◆ THE LAB · Models
Moonshot's Kimi K3.1 Pro ranks third on Vals' coding test
The News
Moonshot AI released Kimi K3.1 Pro, a post-trained version of Kimi K3, on October 23, 2026, across its Kimi products, including the API, alongside a Highspeed variant it says runs 2–3 times faster. Moonshot charges $2 per million input tokens and $10 per million output, and says full weights will be released by November 4.
Chandni Melwani
Founder & Editor
Oct 29, 2026 · 1 MIN READ
Know More
- — Highspeed: $4 in and $20 out per million tokens; cached input $0.20, the same as standard.
- — Context: 1,048,576 tokens; output defaults to 131,072 and can be set up to 1,048,576.
- — Reasoning effort: low, high or max, default max; thinking can't be turned off.
- — Vals Index: 62.52%, ninth of 29 models, at $6.44 a test.
Moonshot’s price list puts standard K3.1 Pro a third below Kimi K3, which costs $3 in and $15 out per million tokens.
Vals.ai ran K3.1 Pro through Terminal-Bench 4.0, the coding measure in New in AI’s model comparison: 66 long tasks an AI agent does in a terminal. It scores 63.64%, third of 57, at $4.57 a test. Claude Opus 5.5, priced at $4 in and $20 out, leads at 65.15% and $13.20; Vals ran Opus with older Claude models as server-side fallbacks for refusals and lists 22 of its 198 attempts as fallbacks. K3 scores 17.17% at $12.48.
LiveBench’s reasoning suite, the table’s second measure, puts K3.1 Pro at 91.1 to K3’s 90.7, Opus 5.5’s 92.2 and GPT-6 Astra’s 92.7. Artificial Analysis gives it 46 on version 5.0 of its Intelligence Index, against 58 for Opus 5.5 at max with fallback, and measures 41 tokens a second on Kimi’s API, which it calls notably slow.
The Read
Moonshot says it post-trained K3 with 'substantially larger post-training compute', and the gain shows more on Vals' terminal tasks than on LiveBench reasoning. Artificial Analysis lists it as proprietary, saying its weights aren't publicly available; if they ship, its 46 would probably top the board's open-weights models, whose top score is 32.
Room for Disagreement
The Vals Index and LiveBench coding don't show Artificial Analysis's 14-point gap. Vals puts K3.1 Pro ahead of DeepSeek V4.1 Flash 1017 on the Vals Index by 62.52% to 61.26%, and LiveBench scores DeepSeek's coding 96.0 to K3.1 Pro's 81.8, so the open-weights lead depends on the board.
Related
Frequently Asked Questions
Why does Moonshot report a higher Terminal-Bench 4.0 score than Vals?
Moonshot ran its own test. It reports 65.7% for K3.1 Pro, evaluated with the Claude Code harness at max reasoning effort, and takes most rivals' scores as the maximum reported in their blog posts or on the official Terminal-Bench 4.0 leaderboard, putting Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%. Vals ran K3.1 Pro itself, through Moonshot AI's API at max effort, with mini-swe-agent, the minimal, model-agnostic harness it uses for every model.
Did K3.1 Pro improve on Kimi K3 across the board?
Not on Vals' legal and tax tests. Vals' October 24 update lists it below K3 on Legal Research Bench, 37.50% to 46.15%, Tax Agent Bench, 24.87% to 33.37%, and HLAB-Vals-1.1, 11.67% to 12.92%. On LiveBench's research suite it scores 77.9 to K3's 79.0. Moonshot says K3.1 Pro 'achieves consistent gains in every domain'.
What has Moonshot said about cyber risk in the open weights?
Moonshot says it removed cyber-related data from both mid-training and reinforcement learning for the open-weight release. On xdev-bench-hard, its in-house test of vulnerability discovery and exploitation, it says a checkpoint without those measures solved 90.5% of problems at pass@3 and the final K3.1 Pro 81.0%. It says K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months.
