Kimi K3 vs Claude: K3.1 Pro, Opus 5.5 and price

Claude still leads Kimi on the hardest coding tests, but Moonshot's newest model, Kimi K3.1 Pro, comes close for far less money. On Vals AI's independent Terminal-Bench 4.0 run, Claude Opus 5.5 scored 65.2%, Claude Sonnet 5.5 64.1% and Kimi K3.1 Pro 63.6%, ahead of Claude Fable 5.1 at 58.1%, and a K3.1 Pro task cost $4.57 against Opus 5.5's $13.20. Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output, the same as Sonnet 5.5 and half of Opus 5.5. The original Kimi K3, released in July 2026, scored 17.2% on the same run and costs $3 and $15. Kimi's weights are open under Moonshot's own license, K3's now and K3.1 Pro's promised by November 4; Claude's are not.

Key takeaways

• Kimi K3.1 Pro (October 23, 2026): $2 / $10 per million tokens, 63.6% on Vals AI's Terminal-Bench 4.0, eighth in LMArena's WebDev votes (preliminary).
• Kimi K3 (July 2026): $3 / $15, 17.2% on the same run, 17th in WebDev votes.
• Claude: Opus 5.5 $4 / $20 (65.2%, first in WebDev), Sonnet 5.5 $2 / $10 (64.1%, third), Fable 5.1 $10 / $50 (58.1%, fifth).
• Open weights: Kimi K3 is downloadable under the Kimi K3 License, which allows commercial use with conditions for very large services.

Sources: Moonshot's pricing page and K3.1 Pro guide, the Kimi K3 model card and license; Anthropic's pricing; Vals AI and LMArena, opened October 27, 2026.

What are Kimi K3 and K3.1 Pro?

Kimi is Moonshot AI's model family. Kimi K3, released in July 2026, is a mixture-of-experts model with 2.8 trillion parameters, 104 billion of them active for each token, and a 1M-token context window. It reads images as well as text.

Kimi K3.1 Pro, released October 23, 2026, is K3 post-trained with much more reinforcement learning, at the same size. Moonshot calls it its most capable open-weight model and says the weights will be out by November 4. A Highspeed version runs two to three times faster. Its thinking is always on, with low, high and max effort (max by default).

Moonshot is closing its older K2 models: from October 31, 2026 new accounts can't use them, and all K2 models are due to retire by the end of December.

How much do Kimi and Claude cost?

Per million input and output tokens on each company's API:

• Kimi K3.1 Pro: $2 and $10; cached input $0.20. Highspeed: $4 and $20.
• Kimi K3: $3 and $15; cached input $0.30.
• Claude Sonnet 5.5: $2 and $10; cached input $0.10.
• Claude Opus 5.5: $4 and $20; cached input $0.20.
• Claude Fable 5.1: $10 and $50; cached input $0.25.

Both bill the whole 1M-token window at the standard rate. Moonshot's API is pay-as-you-go with no subscription, and K3.1 Pro unlocks after a top-up of at least $1. Moonshot also sells Kimi memberships and Kimi Code separately.

Per task, the gap is wider than per token. On Vals AI's Terminal-Bench 4.0 run, an average task cost $4.57 on K3.1 Pro, $13.20 on Opus 5.5, $16.51 on Sonnet 5.5 and $17.18 on Fable 5.1. K3 cost $12.48 a task and took over three hours.

How do they compare on coding?

Independent results:

• Vals AI, Terminal-Bench 4.0 (one shared harness, updated October 26): Opus 5.5 65.15%, Sonnet 5.5 64.14%, Kimi K3.1 Pro 63.64%, GPT-6 Astra 59.60%, Fable 5.1 58.08%, Kimi K3 17.17%.
• LMArena WebDev (people voting on web builds, October 27): Opus 5.5 first (1809), Sonnet 5.5 third (1769), Fable 5.1 fifth (1745), Kimi K3.1 Pro eighth (1695, marked preliminary), Kimi K3 17th (1656).

Moonshot's own numbers for K3 compare it with Claude Fable 5 and Opus 4.8, not the current Claude models: on Terminal-Bench 2.1, K3 88.3% and Fable 5 88.0%; on DeepSWE, 67.5% and 70.0%. Moonshot notes the models ran in different harnesses. Its K3.1 Pro guide gives no benchmark scores yet.

What does Kimi's open license allow?

Kimi K3's weights and code are released under the Kimi K3 License. It allows commercial use, modification and redistribution, with three conditions:

• Keep Moonshot's copyright notice.
• A company running the model as a service with more than $20 million in revenue over 12 months needs a separate agreement with Moonshot.
• A product with more than 100 million monthly users or $20 million in monthly revenue must display "Kimi K3" in its interface.

Internal use, and use through Moonshot's own products or certified partners, are exempt from the last two. At 2.8 trillion parameters, running it yourself takes a multi-GPU server. Claude models can't be downloaded; they run only on Anthropic's API and its cloud partners.

When would you choose Kimi over Claude?

• Cost at volume: K3.1 Pro's cost per task on Vals AI's run was about a third of Opus 5.5's.
• Self-hosting or fine-tuning: only Kimi offers weights.
• Claude-compatible tooling: Moonshot's API has an Anthropic-compatible endpoint, which is how tools built for Claude can run Kimi.

Claude is the safer choice for the hardest agentic coding, where Opus 5.5 and Sonnet 5.5 lead both independent rankings, and where data handling matters: Anthropic offers US-only inference and its own published retention terms, and Kimi's API is run by a company based in China.

Can you use Kimi and Claude together?

Yes. Cline, Bolt and others offer Kimi alongside Claude. Poly (usepoly.co) offered Kimi from September to October 2026 and retired it: Moonshot's API terms let it use what customers send to improve its models unless agreed otherwise in writing, and no other lab in Poly does that by default. A Poly room now mixes Claude with GPT, Grok and Muse, each person picking their own model (multi-model AI coding). More comparisons in the best AI model for coding.