ModelsKimi K3.1 Pro
Kimi K3.1 Pro
Moonshot AI’s further post-trained Kimi K3, with independent Artificial Analysis measurements and a complete publisher comparison covering coding, agentic work, reasoning and vision.
Model specifications
- Context window
- 1,048,576 tokens
- Maximum output
- 1,048,576 tokens
- Input / output price
- $2.00 / $10.00USD / million tokens · check tiers below
- AA Intelligence Index
- 55.09Evaluator composite, on its own scale
Test coverage
- 14aa
- 45self
- 106 no result
165 catalogue tests. Each test counts once; evaluator results take precedence where sources overlap.
tests with a result
source-specific results
Evaluation results
Filter the observations. Expand a row to inspect its run conditions.
73 results
| Benchmark | Result | Source | Run |
|---|---|---|---|
| AA-AnalystAgent | 38.75% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| AutomationBench-AA | 73.30% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 73.3% | self | DetailsMoonshot quotes Artificial Analysis. Rounded publisher values remain separate from direct AA measurements. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| SciCode | 60.76% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| GDP.pdf | 31.40% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 34.6% | self | DetailsNo Python: PDF pages resized to 1,152-pixel longest edge and interleaved with Docling text. Gemini-3.5-Flash medium grades official rubrics; strict all-pass rate among fully scored responses, averaged over five runs. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| CritPt | 24.57% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| AA-Omniscience | 29.52 score | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| AA-LCR v1.1 | 86.33% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| Terminal-Bench 4.0 | 67.17% | aa | DetailsAA mini-swe-agent results, highest published effort per release. Grouped here from the published configurations; ranks are derived, not publisher-issued positions. Vals runs remain in a separate table.
|
| 65.7% | self | DetailsClaude Code for Kimi. Comparison figures are the maximum cited by Moonshot from model reports or tbench.ai, so fallback and harness settings differ. The separate cost chart is not this headline comparison. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| 52.8%cost chart, high | self | DetailsExact labelled chart coordinates at high effort. These can differ from the release table, which takes the maximum of reports and leaderboard results. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison.
| |
| 65.7%cost chart, max | self | DetailsExact labelled chart coordinates at max effort. These can differ from the release table, which takes the maximum of reports and leaderboard results. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison.
| |
| MMMU Pro | 83.87% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 84.1% | self | DetailsMMMU-Pro without Python, official input order with images before text; mean over four runs. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| FrontierCode v1.1 | 47.07%Main | aa | DetailsAA public FrontierCode results, highest published effort per release. Grouping ranks are derived here; model-publisher reports retain separate harnesses and configurations.
|
| GDPval-AA v2.1 | 1,845.72 Elo | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 1,846 Elo | self | DetailsGDPval-AA v2.1 Elo quoted by Moonshot from Artificial Analysis; not a percentage. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| AA-Briefcase v1.1 | 1,677.79 Elo | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 1,678 Elo | self | DetailsAA-Briefcase v1.1 Elo quoted by Moonshot from Artificial Analysis; not a percentage. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| Humanity's Last Exam | 52.46% | aa | DetailsHighest published effort per model release, grouped here from AA’s public results. Grouping ranks are derived here; source settings and model-publisher reports stay separate.
|
| 51.1%HLE-Full | self | DetailsFull HLE without tools, 2,500 questions. GLM’s starred 43.0 is text-only and belongs to a separate subset row, not this full-set field. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
| |
| Terminal-Bench Science 0.1 | 38.1% | self | DetailsKimi uses Claude Code. The 15.7 printed under DeepSeek 1017 is explicitly attributed in footnote 2 to the September DeepSeek V4.1 Flash. It is recorded under that older model identity. Moonshot also discloses its 1017 internal score of 9.3 separately. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| SWE-Marathon | 51.7%SWE Marathon v1.1 | self | DetailsSWE-Marathon v1.1; Kimi and DeepSeek use Claude Code. Other results are quoted from the official leaderboard. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| BenchCAD w/o tools | 72.8 IoU × 100 | self | DetailsVoxel IoU expressed on a 0–100 scale, not percent-correct accuracy. Open models use a random 1,000-case subset of the 17,900-case Vision2Code dataset; single-shot official prompts, max effort. GPT has no published no-tools result. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| BenchCAD w/ tools | 92.7 IoU × 100 | self | DetailsVoxel IoU on a 0–100 scale, not accuracy. The open-model 1,000-case Vision2Code subset uses Claude Code and tools. Proprietary scores are quoted from their reports; subset differences remain a limitation. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| ProgramBench (binary threshold 95%) | 31.5%Program Bench | self | DetailsBinary task success: 1 only when the underlying ProgramBench score is at least 95%, otherwise 0. Kimi uses Claude Code. This is distinct from hidden-test pass rate and fully resolved task accuracy. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| DeepSWE v1.1 | 75.5% | self | DetailsKimi uses Claude Code; comparison scores come from each model technical report and can use other harnesses. Datacurve independent runs stay separate. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| PostTrain Bench v1.2 | 36.5% | self | DetailsKimi and DeepSeek were run internally on v1.2 with Claude Code and H20 GPUs rather than H100s, using DeepSeek 1017 max as judge in place of the official GPT-5.6-Terra xhigh judge. Other scores quote the official board. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| Kimi Code Bench 2.0 (Internal) | 77.1% | self | DetailsInternal Kimi Code Bench 2.0. Kimi uses Claude Code; missing proprietary-model cells are unreported. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| Kimi Code Bench 3.0 Long-horizon (Internal) | 61.2% | self | DetailsInternal Kimi Code Bench 3.0 Long-horizon; Kimi uses Claude Code, all other models max effort, with possible proprietary-model fallback. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| 61.2%Kimi Code Bench 3.0 Long-horizon (predecessor) | self | DetailsExact figures stated in the release prose, including K3’s two decimal places. Chart labels and explicit release prose, read 24 October 2026. These are internal publisher evaluations. The finance-chart Business Operation value differs from the full-table value and retains its own row. No Highspeed scores are inferred.
