The Model Gap

Moonshot AI

Kimi K3.1 Pro

On sale since 2026-10-23 at $2/$10 per million tokens, one rate on a 1,048,576-token window. Weights have been public since 2026-11-04 at moonshotai/Kimi-K3.1-Pro under the Kimi K3.1-Pro License, not MIT. Artificial Analysis, no effort suffix: HLE 52.46, Terminal-Bench 4.0 67.17, Analyst 38.75. Launch-chart rows stay self-reported: HLE with tools 70.0, DeepSWE 75.5, GPQA 94.2, OSWorld 2.1 79.6.

Kimi K3.1 Pro benchmarks and pricing, every number sourced: 7 tracked Kimi K3.1 Pro benchmark scores (3 independently run, 4 still resting on a vendor’s own claim), priced at $2.00 per million input tokens and $10.00 per million output.

Kimi K3.1 Pro architecture: Post-trained Kimi K3, natively multimodal; 2.8T total parameters (104B activated per token); 1M-token context window.

Released
2026-10-23
Availability
API since 2026-10-23 as kimi-k3.1-pro; open weights as of 2026-11-04 at huggingface.co/moonshotai/Kimi-K3.1-Pro
License
open-weights
Context window
1M tokensvendor-stated (platform.kimi.ai/docs/pricing/chat: all K3-series models support a 1,048,576-token context window; checked 2026-10-24)
Knowledge cutoff
Not disclosed
Verified
sources checked 2026-10-24–2026-10-25
Parameters
2.8T (104B active)
Architecture
Post-trained Kimi K3, natively multimodal

Kimi K3.1 Pro’s verified record

Kimi K3.1 Pro’s most-compared rival is GLM-5.3-Flash: 3 leads and 3 not callable across their 6 shared comparisons. Kimi K3.1 Pro is announced at $2.00/$10.00 per 1M tokens (in/out) vs GLM-5.3-Flash’s $0.15/$0.50.

Against the 167 head-to-head comparisons Kimi K3.1 Pro shares with other tracked models:

53 real gaps17 inside the band97 not called

A gap counts for Kimi K3.1 Pro only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Kimi K3.1 Pro trails on 16 of them.

HLE · no tools Reasoning

±2 is noise
Behind
GPT-6 Astra −2.2 · Claude Sonnet 5.5 −2.5 · Gemini 4 Argon −4.6 · Claude Fable 5.1 −6.6 · Claude Opus 5.5 −8.9 · 2 previous: Claude Opus 5 −2.4 · Claude Fable 5 −3.0
Ahead
Muse Wonder +2.4 · MiMo-V2.6-Pro +3.1 · Muse Spark 1.3 +3.8 · MiniMax M3.1 Flash +4.5 · Gemini 3.8 Flash +4.7 · Gemini 3.1 Pro Preview +5.5 · Kimi K3 +5.6 · Step 5 Preview +6.0 · Claude Haiku 5.5 +8.1 · Muse Spark Open +8.6 · DeepSeek V4.1 Flash +9.1 · Qwen3.8-Max +9.4 · Grok 4.7 +9.4 · GLM-5.3 +10.2 · DeepSeek V4 Pro (0813) +11.5 · GLM-5.3-Flash +12.6 · MiniMax M3 +13.5 · GPT-6 Luna +14.0 · Qwen3.8-Flash-Next +14.5 · Mistral Large 4 +17.4 · Beam 1.0 +17.8 · Nemotron 3.5 Super VL +21.7 · 10 previous: GPT-5.6 Sol +3.0 · Claude Opus 4.8 +3.8 · Gemini 3.7 Flash +4.6 · GPT-6 Sol +4.6 · Muse Spark 1.2 +7.0 · Grok 4.6 +9.6 · Claude Sonnet 5 +11.2 · GLM-5.2 +11.4 · GPT-5.6 Luna +13.0 · DeepSeek V4 Flash (0731) +13.9
Tie
1 model within ±2
Unverified
1 model — vendor-reported on one side

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Behind
GPT-6.1 Sol −11.3 · GPT-6 Astra −12.5 · Grok 4.7 −13.8 · Claude Haiku 5.5 −15.0 · Claude Opus 5.5 −17.5 · Claude Fable 5.1 −18.8 · Claude Sonnet 5.5 −18.8 · 2 previous: Claude Opus 5 −15.0 · Gemini 3.7 Flash −21.3
Ahead
GLM-5.3-Flash +13.8 · MiniMax M3 +28.8
Tie
16 models within ±11.2

Terminal-Bench 4.0 Terminal ops

±12.4 is noise
Ahead
GLM-5.3-Flash +34.3 · MiniMax M3.1 Flash +41.4 · Beam 1.0 +64.1
Unverified
1 model — vendor-reported on one side
Setup-dependent
6 models — scored on a different harness or effort tier

No verdict for Kimi K3.1 Pro anywhere on DeepSWE, HLE · with tools, OSWorld 2.0 · v2 1 full ( nothing independently confirmed on both sides); GPQA Diamond ( saturated).

