Tech • AI • Robotics • Game
Kimi K3.1 Pro: a new top open-source model has just arrived
7/10TL;DR
Kimi K3.1 Pro has emerged as a strong open-weight AI model with benchmark results near Opus-class systems, better performance than several leading rivals on coding and agentic tasks, and pricing around $2 per million input tokens and $10 per million output tokens.
KEY POINTS
Benchmark leap
Kimi K3.1 Pro is described as a major step up from earlier Kimi, GLM, and DeepSeek generations, with roughly a 10-point gain over prior family models in the cited comparisons. Across the reviewed charts, it was positioned above Astra, DeepSeek, and other mainstream competitors on several intelligence measures, while trailing only top Opus configurations in some aggregate rankings.
Strong results on coding and agent tasks
On specialist evaluations such as Terminal Bench, SWE Marathon, and OS World, the model posted standout scores. It outperformed Astra and DeepSeek repeatedly and, in some cases, edged close to Fable and Opus, suggesting particular strength in code execution, task completion, and operating-system style control.
Pricing undercuts premium rivals
The model was presented as significantly cheaper than top-end premium systems while preserving high capability. On OpenRouter, the cited rate was $2 for input and $10 for output, roughly half the cost of Opus in the comparison used. Subscription tiers were also listed at $15 for Plus, $31 for Pro, $79 for Max, and $159 for Ultra.
Speed has improved sharply
Earlier Kimi versions were criticized for slow execution, but K3.1 Pro showed a clear gain in task time. It was described as operating at roughly the same speed as GPT-6.1 Sol Max and close to Opus Extra High in the cited tests. Throughput figures of about 37 to 57 tokens per second were highlighted as competitive for a model at this capability level.
Token efficiency stands out
A notable claim was that Kimi K3.1 Pro matched Opus Extra High on one intelligence graph while using about 10,000 fewer output tokens. That lower token consumption helps explain its stronger cost-performance profile even when raw pricing is no longer dramatically lower than discounted GPT options.
Mixed practical results in visual coding tests
In hands-on tasks, results were uneven but often strong. A car-crash simulation showed the model could build visually convincing scenes and functioning sites, but physics remained unreliable, with impacts and movement behaving unrealistically. Interestingly, the high reasoning mode sometimes outperformed max, reinforcing a recurring pattern that heavier reasoning can worsen creative or simulation-style outputs.
Better at 3D interfaces than at motion design
A lava lamp and other 3D-style interface tests produced persuasive visual results, including color controls and lighting changes. However, Opus remained clearly ahead in motion design, video generation, and editing taste. In side-by-side comparisons, Kimi delivered usable outputs, but Opus was judged more polished in layout, animation, and overall presentation.
Cost-quality ratio was competitive in real runs
Several direct runs were notably cheaper than premium alternatives. One test was cited at $1.74 for 25 minutes of work, while more expensive competing runs reached about $8.75. In longer video-generation comparisons, Kimi runs were listed at $16 to $27 and completed in 2.5 to 4 hours, versus roughly $44 and 7 hours for an Opus run.
Guardrails are tighter than before
The model no longer appears fully permissive on questionable requests. It refused to provide links to illegal streaming sites, declined to build a fake scarcity counter for email marketing, and initially resisted unauthorized penetration testing. However, the tests also suggested that simple prompting could still weaken those safeguards, indicating protections are present but not robust.
Long context may help enterprise-style work
The model offers a stated context window of 1 million tokens, which could make it attractive for large codebases, long documents, and administrative or workflow-heavy tasks. That combination of long context, solid speed, and strong coding ability positions it as a practical alternative to higher-priced proprietary systems, even if it does not fully surpass the best design-oriented models.
CONCLUSION
Kimi K3.1 Pro appears to be one of the strongest current options for coding, agent workflows, and large-context tasks at a relatively aggressive price. It does not displace Opus on design and video craft, but it raises the pressure on premium proprietary models by narrowing the performance gap at lower cost.
Ask a question