Moonshot AI, released Jan 1, 2026

Kimi K2.5price, context, benchmarks and release details

Provisional: not enough results to rank yet Open weights 100% confidence 100 percent, Full confidence, 6 of 8 expected sources in
59.0
SI Score
#66 of 142 ranked models
Input, per 1M tokens
$0.60LiteLLMFirst-party API Standard token rate; MIT LiteLLM transcription. Provider documentation: https://platform.moonshot.ai/docs/guide/kimi-k2-5-quickstart. Exact endpoint only; cache/batch/long-context rates excluded Retrieved Oct 9, 2026 · MIT
Open source ↗
Output, per 1M tokens
$3.00LiteLLMFirst-party API Standard token rate; MIT LiteLLM transcription. Provider documentation: https://platform.moonshot.ai/docs/guide/kimi-k2-5-quickstart. Exact endpoint only; cache/batch/long-context rates excluded Retrieved Oct 9, 2026 · MIT
Open source ↗
Context window
262Kmodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
Max output
262Kmodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
Released
Jan 1, 2026models.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗

How the score breaks down

Coding (weight 40 percent) 72.4
Math (weight 15 percent) —
Preference (weight 15 percent) 77.3
Reasoning (weight 30 percent) 37.3

Weights: reasoning 30%, math 15%, coding 40%, preference 15%. Results use fixed 0–100 scales before averaging, and thin evidence is pulled toward 50. Method si-v3-retained-evidence-2, computed Oct 9, 2026, 05:13 UTC.

Around it on the leaderboard

  1. 64 Claude Haiku 5.5 59.4
  2. 65 GLM-5 59.4
  3. 66 Kimi K2.5 59.0
  4. 67 Qwen3.7 Max 58.6
  5. 68 Gemini 3 Pro Preview 58.6

Full leaderboard

Benchmark results

7 benchmarks, 9 results

Each row shows the best published result. Where a model was tested at several settings, such as reasoning effort, open the row to see each one. Hover or tap a value for its source.

ARC-AGI-1 (public eval)reasoning 73.1%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. [variant] kimi-k2.5Published Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 73.1
ARC-AGI-1 (semi-private)reasoning 65.3%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. [variant] kimi-k2.5Published Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 65.3
ARC-AGI-2 (public eval)reasoning 12.1%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. [variant] kimi-k2.5Published Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 12.1
ARC-AGI-2 (semi-private)reasoning 11.8%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. [variant] kimi-k2.5Published Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 11.8
Humanity's Last Exam (Scale AI)reasoning 24.4%Humanity’s Last ExamPotential contamination warning: This model was evaluated after the public release of HLE, allowing model builder access to the prompts and solutions. [variant] Published Feb 13, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 24.4
SWE-bench Verifiedcoding 73.8%Epoch AI BenchmarkingEpoch-owned evaluation mean score; scores only, no benchmark questions [variant] Published Feb 17, 2026 Retrieved Oct 9, 2026 · CC-BY
Open source ↗ 73.8 3 settings
  • Setting 1 73.8%Epoch AI BenchmarkingEpoch-owned evaluation mean score; scores only, no benchmark questions [variant] Published Feb 17, 2026 Retrieved Oct 9, 2026 · CC-BY
    Open source ↗
  • Setting 2 70.8%Official model cards via models.devLab-reported; metric resolved; transcribed by MIT models.dev catalog; not independently evaluated [variant] Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 3 70.8%SWE-bench VerifiedSWE-bench published model plus agent result; harness retained, not a base model evaluation [variant] mini-SWE-agent; high; 2.0.0Published Feb 17, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
About SWE-bench Verified
LMArena Textpreference 1445.2 eloLMArena / ArenaPublished source fact [variant] text / overallPublished Oct 8, 2026 Retrieved Oct 9, 2026 · CC-BY-4.0
Open source ↗ 77.3

Normalization uses a fixed 0–100 scale for each unit, independent of other models. Compare evaluation conditions before reading a small gap as decisive. “Lab-reported” marks the provider's own published figure.

Details and sources

Open weights
YesHugging Face HubPublic Hub repo with weight files; gating/repo upload date is not release date Retrieved Oct 9, 2026 · factual metadata; model-specific licenses
Open source ↗
License
otherHugging Face HubPublished source fact Retrieved Oct 9, 2026 · factual metadata; model-specific licenses
Open source ↗
Input modalities
text, image, videomodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
First seen by SuperIndex
Oct 8, 2026
Coverage
83% of expected source weight

Reported (6)

  • ARC PrizeOct 8, 2026
  • Epoch AI BenchmarkingOct 8, 2026
  • Humanity’s Last ExamOct 8, 2026
  • Official model cards via models.devOct 8, 2026
  • LMArena / ArenaOct 8, 2026
  • SWE-bench VerifiedOct 8, 2026

Awaiting (2)

  • LiveBench11% of weight
  • Terminal-Bench6% of weight

Confidence rises as pending sources publish. Some sources never cover some models, so confidence reaches 100% at 80% of expected weight.

What changed

All releases · RSS feed

Alerts on this device

What to be alerted about
RSS feed