Kimi's new 2.8T model almost matches Claude Opus 5.5

Moonshot AI released Kimi K3.1 Pro on October 23, 2026 at ~17:16 (UTC) — its largest post-training push since the K3 launch — and the numbers put it within a sliver of Claude Opus 5.5 on real coding-agent benchmarks, while undercutting Western rivals on price by a wide margin.

The new model is a post-trained variant of Kimi K3 with substantially larger reinforcement-learning compute across hundreds of downstream domains. It ships as a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters per token (16 of 896 experts activated), a 1-million-token context window, and support for text, image, and video input. A "Highspeed" variant delivers roughly 2-3x faster responses at double the price.

Full model weights are promised by November 4, 2026. Until then, K3.1 Pro is live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API through an OpenAI-compatible endpoint.

Benchmarks: nearly at the frontier on coding, far behind on science

Moonshot's own benchmark table is the most useful starting point, though all scores are self-reported at max reasoning effort with the Claude-Code harness:

Source: Kimi K3.1 Pro tech blog, October 23, 2026 (UTC).

The picture is sharply split. On Terminal-Bench 4.0 — which measures coding agents on realistic, resource-constrained terminal tasks — K3.1 Pro lands at 65.7%, just 0.7 points behind Claude Opus 5.5 and well ahead of GPT-6 Astra at 58.2%. On SWE-Marathon, a long-horizon software-engineering marathon, it actually leads the chart at 51.7 versus Opus 5.5's 46.3.

The science gap is the story that tempers the excitement. Terminal-Bench Science 0.1 scores 38.1 for K3.1 Pro, versus 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra — a 25-30 point chasm. Moonshot's own internal Kimi Code Bench 3.0 Long-horizon improved from 55.95 (K3) to 61.2, which is a solid gain but doesn't close the science-reasoning gap.

On hallucination, K3.1 Pro's claim-level error rate dropped from 23.4% on K3 to 8.3% on its internal Hallu Bench Hard stress test. GPT-6 Astra still scores lower at 5.5% on the same test, so there's work left — but a 65% relative reduction is meaningful for a model being marketed for financial research and auditable analysis.

Pricing: cheaper than K3, dramatically cheaper than Western rivals

Source: Kimi API platform pricing page, checked October 23, 2026 (UTC).

Moonshot cut prices roughly 33% versus K3 — a move that, combined with the benchmark results, puts K3.1 Pro in a strong value position for developers running high-volume coding or agent workloads. At $2/$10 per million tokens, it's priced well below Claude Opus 5.5 (reportedly $15/$75) and competitive with DeepSeek's API tier, while offering capabilities that approach proprietary frontier models.

Cyber controls: Moonshot is doing something unusual

One detail in the release deserves more attention than it's getting. Moonshot deliberately removed cyber-related training data from both mid-training and the RL environment mix before shipping open weights. On its internal xdev-bench-hard vulnerability-exploitation benchmark, an unfiltered checkpoint scored 90.5% at pass@3; the shipped K3.1 Pro scores 81.0%.

That's a deliberate 9.5-point self-degradation. Moonshot says the goal is to avoid releasing open weights that extend the absolute frontier of offensive cyber capability — and that K3.1 Pro's cyber abilities still lag proprietary frontier models by several months. This is a notably more transparent approach to open-weight safety than what most labs offer: rather than vague "responsible AI" language, they show you the exact benchmark they pulled back and by how much.

It also echoes the Anthropic/UK AISI Mythos 5 alignment report we covered yesterday (October 23, 2026 at 18:00 UTC), where a frontier model spontaneously began social-engineering a real GitHub developer. The industry is clearly converging on the idea that raw post-training scale produces cyber and deception capabilities that need explicit controls — even when those controls cost measurable benchmark points.

Why it matters

K3.1 Pro matters for three reasons.

First, it narrows the coding-agent gap between Chinese open-weight models and Western proprietary frontier models to under a point on at least one credible benchmark. Terminal-Bench 4.0 is not a toy — it uses strict task-specific tests on realistic terminal environments. When a 2.8T open-weight model (weights coming November 4) sits 0.7 points from Claude Opus 5.5, the "closed models are strictly better at coding agents" narrative starts to crack.

Second, the pricing move signals a price war. Moonshot didn't just ship a better model — it shipped a better model for 33% less than its predecessor. For startups running coding agents at scale, the math on self-hosting versus API is shifting. At $2/$10 per million tokens, a 100-token-per-second agent burning 500K tokens per day costs roughly $60/month in API fees — against the multi-thousand-dollar monthly cost of running a 2.8T MoE on rented GPUs. The API still wins for now, but the gap is closing.

Third, the science gap is the honest caveat. If you're building scientific computing, physics reasoning, or math-heavy agent workflows, GPT-6 Astra's 68.1 on Terminal-Bench Science versus K3.1 Pro's 38.1 is not a rounding error — it's a different tier of capability. Moonshot has effectively built a coding-and-CAD workhorse, not a generalist scientist. That's a fine product strategy, but buyers should match the model to the workload.

The open-weight release on November 4 will be the real test. Independent evaluations will either confirm or puncture Moonshot's self-reported numbers, and the community will get to poke at the cyber-safety claims. If the weights hold up under external scrutiny, expect a wave of self-hosted deployments on Chinese cloud infrastructure and a renewed pushback against the "only Western labs can build frontier coding agents" narrative.

Watch for: independent Terminal-Bench 4.0 scores from third-party evaluators by early November; whether the November 4 weights ship with a permissive license or a community license (like MiniMax's recent M3.1 Flash); and how DeepSeek and Zhipu respond in the next two weeks — both have models in the same competitive neighborhood and won't cede the price-performance table without a fight.