Vals reports Kimi K3.1 Pro runs Terminal-Bench 4 about 4× faster than K3

• Terminal-Bench 4 got substantially cheaper and faster: Vals AI reports roughly 4× faster completion, half the output tokens and about one-third the cost versus K3.
• Gains were uneven: legal research and tax scores fell in Vals AI's tests, with legal research consuming roughly 8× the output tokens.
• Agent Arena placed K3.1 Pro eleventh across more than 4,000 sessions, with a $1.03 launch median task cost, according to Arena's announcement.
• Hosted access shipped ahead of downloadable weights: the launch details promise full weights by November 4, 2026.

Moonshot used Claude Code for most of its own coding evaluations. Its launch post describes K3.1 Pro as a post-trained K3, plus a separate Highspeed variant promising 2–3× faster output generation.

K3.1 Pro reached 62.52% on the Vals Index, up from K3's 52.95%, moving from twentieth to eighth. Almost all that improvement came from agentic coding, according to Vals.

The Terminal-Bench 4 comparison breaks down as follows:

• Accuracy: 17.2% → 63.6%, a 46.4-point gain, placing K3.1 Pro third overall.
• Cost per test: $12.48 → $4.57, approximately 63% lower.
• Completion time: roughly 4× faster.
• Output tokens: approximately half as many.

Code Migration also moved from near the bottom of the index into the top ten.

More solved tasks on half the output tokens is a proper upgrade. Opus 5.5 scored 65.2% at $13.20 per test in Vals' Terminal-Bench comparison.

Vals reported max reasoning effort, temperature 1, top-p 0.95 and a 262K output-token cap for these runs. Its detailed model results list Moonshot AI as the default provider, while noting that individual benchmarks can use different settings.

Scores fell on legal research and tax tasks in Vals AI's evaluation:

• Legal Research: lower accuracy despite roughly 8× K3's output tokens, mostly reasoning.
• Tax Agent: lower accuracy than K3.
• Finance Agent: improved accuracy, but approximately 3× the output tokens.
• EMB: weaker results, particularly on comparable-company analysis and leveraged-buyout modeling, leaving K3.1 Pro below K3 on the overall Finance Index.

Moonshot reports improvements across its own financial modeling, research and business-operation tests in the launch post. The external finance results are considerably less uniform.

K3.1 Pro (Max) rose from K3's twenty-fifth place to eleventh, with net improvement increasing from +1.6% to +7.6% across more than 4,000 sessions, according to Arena's launch results. It ranked eighth in Work, ninth in Code and tenth in Chat.

Its individual signals spread much wider:

• Praise vs. Complaint: +19%, eighth.
• Steerability: +8.3%, eleventh.
• Bash Reliability: +8.3%, fifteenth.
• Confirmed Success: +2%, twenty-sixth.

Confirmed Success measures users affirming that a task is done. The live leaderboard lists that signal at +2.03% ±5.86%, alongside net improvement of +7.61% ±3.00%.

Arena's announcement quoted a $1.03 median task cost, 13% below K3, and placed the model just off its price-performance Pareto frontier. The linked leaderboard displays a different median, $1.18.

A pagoda-garden build took five minutes and about $0.55, running 2–3× faster than K3, according to cedric_chee. He described its design taste as closer to Astra-class models in the same test thread.

A heavier Blender task produced a different cost profile: 73 minutes and $9.04, more expensive than GPT-6.1 Sol, in cedric_chee's separate test. He judged K3.1 slightly better than K3 at that task, while still behind the current frontier.

Both Pro and Highspeed are available through Kimi's website, apps, Work, Code and API, according to the launch details. The model has a 1M-token context window; full weights are promised by November 4.

On OpenRouter, standard pricing is $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens.

Moonshot's availability notes list the API IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise access is off by default at launch, and Zero Data Retention is available for API customers.