| |
| HLE-Full w/ tools | 70.0% | self | DetailsFull 2,500-question HLE with tools, separate from no-tools and text-only results. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| BrowseComp (context management) | 92.1%BrowseComp | self | DetailsContext compaction at 200K tokens, maximum 500 steps followed by a forced answer; temperature 1.0 and top-p 1.0. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| BioMysteryBench | 80.6% | self | DetailsOfficial v11 dataset with Kimi K3 as judge. Kimi uses Kimi Code, GLM and DeepSeek Claude Code, GPT Codex; Claude results quote system cards. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| CharXiv (RQ) w/ python | 93.4% | self | DetailsPython tools; four-run mean. Reasoning-question subset of CharXiv. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| MMMU-Pro w/ python | 90.0% | self | DetailsPython tools; four-run mean. Original input order is retained, with images prepended to text. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| GDP.pdf w/ python | 85.7% | self | DetailsPython receives the original PDF in an IPython sandbox. Five-run task-macro rubric pass rate, judged by Gemini-3.5-Flash medium; this is not strict all-pass GDP.pdf. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| ChartoGraphy w/ python | 90.2% | self | DetailsPython tools, mean over five runs; separate from no-tools ChartoGraphy. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
| BabyVision w/ python | 95.4% | self | DetailsPython tools, mean over four runs. Read 24 October 2026. All 47 source rows and six comparison columns are retained, with text-only HLE and both ZeroBench metrics split explicitly. Kimi uses max reasoning effort, temperature 1.0 and top-p 0.95 unless a row states otherwise; proprietary comparison columns are labelled max with fallback. Comparison figures include quoted external results and mixed harnesses. Missing cells remain null and do not enter ranks. Lower hallucination rates rank first; every other rank compares larger values within the same row. Terminal-Bench Science’s September DeepSeek score is attributed separately according to footnote 2. These publication claims do not populate independent boards. Highspeed has no independently identified results and inherits none of these scores.
|
API, pricing & runtime
Published prompt tiers, configuration notes, and endpoint measurements.
| Prompt tier | Input | Output | Cache read | Write 5m | Write 1h |
|---|---|---|---|---|---|
| All prompt lengths | $2.00 | $10.00 | $0.20 | $2.00 | $4.00 |
Release, availability and weights
Released 23 October 2026 across Kimi web/mobile, Kimi Work, Kimi Code and the API. Full weights are promised by 4 November 2026 and were not public in the sources checked on 24 October; the licence is not yet published. Enterprise administrators must opt in at launch. Kimi Work requires version 3.1.0 or later.
Read source ↗API identity and Highspeed
API identifier kimi-k3.1-pro. Highspeed uses kimi-k3.1-pro-highspeed, with its own catalogue entry and pricing. The official 2–3× speedup claim does not establish measured Highspeed performance.
Read source ↗Context and completion limits
The official K3 API supports a 1,048,576-token shared context window. max_completion_tokens defaults to 131,072 and can be set up to 1,048,576, subject to the combined input/output budget; this is a request upper bound, not an additional output window. OpenRouter’s completion cap is recorded separately under Hosted runtime.
Read source ↗Architecture and modalities
Moonshot calls this a 3T-class post-trained Kimi K3 model with native visual understanding. Artificial Analysis lists 2.8T total parameters and 104B active; these exact counts are evaluator metadata, not a new official architecture disclosure. The API guide supports text, image and video input with text output; no audio input/output is documented.
Read source ↗Thinking and sampling
Always-on thinking with reasoning_effort low, high or max (default max). Temperature 1.0, top_p 0.95, n=1, presence_penalty=0 and frequency_penalty=0 are fixed and should be omitted. Return complete assistant messages unchanged in multi-turn and tool workflows.