  • None of Kimi K3.1 Pro’s coding comparisons are independently confirmed on both sides yet.

Kimi K3.1 Pro announced pricing

$2.00 in / $10.00 out per 1M tokens — official pricing source

What Kimi K3.1 Pro costs per job

Kimi K3.1 Pro cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.300
A codebase review1,000K / 100K$3.00
A day of agent work10,000K / 1,000K$30.00

Computed from Kimi K3.1 Pro’s list rates above — cache discounts and batch tiers are not applied.

Kimi K3.1 Pro is one of 2 Moonshot AI models tracked on this site, at these official list prices.

Moonshot AI model pricing, official list rates
ModelIn / 1MOut / 1M
Kimi K3.1 Pro$2.00$10.00
Kimi K3$3.00$15.00

Kimi K3.1 Pro benchmark scores

Kimi K3.1 Pro benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
52.5artificialanalysis.ai · 2026-10-24
Terminal-Bench 4.0[2]
Terminal ops · ±12.4 is noise
67.2artificialanalysis.ai · 2026-10-24
AnalystAgent[3]
Spreadsheet & document analysis · ±11.2 is noise
38.8artificialanalysis.ai · 2026-10-24
HLE(with tools)[4]
Reasoning · ±2 is noise
70.0kimi.com · 2026-10-25
DeepSWE[5]
Long-horizon coding · ±9.5 is noise
75.5kimi.com · 2026-10-25
GPQA Diamondsaturated[6]
Expert science Q&A — not ranked at any gap size
94.2kimi.com · 2026-10-25
OSWorld 2.0(v2 1_full)[7]
Long-horizon computer use · ±9.7 is noise
79.6kimi.com · 2026-10-25

Who ran these numbers: 3 of 7 independent — artificialanalysis.ai (3); vendor self-reported (4).

  1. HLE: AA's own run, board row 'Kimi K3.1 Pro' with no effort suffix: hle 0.524559777571826 = 52.46, no tools. Read from the Kimi K3.1 Pro page on 2026-10-24. Moonshot's chart lists 'HLE-Full' (no tools) at 51.1 — also not stored as a row, same reason. The two are not directly comparable anyway: the vendor's no-tools figure covers the full question set, AA's run is text-only.
  2. Terminal-Bench 4.0: AA's own run, board row 'Kimi K3.1 Pro' with no effort suffix: terminalBench40 0.671717171717172 = 67.17. The variant is default because the row does not name an effort. Read from the Kimi K3.1 Pro page on 2026-10-24. Moonshot's own chart lists 65.7 at max effort through the Claude Code harness — not stored as a row: this cell holds AA's independent run, and vendor figures yield to independent ones.
  3. AnalystAgent: AA's own run, board row 'Kimi K3.1 Pro' with no effort suffix: analystAgent 0.3875 = 38.75 pass^5. Read from the Kimi K3.1 Pro page on 2026-10-24.
  4. HLE: Moonshot's launch chart, row 'HLE-Full w/ tools'; max thinking effort per the chart-wide footnote. Every HLE-with-tools number on this site is vendor-supplied the same way — the predecessor K3's 56 came from its model card — because no third party runs this variant.
  5. DeepSWE: Moonshot's launch chart, row 'DeepSWE v1.1', run through the Claude Code harness per the chart's own comment; max thinking effort. Same version the tracked column serves — deepswe.datacurve.ai's board is v1.1 (checked 2026-10-25) and lists no K3.1 Pro entry, so this figure holds the cell until the board does. A board row would use the standardized mini-swe-agent harness, a different setup the harness gate already accounts for.
  6. GPQA Diamond: Moonshot's launch chart, row 'GPQA-Diamond'; max thinking effort. vals.ai, this column's independent source, archived its GPQA board on 2026-09-01 and does not score models released after that date, so no independent value is expected. GPQA is saturated on this site's trust grades either way, so this number feeds no gap calls.
  7. OSWorld 2.0: Moonshot's launch chart, row 'OSWorld V2.1'; max thinking effort per the chart-wide footnote. Recorded under the v2_1_full variant — the slice vendor launch charts cite (this site's other v2_1_full rows are launch-chart numbers the same way). The chart does not state full-vs-offline; full is the pinned default for exactly this reason. No independent v2_1 row exists for K3.1 Pro yet.

Notes on the record

Moonshot put Kimi K3.1 Pro on the API on 2026-10-23 (Kimi Code changelog and platform homepage, checked 2026-10-24). The model id is kimi-k3.1-pro. Highspeed, id kimi-k3.1-pro-highspeed, is the same model sold for speed, not a second record.

