Kimi K3.1 Pro is Moonshot AI’s new flagship model and the best value pick for teams that run coding agents every day. It scores 65.7% on Terminal Bench 4.0 and starts at $2 per million input tokens. That puts it close to Claude Opus 5.5 on terminal coding.
Moonshot also removed cyber related data from training before releasing the full model weights. The cut is real but modest. Its own test shows a drop from 90.5% to 81%.
Below you will see why Moonshot did it and how the model compares on benchmarks and cost.
Why Moonshot cut hacking skills before open weighting
Moonshot says its vulnerability research scores kept rising as post training scaled up. It traced the gain to cyber related data hidden in the general mid training corpus. Cyber tasks also sat inside its reinforcement learning environments.
So the team removed that material from both stages and kept the general coding tasks. The stated goal is to avoid pushing the absolute frontier of cyber capability into an open weight model.
The concern is easy to see. Open models can’t be recalled once downloaded. Safety researchers also warn that safeguards on open models lag behind closed ones. The full model weights are due by November 4.
Outside pressure plays a part too. China proposed export controls last week on AI driven malware tools. The White House accused Moonshot in July of using banned Nvidia chips and copying an Anthropic model but published no evidence. US labs have also tightened distillation safeguards.
Open releases are getting attention across the industry. Our look at Hugging Face Carbon-X covers another one from this week.
Kimi K3.1 Pro vs Claude Opus 5.5
| Measure | Kimi K3.1 Pro | Claude Opus 5.5 | GPT 6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% |
| BenchCAD (with tools) | 92.7% | 96.2% | 95.9% |
| Terminal-Bench Science | 38.1 | 63.3 | 68.1 |
| Kimi Hallu Bench Hard (lower is better) | 8.3% | 15.0% | 5.5% |
| Weights | Open (due Nov 4) | Closed | Closed |
Kimi K3.1 Pro sits within a point of Opus 5.5 on terminal coding and ahead of GPT 6 Astra. It falls far behind on science terminal work. Artificial Analysis ranks it fourth of 215 models with an index of 55 against 58 for Opus 5.5.
Moonshot ran most coding tests inside Claude Code. Treat the chart as a guide and not a verdict.
Is the nerf real? 90.5% vs 81%
Moonshot tested xdev bench hard. It asks an agent to build a working exploit from real software bugs and scores success within three tries.
- Build without risk reduction measures: 90.5%
- Final Kimi K3.1 Pro: 81%
- DeepSeek V4.1 Flash: 85.7%
- Kimi K3: 33.3%
The measures cost 9.5 points. K3.1 Pro still scores more than double K3 because stronger general coding carries over to cyber tasks. Calling it a nerf oversells it. Moonshot says its cyber skill stays behind frontier proprietary models.
Three limits matter:
- The test is unpublished so outsiders cannot check it.
- The blog does not say whether fine tuning could bring the skill back.
- These figures are Moonshot’s own and not independently verified.
Benchmarks: coding, context and speed
- Coding agents: Kimi Code Bench 3.0 Long horizon shows 61.2% against 55.95% for K3. That test is internal.
- Hallucination rate: On Kimi Hallu Bench Hard the share of false claims fell from 23.4% to 8.3%. The API quickstart gives 6.85% on a different test set. Neither is a daily error rate.
- Context window: A 1 million token context window on a model with 2.8T total and 104B active parameters.
- Speed: Kimi K3.1 Pro Highspeed runs 2 to 3 times faster than the standard model. Artificial Analysis measured about 81 output tokens per second for standard.
Price
Token rates
Prices are per million tokens on the Kimi API:
- Kimi K3.1 Pro: $2 input and $10 output and $0.20 cached input
- Kimi K3.1 Pro Highspeed: $4 input and $20 output
- Kimi K3: $3 input and $15 output
One review puts the standard rate at half of Opus 5.5’s list price. Check current rates before you budget.
Real cost and setup
The model is wordy. Artificial Analysis measured $2.11 per task against $0.72 for GPT 6.1 Sol at the same token prices. Test your own workload first.
- Model ID:
kimi k3.1 pro - Thinking mode is always on. Set
reasoning_effortto low to cut tokens. - API access needs a $1 minimum top up.
- Zero Data Retention is available for API customers.
Should you use Kimi K3.1 Pro?
Use it if:
- You run coding agents and want near Opus results on terminal tasks.
- You want open weights you can host yourself after November 4.
- You are pairing it with the best AI tools for SEO and blogging and need cheap long context.
Skip it if:
- You need top science or CAD scores.
- You need cyber testing that outsiders have verified.
Kimi K3.1 Pro is a strong value pick with a smaller hacking edge than Moonshot’s pipeline could have produced. If you handle sensitive code read our notes on AI wearable privacy risks and review the best cybersecurity gadgets for your own setup.
FAQs about Kimi K3.1 Pro
What is Kimi AI
Kimi is the AI assistant and model family from Moonshot AI in Beijing. You can use it on the web and in mobile apps. It also runs in Kimi Code and through the Kimi API.
Does Kimi AI have a limit?
Yes. API limits depend on your top up tier and cover concurrent requests and requests per minute and tokens per minute and tokens per day. Kimi Code adds a 5 hour rolling quota and a monthly quota.
How secure is Kimi AI?
API customers can use Zero Data Retention. Kimi Business offers enterprise data privacy with separate personal and organization accounts. Once weights are released you can also run the model on your own servers.
Is Moonshot AI secure?
Moonshot says it removed cyber related data to lower misuse risk. That claim rests on an unpublished internal test so outside testing is still needed. The company also faces US accusations that came without published evidence.
What are the usage limits for Kimi K3?
Kimi K3 opens after a $1 minimum API top up and your total top up sets your rate tier. Kimi Code reportedly starts K3 access at the Moderato plan and needs Allegretto or higher for the full 1 million token context. Limits can change so check your console.
