Moonshot’s Kimi K3.1 Pro Just Beat GPT-6 Astra on Coding – But Deliberately Made Itself Worse at Hacking
TL;DR
- Moonshot released Kimi K3.1 Pro on October 23-24, 2026 – a 2.8-trillion-parameter open-weight model (104B active) that scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%).
- Moonshot deliberately stripped cyber-offensive training data before release: an unfiltered checkpoint scored 90.5% on its internal hacking benchmark, the shipped model scores 81.0% – a 9.5-point cut the company made on purpose and disclosed.
- It’s genuinely mixed, not a clean win: K3.1 Pro leads on some knowledge-work benchmarks but trails badly (38.1 vs 68.1/63.3) on a harder science-coding test called Terminal-Bench Science 0.1.
- Pricing dropped to $2/$10 per million tokens (input/output), well under GPT-6 Astra’s $10/$50 – full weights are reportedly planned for November 4, 2026, not yet confirmed as shipped.
Reviewed by Imran Khan Pathan, Editor at AI Tech Safar. An earlier draft of this piece said K3.1 Pro has “3 trillion parameters” and cited a benchmark source called “RDWorld” for its weakest result. Neither checked out – the real parameter count is 2.8T total/104B active (same scale as K3), and the benchmark in question is Moonshot’s own “Terminal-Bench Science 0.1,” confirmed via MarkTechPost’s coverage of Moonshot’s release materials. Numbers below are cross-checked against Moonshot’s own published comparison tables rather than taken from a single secondhand writeup.
Last updated: October 26, 2026.
Moonshot AI released Kimi K3.1 Pro on October 23-24, 2026 (Moonshot’s own materials carry the 23rd; most press coverage landed the next day), and the numbers are genuinely striking. The 2.8-trillion-parameter open-weight model (104 billion active per token, same architecture scale as its predecessor) scored 65.7% on Terminal-Bench 4.0, a benchmark that tests AI coding agents on realistic terminal tasks. That puts it ahead of OpenAI’s GPT-6 Astra (58.2%) and just 0.7 points behind Anthropic’s Claude Opus 5.5 (66.4%), per Moonshot’s own comparison chart.
But the more interesting story isn’t what K3.1 Pro can do. It’s what Moonshot deliberately stopped it from doing.
The Hacking Downgrade Nobody Else Is Doing
Moonshot confirmed it had removed cyber-related data from the model’s training dataset and reinforcement-learning stages specifically to reduce its ability to conduct offensive cyber operations, while keeping the general coding ability that makes the model useful in the first place.
The results were measurable. On Moonshot’s internal benchmark, xdev-bench-hard – expert-curated, source-code-based vulnerability discovery and exploitation tasks, scored at pass@3 – Kimi K3.1 Pro reached 81.0%. A checkpoint trained on the exact same pipeline without the risk-reduction measures scored 90.5%.
That’s a deliberate 9.5-point reduction in offensive cyber capability, and some useful context that got lost in most coverage of this: Kimi K3, the previous generation, scored just 33.3% on the same test, and rival Chinese model GLM-5.3 scored 57.1%. So K3.1 Pro’s “nerfed” 81.0% is still dramatically higher than K3 ever managed – Moonshot didn’t make the model less capable than its predecessor, it made it less capable than it could have been, while still landing well above most of the field.
Moonshot’s own framing is specific rather than vague: the company says it doesn’t release open weights that would extend the absolute frontier of cyber capability, and describes K3.1 Pro’s offensive capability as staying behind what frontier proprietary models have had publicly available for at least three months. This is a notable move in an industry where labs typically market raw capability gains without qualification – Moonshot shipped a model weaker than the one it could have shipped, and published the exact cost of that choice.
What K3.1 Pro Actually Is
Kimi K3.1 Pro is a post-trained successor to Kimi K3, which launched in July 2026 at the same 2.8T/104B parameter scale. It ships with a 1-million-token context window and native visual understanding, and comes alongside a faster companion variant, Kimi K3.1 Pro Highspeed.
The model is available now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Full model weights are reportedly planned for release by November 4, 2026 – treat that as Moonshot’s stated intention rather than a confirmed date, since it hasn’t shipped yet as of this writing. Once the weights are out, K3.1 Pro is positioned to become one of the strongest open-weight models available, if not the strongest.
Worth knowing: Kimi K3’s weights already got real third-party use once released – Cognition AI’s Devin AI and Fireworks AI’s Ember-1 were both post-trained on top of K3. That’s a genuine signal that Moonshot’s open releases aren’t just PR exercises; other labs are actually building on them.
The Pricing
Moonshot cut prices compared to Kimi K3. K3.1 Pro runs $2 per million input tokens and $10 per million output tokens, down from K3’s $3/$15 – a real cut on both ends. The faster Highspeed variant costs more, at roughly $4/$20 per million tokens, and cached input tokens on the standard model are billed separately at a lower rate. Compared to GPT-6 Astra’s $10 input / $50 output pricing, K3.1 Pro is dramatically cheaper either way – Moonshot is clearly betting that cost wins over developers who don’t specifically need Astra’s marginal performance edge.
