's pricing page lists MiniMax-M3.1-Flash at $0.15/$0.60 with 1M context and no 512K step (checked 2026-10-19). AA HLE 47.96 vs MiniMax M3's 39.0 is a real gap on that matched pair. AA Terminal-Bench 4.0 is 25.76 vs M3's AA 2.02 (23.74, past the 12.4 band); M3's stored cell remains vals.ai 1.01, a different harness. models.dev commit fbfb5229 (2026-10-19) titles the OpenCode free-route reveal "Reveal Space Bunny as MiniMax M3.1 Flash Free." No HF weights found under that name the same day.
- 2026-10-16Release
Reflection launches Beam 1.0 — 501B/23B Apache 2.0 open weights
Official source →First model from Reflection. Artificial Analysis (checked 2026-10-17): HLE 34.66, Terminal-Bench 4.0 3.03, Intelligence Index 30 (pre-v5 reading; Beam's absolute score is not comparable to Index v5, shipped 2026-10-28 — this site does not publish Beam 1.0's Index). Reflection's own HLE no-tools table reads 36.0.
- 2026-10-15Release
NVIDIA ships Nemotron 3.5 Bolt 74B-A9B — open agent model aimed at 24GB desktops
Official source →Recorded 2026-10-25, ten days after the fact: this entry backfills the second member of the Nemotron 3.5 trio that NVIDIA announced on 2026-10-15 — this site logged Super VL that day and missed this one until a 10-25 sweep. 74B total with 9B active parameters per token, an open mixture-of-experts model positioned for agent workloads on roughly 24GB of desktop memory. Weights sit under the nvidia org on Hugging Face: the 24GB and GGUF repositories carry a 2026-10-07 createdAt and the BF16 repository 2026-10-09 (checked 2026-10-25) — those are repo-creation stamps, not verifiable public-release dates, so this entry uses the announcement date. Artificial Analysis has no page for the model (checked 2026-10-25, 404), so no board this site tracks has scored it — tracker-only, no model page, the same rule as Ling 3.1 Flash. The third trio member, Bolt E4B-4GB, is deliberately not tracked: described as 8B total / 4B active on a Gemma 4 E4B base for 4GB-class edge deployment, below this site's frontier tracking bar; its weights landed 2026-10-23 (repo lastModified, checked 2026-10-25). Marked minor rather than major: the maker is known, there is no benchmark signal, and the trio's measured headline — the AA Intelligence Index lead among US open-weight models — referred to Super VL's pre-v5 Index standing, already logged on 2026-10-15. Under Intelligence Index v5 (2026-10-28), Artificial Analysis's released-weight open leaders are DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32, and this site still does not publish that Index for Nemotron 3.5 Bolt.
- 2026-10-15Release
NVIDIA ships Nemotron 3.5 Super VL — 120B/12B open multimodal under OpenMDW-1.1
Official source →Open weights for multimodal agents. Artificial Analysis HLE 30.72 (checked 2026-10-17). Terminal-Bench 4.0 not recorded: AA's embedded terminalBench40 field reads 0. Bolt edge models announced the same day are tracker-only here.
- 2026-10-15Release
Meta releases Muse Spark Open weights — 816B/42B, Muse Model License, hosted at $1.25/$4.25
Official source →Meta's 2026-10-15 post opens Muse Spark Open weights (816B total / 42B active, 1M context) under the Muse Model License. OpenRouter created meta/muse-spark-open the same day at $1.25/$4.25. The Hugging Face repo's createdAt is 2026-09-15. Artificial Analysis xhigh: HLE 43.88, Terminal-Bench 2.1 76.40, Terminal-Bench 4.0 8.08, GPQA 88.99 (checked 2026-10-19). Tracked as its own family, not a Spark 1.3 point release.
- 2026-10-15Release
StepFun publishes Step 5 Preview weights — StepFun Model License 1.0, not MIT
Official source →StepFun posted the weights on the day it had named. Hugging Face now has stepfun-ai/Step-5-Preview-BF16, an FP8 repo, and an NVFP4 repo, under the StepFun Model License 1.0: Apache-style use and redistribution, plus an attribution duty above 100 million monthly users or $20 million revenue, and a separate commercial license for device makers and large model-as-a-service operators.
