Updated
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI's flagship language model and the first model in the K3.1 series, released on October 23, 2026. It is not a new architecture: Moonshot took Kimi K3, a 2.8-trillion-parameter sparse mixture-of-experts model with about 104B active parameters per token, and spent substantially more post-training compute on it, most of it reinforcement learning. The result lands within a point of Claude Opus 5.5 on Terminal-Bench 4.0 and cuts the hallucination rate of K3 by almost two thirds on Moonshot's own stress test. On UnificAlly it runs as moonshotai/kimi-k3.1-pro with a 1,048,576 token context, 131,072 output tokens by default, and thinking on every request.
Kimi K3.1 Pro key features
65.7% on Terminal-Bench 4.0
Moonshot reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5, 58.2% for GPT-6 Astra, and 42.3% for DeepSeek V4.1 Flash. Kimi Code Bench 3.0, Moonshot's long-horizon coding suite, rises from 55.95% on K3 to 61.2%. The vendor ran K3.1 Pro in its own harness, so read the gap to other labs as a direction, not a precise margin.
55 on the Artificial Analysis Intelligence Index
The independent Artificial Analysis index scores K3.1 Pro at 55, #4 of 215 models as of October 25, 2026. Only Claude Opus 5.5 at 58 sits above it, and it edges Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon, all at 53. Output speed is 73.5 tokens per second, below the median of about 82.
Hallucination rate cut from 23.4% to 8.3%
On Kimi Hallu Bench Hard, Moonshot's claim-level stress test with tools, K3.1 Pro produces false claims 8.3% of the time against 23.4% for K3, 15.0% for Claude Opus 5.5, and 5.5% for GPT-6 Astra. Without tools it lands at 5.4%, between GPT-6 Astra at 4.2% and Claude Opus 5.5 at 6.8%.
Finance and office work
Moonshot reports 64.4 on OfficeQA Pro v2 and 58.6 on Frontier Finance, and its internal finance suites move from 50.4 to 57.6 on financial modeling, 58.8 to 70.4 on financial research, and 8.9 to 37 on business operation. BenchCAD, a 3D CAD generation test, lands at 92.7 with tools, level with Claude Fable 5.1 at 92.6.
1M token context, 131k default output
The window holds a large repo, a long filing pack, or a full agent transcript at once, at one flat rate across the whole window. A single response defaults to 131,072 tokens and can be raised to the full 1M. Context caching is automatic, with a 5-minute default TTL and a 1-hour option.
Full agent and tool surface
Tool definitions, tool_choice including required, dynamic tool loading, strict json_schema structured output, and partial mode all run through the same endpoint. Streaming returns reasoning and answer text as separate deltas, so you can show thinking without parsing it out of the response.
What Kimi K3.1 Pro is best for
Long coding sessions
Multi-hour work across a repo where the model has to hold context and keep going without supervision.
Agentic web research
Multi-step search and synthesis with a low false-claim rate when the model has tools in hand.
Finance and document work
Filings, models, and spreadsheets read in one pass and returned as a brief, where the K3.1 gains are largest.
Tool-heavy agents
Strict tool choice and dynamic tool loading, for agents that carry a large tool surface.
Single-artifact generation
Reports, migrations, and drafts that need one very long response instead of stitched calls.
Kimi K3.1 Pro use cases
Build a coding agent that plans a migration, edits files across a repo, runs tests, and keeps going across a long session without restarting. Use it as the research layer in a product that has to plan queries, read what comes back, and synthesise an answer with sources, where the lower hallucination rate matters most. Point it at financial filings, spreadsheets, and office documents, the area Moonshot spent the most post-training on. For teams already running an OpenAI-compatible agent loop on Kimi K3, K3.1 Pro is a model id change.
Kimi K3.1 Pro limitations
Thinking is always on and reasoning_effort defaults to max. A request that leaves the field out pays full reasoning cost, so simple chat and extraction carry latency and token spend they do not need unless you set low. Changing the effort between turns invalidates prefix-cache hits.
Speed is a real constraint. Artificial Analysis measures 73.5 output tokens per second, below the median for its class, and the model is verbose: about 150M output tokens across the index against a median of 87M. The Highspeed serving exists for exactly this reason.
Sampling is fixed upstream. temperature is 1.0 and top_p is 0.95, and a request that sets other values is rejected, so steer the model with the system prompt and a JSON schema.