Read source ↗Capabilities and integration
Streaming, tool calling, forced tool choice, dynamic tool loading, JSON/strict JSON Schema and partial/prefix completion are documented. Chat Completions, Responses and Anthropic Messages-compatible APIs are available. Vision uses base64 or ms:// file identifiers; public image URLs are unsupported. Built-in web search is being updated and is not currently recommended for production. Zero Data Retention is available to API customers.
Read source ↗USD pricing and caching
USD per million tokens, excluding tax: input $2.00, output $10.00, cached input $0.20, cache writes $2.00 for 5m and $4.00 for 1h. Flat rates across context lengths; automatic repeated-prefix caching defaults to a 5m TTL. Writes are billed separately; hits refresh the TTL.
Read source ↗Independent measurements
Artificial Analysis reports an Intelligence Index of 55.0869210959395 (55.09 displayed), weighted index cost $2.1145/task and 608.83 seconds/task. Independent Terminal-Bench 4.0 is 67.1717%, distinct from Moonshot’s 65.7 Claude Code report. AA publishes 13 task metrics plus the index in this snapshot. Vals’ public model directory and Datacurve’s DeepSWE v1.1 artifact contain no separately identified K3.1 Pro result.
Read source ↗Independent timing and output usage
AA’s Kimi API long-prompt measurements: 77.09 output tokens/s; first streamed chunk 2.888s, first answer token 28.831s including thinking, and a 500-token end-to-end response 35.317s. These timings differ from the index task duration. Index-task output averages 55,194.60 tokens, including 34,882.70 reasoning and 20,311.90 answer tokens.
Read source ↗Additional evaluator statistics
AA-Briefcase overall Elo 1677.79 (95% CI 1667.35–1688.23), analytical quality 1995.07, presentation 1665.70; rubric pass rate 50.40%. GDPval-AA Elo 1845.72 (interval 1820–1872). AA 3D Elo 1110.16 (interval 1078–1143). AA-Omniscience accuracy 51.02% and hallucination rate 43.89% use AA’s own definitions, distinct from Kimi Hallu Bench. Capability indices: finance/accounting 56.52, strategy/operations 59.39, legal 55.84, engineering 55.86 and economics 60.03.
Read source ↗Publisher evaluation scope
The complete 47-row, six-model comparison retains separate tool settings, HLE subsets, ZeroBench avg@5/pass@5, and lower-is-better hallucination rates. Kimi uses max effort unless stated; proprietary columns are labelled max with fallback. Internal benchmarks, quoted external results and modified judges remain publisher claims. A Science footnote attributes DeepSeek’s 15.7 to the September model and discloses 9.3 for 1017.
Read source ↗Publisher chart supplements
The Terminal-Bench 4.0 cost chart reports Kimi high at 52.8% and $4.16/task, max at 65.7% and $5.25/task. Kimi Code Bench 3.0 Long-horizon improves from K3’s 55.95 to 61.2. The internal Hard hallucination rate improves from 23.4% to 8.3%; responses with an error are 42.4% of the prior rate, not a 42.4% absolute error rate. Internal stress tests do not describe everyday accuracy.
Read source ↗Cybersecurity evaluation
The final checkpoint scores 81.0% pass@3 on internal xdev-bench-hard, versus 90.5% for a checkpoint without cyber-data filtering; the unreleased checkpoint is not a selectable model. The figure reports Kimi’s average success rate 64.5%, with 159.8M input and 546K output tokens per successful trajectory. Kimi uses Kimi Code, without a hard time or output-token limit.
Read source ↗Documentation read 24 October 2026.
Hosted runtime: OpenRouter
- Context
- 1,048,576 tokens
- Modalities
- text, image, video
- Capabilities
- vision, tools, thinking, cloud
- Parameters / active
- Not published / Not published
| Rate | Input | Cached input | Output |
|---|---|---|---|
| Peak / standard | $2.00 | $0.20 | $10.00 |
| Off-peak | Not reported | Not reported | Not reported |
OpenRouter advertises text, image and video input with text output. Its public catalogue lists a 943,718-token maximum completion, separate from the official API’s upper parameter limit. Model identity: moonshotai/kimi-k3.1-pro-20261023. Host capabilities do not establish a separate benchmark result.
Runtime source · read 24 October 2026 ↗AA API performance and evaluation cost
- Output speed
- 77.0917244477023 tokens/s
- First answer token
- 28.831169042873917 seconds
- 500-token response
- 35.316949502217405 seconds
- Index cost per task
- $2.1145052918767275
AA measures the Kimi API. Long-prompt medians; first answer token includes input processing and thinking. First streamed chunk is 2.888 seconds and is a different metric. Index cost includes cache and reasoning tokens. No separately published effort label is attached to this AA record.
Evaluator source ↗Original sources
- Moonshot release announcement and full evaluation table ↗Where this source is used →
- Kimi K3.1 Pro API guide and limits ↗Where this source is used →
- Official USD API and cache pricing ↗Where this source is used →
- Artificial Analysis independent model evaluation ↗Where this source is used →
- OpenRouter hosted model catalogue ↗Where this source is used →