Kimi K3.1 Pro closes gap to Anthropic’s Claude Opus 5.5

The Chinese AI model Kimi K3.1 Pro from Moonshot AI scored 55 points on the Artificial Analysis Intelligence Index, three points behind Claude Opus 5.5 and fourth overall. Moonshot released the model as a post-trained version of Kimi K3, using significantly more post-training compute (additional training after the base model’s initial phase).

Artificial Analysis places only Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56 above K3.1 Pro. The model’s score rose 11 points from K3, which matters because Moonshot plans to publish the weights by November 4, 2026.

If that release proceeds and the rankings hold, K3.1 Pro would become the highest-scoring open-weights model, meaning its parameters would be available for download rather than restricted to a hosted service. Moonshot also launched Kimi K3.1 Pro Highspeed, which the company says runs two to three times faster.

Moonshot offers both versions through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The commercial contest now extends beyond benchmark position because buyers can compare the standard model with a faster, higher-priced option.

What do the agentic benchmarks show?

Moonshot reports a GDPval-AA score of 1,846 Elo (a relative performance scale), 312 points above K3. That puts K3.1 Pro behind Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857, according to Artificial Analysis.

On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and narrowly behind Claude Opus 5.5 at 66.4 percent. K3 managed only 13 percent on the same terminal-use test.

Claim-level hallucinations—unsupported factual claims in model output—fell from 23.4 percent with K3 to 8.3 percent with K3.1 Pro on Moonshot’s internal Kimi Hallu Bench Hard. Moonshot describes that result as “a stress test, not a measure of general error rates,” and GPT-6 Astra still scores lower at 5.5 percent.

Moonshot says K3.1 Pro gained 7.2 points on its internal Financial Modeling benchmark and 11.6 points on Financial Research. For general office tasks, the model ranks first among the tested models on OfficeQA Pro v2.

Moonshot’s benchmark table reports every result with reasoning effort set to “max.” The full comparison is:

Benchmark	Kimi K3.1 Pro	Claude Opus 5.5	GPT-6 Astra	Claude Fable 5.1	DeepSeek-V4.1-Flash	GLM-5.3
Terminal-Bench 4.0	65.7	66.4	58.2	57.9	42.3	41.8
SWE Marathon v1.1	51.7	46.3	42.5	45.6	40.0	42.5
DeepSWE v1.1	75.5	74.2	74.1	67.4	77.0	66.9
BenchCAD (with tools)	92.7	96.2	95.9	92.6	81.6	--
HLE-Full (with tools)	70.0	67.7	57.2	65.6	58.8	62.5
BrowseComp	92.1	90.8	91.5	--	--	--
OfficeQA Pro v2	64.4	62.2	54.4	54.5	40.4	31.1
Automation Bench-AA	73.3	69.5	68.5	59.4	66.6	62.2
GDPval-AA v2.1 (Elo)	1,846	1,866	1,573	1,754	1,746	1,650
GPQA-Diamond	94.2	93.9	96.1	93.7	92.2	91.7
ExploitBench	74.7	91.0	100	83.0	--	54.4

K3.1 Pro leads the supplied field on SWE Marathon v1.1, HLE-Full, BrowseComp, OfficeQA Pro v2, and Automation Bench-AA. Claude Opus 5.5 leads Terminal-Bench 4.0 and BenchCAD, GPT-6 Astra leads GPQA-Diamond and ExploitBench, and DeepSeek-V4.1-Flash leads DeepSWE v1.1.

The overall picture is competitive rather than dominant: K3.1 Pro wins several agentic and office tests but does not lead every test, including BenchCAD, GPQA-Diamond, and ExploitBench.

The evidence supports a major advance, not a clean sweep of every benchmark.

Can Kimi K3.1 Pro build usable 3D projects?

Moonshot says K3.1 Pro works directly with Unity, Unreal Engine, and Godot, handling both quick fixes and full project builds. The company positions this as native game-engine operation, although its announcement offers benchmark results rather than a full audit of completed projects.