The benchmark sheet did not move with the weights. DeepSWE's server-rendered model list on 2026-10-15 still has no Step 5 row. Artificial Analysis's existing rows are unchanged: Humanity's Last Exam 46.48 and AA-AnalystAgent 35.0.
- 2026-10-13Release
Gemini 4 Argon API is now available — endpoint gemini-4-argon, not on the Free tier, intro $2/$10 through 2026-11-30
Official source →Google's Gemini API model page says gemini-4-argon is now available, and the pricing page says it is not on the Free tier (rechecked 2026-10-19). OpenRouter created google/gemini-4-argon on 2026-10-13. Introductory paid pricing is $2/$10 through 2026-11-30, then $4/$20. Vendor-stated context window is 1,048,576. This updates the September 30 announced-only record; scores on the model page are unchanged.
- 2026-10-12Release
Anonymous "Cat Jam Free" appears on OpenCode Zen — free, 262K listed context, maker undisclosed
Official source →An unattributed free route on OpenCode Zen, id cat-jam-free. The operator-supplied models.dev file dates it 2026-10-12. OpenCode's docs commit adding the row (b4f19bfe, #55029) is timestamped 2026-10-12T21:57:28-07:00, which is 2026-10-13 04:57 UTC. This entry uses 2026-10-12, the date on the route file.
OpenCode's docs, checked 2026-10-13, call it a stealth model that is free for a limited time, and say the team is using that time to collect feedback and improve the model. The same page's data-use list says that during its free period, collected data may be used to improve the model. No end date is published there.
The toml lists text-only input, a 262,144-token context, a 131,072-token output limit, tool calling, and reasoning-effort values low through max, with no default named, at $0. Those are listed route facts. No maker, licence or weights. No model page: Artificial Analysis returns 404 for /models/cat-jam-free (checked 2026-10-13). Other boards were not re-read in a browser today, so this entry does not call them empty.
- 2026-10-12Release
Meta launches Muse Wonder — $4/$20 per million tokens, 1M context, on the Meta Model API
Official source →Meta's model after Muse Spark, launched 2026-10-12 on the Meta Model API and Muse Code at $4.00 in / $20.00 out per million tokens, cached input $0.20, with a 1M context and no length-based price step on the model card (checked 2026-10-13). Artificial Analysis's max-effort row reads Humanity's Last Exam 50.09 and Terminal-Bench 4.0 52.02. Against Muse Spark 1.3 at the same effort on that same page, the HLE gap is 1.39 points (Spark stored here as 48.7) and stays inside the 2.0 band, while Terminal-Bench 4.0's 52.02 against Spark's 33.33 on that runner clears the 12.4 band. Muse Wonder Full is the same-day subscription model, with no token price; it is not a separate record here.
- 2026-10-07Release
Anthropic launches Claude Haiku 5.5 — cheapest and fastest small model to date, $0.10/$0.50 up to a 100K prompt
Official source →Anthropic's small model, one day old, and already carrying three independent scores: Humanity's Last Exam 44.39 no tools at max effort (Artificial Analysis), LiveBench 72.07 overall on the board's default xHigh row (its Max Effort row reads 67.88, another tier inversion of the usual order — the same shape as Opus 5.5's ARC-AGI-2 row), and Terminal-Bench 4.0 35.35 at max effort (vals.ai; that row reports 3 provider refusals, 1.51%, scored as failures) against Anthropic's own 39.2. That third pair is not like-for-like: Anthropic ran Terminal-Bench 4.0 with no fallback model and 1.8% of trials stopped on safeguards and failed, where Sonnet 5.5's cyber and frontier blocks fall back to Sonnet 5 by default. The number worth watching is the price step: $0.10/$0.50 per million tokens up to a 100,000-token prompt, then $0.50/$2.50 — the lowest long-context threshold on this site, and past it Haiku 5.5 costs more than GPT-6 Luna. Knowledge cutoff June 2026, default effort medium, no weights.