The Benchmark Breakdown
| Benchmark | Kimi K3.1 Pro | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 58.2% | 66.4% |
| AA Intelligence Index | 55 | 53 | 58 |
| OfficeQA Pro v2 | 64.4 | 54.4 | 62.2 |
| Kimi Code Bench 3.0 Long-horizon | 61.2% (K3: 55.95%) | – | – |
| BenchCAD (with tools) | 92.7% | 95.9% | 96.2% |
| Terminal-Bench Science 0.1 | 38.1 | 68.1 | 63.3 |
| xdev-bench-hard (cyber) | 81.0% (unfiltered: 90.5%) | – | – |
On the Artificial Analysis Intelligence Index, K3.1 Pro scores 55 – ahead of GPT-6 Astra (53), roughly level with Claude Fable 5.1, and behind only Claude Opus 5.5 (58). It also outright leads on OfficeQA Pro v2, a knowledge-work benchmark where it beats both Astra and Opus 5.5.
The genuine weak spot is Terminal-Bench Science 0.1, a harder science-and-coding test where K3.1 Pro scores just 38.1 against GPT-6 Astra’s 68.1 and Opus 5.5’s 63.3 – a 30-point gap that doesn’t show up anywhere in the headline Terminal-Bench 4.0 number. The Terminal-Bench advantage clearly doesn’t carry over to every coding-adjacent task.
The “Bench Maxxing” Problem
Not everyone is convinced by the headline numbers. Developers on OpenCode Go noted that K3.1 Pro “shows signs of bench maxxing gone too far in addition to cybersecurity nerf” – the concern being that Moonshot optimized hard for specific benchmarks like Terminal-Bench 4.0 while leaving real gaps elsewhere, which the Terminal-Bench Science 0.1 result arguably supports.
There’s also a less technical controversy attached to the launch. VentureBeat’s coverage carried the headline “Moonshot’s Kimi K3.1 Pro flirts with open weights leadership amid paid influencer scandal allegations” – Moonshot hasn’t publicly addressed the specifics, and the details of what exactly is being alleged remain unclear from public reporting so far. Worth noting this isn’t Moonshot’s first trust-related friction this year either: in September, Anthropic separately accused the company of routing user requests to Claude models and distilling from them without authorization, a claim Moonshot has disputed. None of this directly concerns K3.1 Pro’s benchmarks, but it’s relevant context for how much weight to put on any single number Moonshot publishes about itself.
What This Means for the Open-Weight Race
When Moonshot releases K3.1 Pro’s weights – if the November 4 timeline holds – it’s positioned to become one of the most capable open-weight models available, giving developers a locally-hostable alternative to API-only models like GPT-6 Astra and Claude Opus 5.5.
But the open-weight release also carries the exact risk Moonshot tried to mitigate. Once weights are public, anyone can fine-tune the model, including on the cyber-offensive data Moonshot deliberately filtered out of training. The 90.5% score of the unfiltered checkpoint is a reminder that the underlying capability hasn’t been deleted, only suppressed in the version Moonshot is shipping – and suppression baked into training data is a choice a well-resourced actor with the raw weights could, in principle, try to reverse.
For enterprises evaluating K3.1 Pro: it’s a capable, cheap, open model with genuinely strong coding and knowledge-work scores, and a real, documented gap on harder scientific-coding tasks. The safety trade-off is honest rather than hidden, which is itself unusual – but “honest about the trade-off” and “the trade-off is permanent once the weights are public” are two different claims, and only the first one is really guaranteed.
Related Reading on AI Tech Safar
- Moonshot’s Kimi Models Were Jailbroken to Discuss Bioweapons: What We Know So Far
- Kimi K3: Inside China’s 2.8-Trillion-Parameter Bet to Beat GPT-5 and Claude
- Anthropic’s AI Sent a Fake Murder Tip to Police – Then It Cut the Internet
Useful Sources
- MarkTechPost – Moonshot AI Releases Kimi K3.1 Pro: 2.8T Total/104B Active Params
- Digital Trends – Moonshot tells Kimi AI to code away, but leave the hacking alone
- Unite.AI – Moonshot AI Releases Kimi K3.1 Pro
- OrcaRouter – Kimi K3.1 Pro Cut Its Own Cyber Capability by 9.5 Points
- VentureBeat – Moonshot’s Kimi K3.1 Pro flirts with open weights leadership
- Shattered.io – Moonshot Curbs Kimi K3.1 Hacking: 81.0% vs 90.5%
For the broader context and related reading, explore our AI Models 2026: Comparing Gemini, GPT, Claude and More.