Sources
- Moonshot AI, “Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence” (checked October 28, 2026): K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute, the Highspeed variant at 2-3× the speed, availability, the weights promised by Nov 4, 2026, Moonshot's own Terminal-Bench 4.0 figures and method, the claim of consistent gains in every domain, and the cyber-data removal and xdev-bench-hard results
- Kimi API Platform, “Kimi K3.1 Pro” quickstart (checked October 28, 2026): reasoning effort low, high and max (default max), thinking always on, and max_completion_tokens defaulting to 131072 and settable up to 1048576
- Kimi API Platform, “Model Inference Pricing” (checked October 28, 2026): per million tokens, kimi-k3.1-pro at $2.00 input, $10.00 output and $0.20 cached input; kimi-k3.1-pro-highspeed at $4.00, $20.00 and $0.20; kimi-k3 at $3.00, $15.00 and $0.30; and all K3 series models support a 1,048,576-token context window
- Vals.ai, Kimi K3.1 Pro model page (checked October 28, 2026): the Vals Index at 62.52%, ninth of 29, at $6.44 per test, and the October 24 update on DeepSeek V4.1 Flash 1017's 61.26%, the legal and tax results against Kimi K3, and the run through Moonshot AI's API at max effort
- Vals.ai, “Terminal-Bench 4.0” leaderboard (updated October 27, 2026; checked October 28): 66 tasks, three runs per model with mini-swe-agent, Kimi K3.1 Pro third of 57 at 63.64% and $4.57 per test, Claude Opus 5.5 first at 65.15% and $13.20, and Kimi K3 at 17.17% and $12.48
- Vals.ai, Claude Opus 5.5 model page (checked October 28, 2026): $4.00 input and $20.00 output per million tokens, Claude Opus 5 and Claude Opus 4.8 as server-side fallbacks for refusals, and 22 Terminal-Bench 4.0 fallbacks (11.11%)
- LiveBench 2.0, October 25, 2026 results table (checked October 28, 2026): reasoning and research scores for Kimi K3.1 Pro and Kimi K3, reasoning scores for Claude Opus 5.5 and GPT-6 Astra, and coding scores for Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017
- Artificial Analysis, Kimi K3.1 Pro (checked October 28, 2026): an Intelligence Index v5.0 score of 46, 41.2 output tokens a second based on Kimi's API, “notably slow”, and its listing as proprietary because the weights are not publicly available
- Artificial Analysis, LLM leaderboard (checked October 28, 2026): Claude Opus 5.5 (max with fallback) at 58, and DeepSeek V4.1 Flash 1017 (Max) as the highest-ranked open-weights model at 32
- New in AI, model comparison table (checked October 28, 2026): the leaderboards it uses for coding, reasoning and speed
Chandni Melwani
Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.
Follow ↗The Lab Briefing
Get the five-beat briefing in your inbox.
The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.
Free · No spam · Unsubscribe in one click
More from The Lab
Models
Reflection's Beam is faster than every model on our table
Artificial Analysis measures Reflection's open-weight Beam at a median 327 tokens a second across its hosts, faster than any model on New in AI's comparison table, but LiveBench scores its reasoning 80.2, below all nine models on that table. Nebius, Beam's cheapest OpenRouter host and the slowest of the five Artificial Analysis tested, charges $0.25 per million input tokens and $1.00 per million output.
Chandni Melwani · US · Oct 23, 2026
Research
Anthropic's Claude Mythos 5 designs cheaper quantum circuits
Anthropic says Claude Mythos 5, working in a loop with an evaluator program, found more efficient quantum circuits. In seven of eight tasks, including chemistry and optimization, the accompanying paper reports both fewer Toffoli gates and fewer qubits than the baseline it compares each against.
Chandni Melwani · Global · Oct 22, 2026
Labs
Anthropic is turning away distillers' paying traffic
Anthropic says early evidence suggests its new defences have cut the data harvested by distillers, who covertly extract Claude's capabilities to replicate them in other models, by around 90 to 99%, and that it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
Chandni Melwani · Global · Oct 21, 2026