The homepage lists $2.00/$10.00 per million tokens, cache writes $2.00, cache hits $0.20. Highspeed is $4.00/$20.00 with the same cache hit — not the cheaper alternate. The model guide says the rate does not step up with context length. The K3-series pricing page states a 1,048,576-token window for every model in the series.

Moonshot calls this a 3T-class post-train of Kimi K3. AA's payload reads 2,800 billion total / 104 billion active, matching K3's stored 2.8T/104B. The homepage calls K3 the previous-generation flagship, and K3 is still on the pricing page. Open weights shipped 2026-11-04 at moonshotai/Kimi-K3.1-Pro, 1.56 TB of MXFP4, under the Kimi K3.1-Pro License rather than MIT. API id and $2/$10 are unchanged. The FAQ names the two commercial conditions.

Artificial Analysis's kimi-k3-1-pro page, checked 2026-10-24, has no effort suffix on the row. Humanity's Last Exam is 52.46. Terminal-Bench 4.0 is 67.17. AA-AnalystAgent is 38.75 pass^5. K3's AA max row the same day is HLE 46.90, Terminal-Bench 4.0 12.63, and Analyst 38.75. K3.1 does not name an effort, so those pairs are not matched-effort calls. Moonshot's own Terminal-Bench 4.0 figure is 65.7 on a Claude Code harness, and it is not stored.

The launch blog's benchmark table ships as both a rendered image and embedded page data; the four vendor scores were taken from the embedded data on 2026-10-25. The chart's Agents' Last Exam (ALL) 43.2 is not recorded: Moonshot's annotation says its ALE data is the official v1.1 release of 122 tasks (86 CLI) — a different set from the tracked leaderboard's rolling Overall. The chart's OSWorld figure (V2.1, 79.6) is recorded under the v2_1_full variant — the same slice every other vendor launch-chart OSWorld number on this site uses.

Compare with

Kimi K3Inkling 2 SmallGPT-6 CyberGemini 4 Carbonany model…

FAQ

What is Kimi K3.1 Pro?

A post-trained Kimi K3 that Moonshot put on the API on 2026-10-23. The id is kimi-k3.1-pro. It has a 1,048,576-token window and native visual input. Highspeed is the same model on a faster, more expensive route, not a separate capability.

What does Kimi K3.1 Pro cost?

The platform homepage, checked 2026-10-24, lists $2 per million input tokens and $10 per million output tokens. Cache hits are $0.20 and cache writes are $2. Highspeed is $4 and $20, with the same cache-hit price. The model guide says the rate does not change with prompt length.

Which Kimi K3.1 Pro scores are independent?

Three, all read from Artificial Analysis on 2026-10-24. The row name has no effort suffix. Humanity's Last Exam without tools is 52.46. Terminal-Bench 4.0 is 67.17. AA-AnalystAgent is 38.75 pass^5. Four more scores come from Moonshot's own launch chart, recorded as self-reported: HLE with tools 70.0, DeepSWE v1.1 75.5, GPQA Diamond 94.2, and OSWorld 2.1 79.6 — all at max thinking effort, each replaced when an independent board publishes. Three launch figures are not rows: Terminal-Bench 4.0 65.7 and HLE-no-tools 51.1 (those cells hold independent runs), and Agents' Last Exam (ALL) 43.2 — Moonshot's page says its ALE data comes from the official v1.1 release of 122 tasks, a different task set from the official leaderboard's Overall column.

How does Kimi K3.1 Pro compare with Kimi K3?

Artificial Analysis was re-read for both on 2026-10-24. K3.1's row has no effort suffix: HLE 52.46, Terminal-Bench 4.0 67.17, Analyst 38.75. K3's max row is HLE 46.90, Terminal-Bench 4.0 12.63, Analyst 38.75. Those gaps are not matched-effort calls. K3's stored Terminal-Bench 4.0 cell remains vals.ai's 17.17.

Is Kimi K3.1 Pro open source?

Yes, as of 2026-11-04. The weights are public at huggingface.co/moonshotai/Kimi-K3.1-Pro: 96 safetensors shards, about 1.56 TB, MXFP4 weights and MXFP8 activations. The Kimi K3.1-Pro License allows use, modification, redistribution and commercial use. It is not MIT or Apache. Two conditions sit on top. A Model-as-a-Service operator whose aggregate revenue exceeds $20 million in any consecutive 12 months must sign a separate agreement with Moonshot before commercial use. A commercial product with more than 100 million monthly active users, or more than $20 million of monthly revenue, must display "Kimi K3.1-Pro" on the interface. Internal use, and use through Moonshot's own products or certified partners, is exempt from both. This page's license field reads open-weights because the files are public.

Does Kimi K3.1 Pro replace Kimi K3?

Moonshot's homepage calls K3 the previous-generation flagship, and K3 is still listed with its own price. This site's previous-version badge does not attach, because the names land in different tiers: Kimi K3 has no tier word, and Kimi K3.1 Pro ends in Pro.

Further reading

Benchmark guides