- 2026-10-06Release
Anonymous "Exo Free" appears on OpenCode Zen — free, 1M-token listed context, maker undisclosed
Official source →A free-tier preview reasoning model listed on OpenCode Zen as `exo-free` on 2026-10-06 — the date on the operator-supplied models.dev entry, the same day OpenCode's own documentation update listing it was merged (2026-10-06T22:27Z), and within half an hour of that an independent models.dev sweep recorded it (2026-10-06 22:54 UTC).
It is the fourth model in Zen's free pool in under two weeks, after Space Bunny Free (09-23), LongCat 2.5 Preview Free (09-25) and Fledge Alpha Free (10-01); three of those four are unattributed, and the exception is LongCat, which Meituan announced.
No lab, developer or country of origin has claimed Exo Free and no weights are published, so this site records it as an appearance rather than a release from a known maker — the treatment already given to Ox Alpha, Space Bunny Alpha and Fledge Alpha.
OpenCode's route metadata gives the specs (models.dev `exo-free.toml`, vendor-supplied and dated 2026-10-06, still matching the live models.dev catalogue on 2026-10-07): a 1,048,576-token context, a 131,072-token output limit, text-and-image input, tool use, reasoning effort fixed at `high` with temperature disabled, and $0 input and output — vendor-listed figures, not third-party measurements.
Two things the listing does say plainly: the free period is limited with no published end date, and Exo Free is named in OpenCode's docs as an exception to its zero-retention policy — during the free period, collected data may be used to improve the model (the other two exceptions named are Big Pickle and Fledge Alpha Free; Space Bunny Free and LongCat 2.5 Preview Free are described as zero-retention).
No benchmark score is recorded here, and none is available: searched directly on 2026-10-07 with nothing found on LiveBench's 2026-06-25 table, the DeepSWE board, MathArena's HMMT table, vals.ai's four benchmark tables, tbench.ai, Toolathlon, Snorkel's Agents' Last Exam board, and Artificial Analysis (no /models/exo-free page — 404). ARC Prize is the one gap: its leaderboard now redirects to Kaggle, so that board is unchecked today rather than clear.
- 2026-10-06Release
Mistral Large 4 enters public preview — 1.05T multimodal MoE, open weights promised for late October
Official source →Mistral's first model on this site's tracker: a 1.05T-parameter (49B active) natively multimodal MoE, API-preview at $0.68/$2.09 per 1M tokens (the $1.36/$4.18 Artificial Analysis bills is the priority tier), with open weights promised for late October 2026.
The launch chart is entirely self-reported — DeepSWE v1.1 61.7% is the only claim on a benchmark this site tracks, filed as a self-reported row — and the one independent read so far is Artificial Analysis's Intelligence Index at 38 (#64 of 225), a pre-v5 reading checked 2026-10-06 whose absolute score is not comparable to Intelligence Index v5 (shipped 2026-10-28); this site does not publish Mistral Large 4's Index. Zero tracked-benchmark independent scores on day one; the provenance story is the thing to watch.
- 2026-10-01Release
Anonymous "Fledge Alpha" appears on OpenCode Zen — free, 1M-token listed context, maker undisclosed
Official source →A free-tier stealth model listed on OpenCode Zen as `fledge-alpha-free`, alongside other unattributed free entries (Big Pickle, Space Bunny Alpha) and one attributed one (LongCat 2.5 Preview, announced by Meituan on 2026-09-25 — corrected 2026-10-07, this entry had grouped it with the unattributed ones).
No lab, developer or country of origin has been claimed, and no weights are published, so this site records it as an appearance rather than a release from a known maker — the same treatment given to Ox Alpha in August.
OpenCode itself supplied the route metadata: its listing on models.dev (checked 2026-10-02) gives a 1,048,576-token context, a 131,072-token output limit, text-and-image input, reasoning effort of low/high/max and $0 input and output, all dated by the operator at 2026-10-01 — so those are vendor-listed figures, not third-party estimates.
What none of that establishes is who is behind the model. An independent tracker (StealthModels) reports roughly 61 output tokens per second in its own tests and tokenizer inconsistencies that suggest a router over unidentified faster models — a hypothesis, not an attribution.
No benchmark score is recorded on this site: none of the benchmarks tracked here lists the model, and a router's output cannot be pinned to one checkpoint.