On BenchCAD, K3.1 Pro scores 92.7 percent with tools, close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Without tools, it reaches 72.8 percent, nearly matching Claude Opus 5.5 at 73.0 percent.

Moonshot says the model is strongest at parametric modeling, where dimensions and features change through defined parameters. It particularly handles mechanical parts, architectural models, and configurable products, and can write and run scripts for Blender and CAD (computer-aided design) software without MCP (Model Context Protocol) connectors.

Moonshot also launched “3D Make” on Kimi.com. Users can preview, edit, and export 3D models through that feature, extending the company’s claim from code assistance into a visible asset workflow.

Moonshot’s 3D pitch is strongest as a workflow claim, while the public scores remain close to Claude and GPT-6 Astra.

What changed in model scale and API price?

K3.1 Pro retains Kimi K3’s dimensions: 2.8 trillion total parameters (learned numerical weights), 104 billion active parameters, and a one-million-token context window (the input the model can process in one request). It accepts both text and images.

Artificial Analysis places K3.1 Pro at the top of its Pareto frontier, meaning the best balance of capability and cost. It says no model with a comparable parameter count scores higher on the Intelligence Index.

The standard API (application programming interface) prices use MTok, meaning one million model tokens, as the billing unit. Cache input refers to context reused from an earlier request.

Variant	Input	Output	Cache input
Kimi K3.1 Pro	$2.00	$10.00	$0.20
Kimi K3.1 Pro Highspeed	$4.00	$20.00	$0.20

All prices are per MTok. Moonshot cut standard token prices by 33 percent from K3: input fell from $3 to $2, output from $15 to $10, and cache hits from $0.30 to $0.20.

Highspeed charges twice as much for input and output, while its cache price stays at $0.20. Moonshot’s two-to-three-times speed claim remains a company claim rather than an independently reported Artificial Analysis result.

Artificial Analysis estimates that each Intelligence Index task costs $1.98, nearly unchanged from K3’s $2.00. The model produces 14 percent more output tokens, offsetting its lower per-token price, according to the analysis.

The model holds its size while cutting standard API prices, making its efficiency claims easier to test than its raw scale.

Why did Moonshot remove cyber capabilities?

Moonshot devoted a full section of its technical blog post to cybersecurity. It says vulnerability-research capabilities kept growing as post-training scaled, because cyber-related content appeared in the general mid-training corpus and cyber tasks featured in the RL (reinforcement learning) environment mix.

For the open-weights release, Moonshot removed cyber-related data from both training phases. On its internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software including the Linux kernel, Node.js, and JavaScriptCore, the uncleaned model solves 90.5 percent of tasks.

After data cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test, while Kimi K3 reaches 33.3 percent, showing that the cleaned model retains much more capability than its predecessor but trails DeepSeek.

Moonshot frames the reduction as a deliberate boundary on offensive use. It argues that no open-weights model should push “the absolute frontier of offensive cyber capabilities,” while also wanting to raise open-model capability gradually because such models help defenders audit legacy systems.

Anthropic previously showed that even GLM-5.3, which scores 45 on the Artificial Analysis Intelligence Index, can write exploits at near-Claude Mythos levels. The claim sets a broader policy problem around releasing capable models, not just a Kimi-specific benchmark result.

The security restriction has a measurable cost: cleaned K3.1 Pro remains far above K3 on xdev-bench-hard, but below both the uncleaned model and DeepSeek-V4.1-Flash.

Will open weights change the race?

Moonshot’s planned November 4, 2026 weight release remains conditional, and the leaderboard can move before then. Artificial Analysis estimates that, if Moonshot delivers as promised and nothing else changes, K3.1 Pro would cut the open-to-proprietary gap from 12 points to three.

The current API release gives developers access now, including the faster Highspeed variant. The open-weights claim still carries a deliberate capability discount because Moonshot removed cyber data before the planned release.

For model buyers, the immediate choice is a standard API at $2 per million input tokens or a faster option at $4. For open-model researchers, November 4 is the date on which Moonshot’s score claim becomes testable at the promised weight release.

Current evidence supports a near tie at the open-model frontier, not a guaranteed lead after November.