The weights are a promise, not a download. Moonshot says they arrive by November 4, 2026, with the licence still unannounced. Moonshot also removed cyber-related data from training before the open release; on its own exploit test the shipped model scores 81.0% where an unfiltered checkpoint scored 90.5%.
Most of the benchmark table is Moonshot's own work: Kimi Code Bench and the finance suites are Moonshot-owned, and the coding rows were run in Moonshot's harness while competitor rows come from other labs. Treat the self-reported numbers as unconfirmed until independent reruns land.
Kimi K3.1 Pro vs Kimi K3
Same architecture, same 1M window, same 131k default output, better model. Terminal-Bench 4.0 is new in the K3.1 table, Kimi Code Bench 3.0 rises from 55.95% to 61.2%, and the claim-level hallucination rate falls from 23.4% to 8.3%. The finance suites move the most, with business operation going from 8.9 to 37. K3.1 Pro is also cheaper per token on UnificAlly. The one thing K3 still has is open weights you can download today; K3.1 Pro's arrive by November 4. For a running K3 deployment, switch the model id and keep the rest.
When to use Kimi K3.1 Pro
Use Kimi K3.1 Pro for work that runs long and mostly unattended: repo-scale coding, multi-step research, and finance or document passes that would otherwise be split across many prompts. It is a poor fit for chat and anything latency-sensitive, because it thinks at full effort by default and generates slower than most models in its class. For interactive use, call moonshotai/kimi-k3.1-pro-highspeed; for the strongest all-round model, Claude Opus 5.5 still leads the index.
Kimi K3.1 Pro API examples
Call Kimi K3.1 Pro by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Parameter docs live at docs.unifically.com/models/llm/moonshotai/kimi-k3.1-pro.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "moonshotai/kimi-k3.1-pro",
"reasoning_effort": "high",
"messages": [
{ "role": "user", "content": "Plan a migration from Express to Fastify for this service and list the risks in order." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate; thinking and answer text arrive as separate deltas. Drop reasoning_effort to low for quick steps.
FAQs
People also ask about Kimi K3.1 Pro
moonshotai/kimi-k3.1-pro, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key. The faster serving of the same weights is moonshotai/kimi-k3.1-pro-highspeed.
On UnificAlly, $1.50 per 1M input tokens, $7.50 per 1M output tokens, and $0.15 per 1M cached input tokens. Context caching is automatic, and reasoning tokens bill as output. Live rates are on the pricing page.
K3.1 Pro is Kimi K3 with much more post-training compute, mostly reinforcement learning, on the same 2.8T-parameter architecture. Moonshot reports Terminal-Bench 4.0 at 65.7%, Kimi Code Bench 3.0 up from 55.95% to 61.2%, and its claim-level hallucination rate down from 23.4% to 8.3%. The 1M-token window and the 131k default output are unchanged.
Moonshot reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, and 92.7 on BenchCAD with tools. The independent Artificial Analysis Intelligence Index puts it at 55, #4 of 215 models as of October 25, 2026, behind only Claude Opus 5.5 at 58 and ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon at 53.
A 1,048,576 token context window at one flat rate, with no long-context surcharge. A single response defaults to 131,072 output tokens and can be raised to the full 1M.
No. Thinking runs on every request. reasoning_effort takes low, high, and max, and max is the default when the field is left out, so set low for quick steps. Changing the effort mid-conversation invalidates prefix-cache hits.
Not yet. Moonshot says the full weights will be published by November 4, 2026. The licence has not been announced. Kimi K3 shipped under the Kimi K3 License, which allows commercial use.
Yes. Tool definitions, tool_choice including required, dynamic tool loading through a system message, strict json_schema structured output, and partial mode all work through the standard chat completions request. Streaming returns reasoning and answer text as separate deltas.
No. Artificial Analysis measures 73.5 output tokens per second, below the median for its class, and the model is verbose, using about 150M output tokens across the index against a median of 87M. For interactive work, call moonshotai/kimi-k3.1-pro-highspeed, which serves the same weights two to three times faster.
temperature, top_p, n, presence_penalty, and frequency_penalty are fixed upstream (temperature 1.0, top_p 0.95), so leave them out. Steer the model with the system prompt, reasoning_effort, and a JSON schema instead.
