Kimi K3.1 Pro vs Claude Opus 5.5
Which should you use: Kimi K3.1 Pro or Claude Opus 5.5?
It depends. Claude Opus 5.5 scores higher on the Artificial Analysis Intelligence Index v5 (58 vs 46), leads on hard science and knowledge tests, and measured over twice as fast, but costs twice as much per token ($4/$20 vs $2/$10 per 1M). Kimi K3.1 Pro leads Artificial Analysis's Terminal-Bench 4.0 (67% vs 60%). Its open weights are promised by November 4 but not released, and its licence is unpublished. Pick Opus for the strongest model; pick Kimi for cheaper terminal and agentic coding work, and test it first.
Kimi K3.1 Pro (Moonshot AI) launched on October 23, and Claude Opus 5.5 (Anthropic) on September 22. “Vendor” marks a claim from the model’s maker; “AA” marks an independent measurement by Artificial Analysis. All facts are as of October 28.
Score change: on October 27 AA moved to Intelligence Index v5, and our earlier issues quoted the old scale. On v5, Kimi K3.1 Pro scores 46 and Opus 5.5 scores 58. The old figures (55 and 58) came from a retired version and do not compare (AA).
Side by side
| Kimi K3.1 Pro | Claude Opus 5.5 | |
|---|---|---|
| Input, $ per 1M tokens | $2.00 (Kimi) | $4.00 (Anthropic) |
| Output, $ per 1M tokens | $10.00 | $20.00 |
| Cache read, $ per 1M tokens | $0.20 | $0.20 |
| Cache write, $ per 1M tokens | $2.00 (5 min) / $4.00 (1 h) | $5.00 (5 min) / $8.00 (1 h) |
| Faster or cheaper variants | Highspeed $4/$20, 2–3× faster | Fast mode $8/$40, up to 2.5× faster; Batch $2/$10; US-only inference 1.1× |
| Long-context pricing | Flat, no tiers | Flat across the full 1M |
| AA Intelligence Index v5 | 46 (AA) | 58, at max effort with fallback (AA) |
| Terminal-Bench 4.0 (AA) | 67%, the top of all models | 60% |
| Terminal-Bench 4.0 (vendor) | 65.7%, max effort (Kimi) | 66.4%, xhigh effort (Anthropic) |
| HLE-Diamond (AA) | 37% | 68% |
| FrontierCode 2.0 / Terminal-Bench-Science (AA) | 25% / 34% | 46% / 59% |
| AA-Omniscience v1.1 / CritPt v2 (AA) | 30 / 45% | 48 / 61% |
| GDPval-AA v2.1 (Elo, AA) | 1,846 | 1,866 |
| ARC-AGI-2 | No result found | 93.3% (High), 91.7% (Max); independent, verified (ARC Prize) |
| LMArena text | Not ranked yet (Arena) | #2, 1507±8 (High) |
| Output speed (AA) | 41.4 tokens/s | 98.4 tokens/s |
| Time to first token (AA) | 2.82 s | AA’s two pages disagree (700 s vs 34 s); not compared |
| Tokens used on the Index (AA) | 110M | 220M |
| Cost per AA Index task | Not found | $10.04 |
| Context window | 1M tokens (Kimi docs) | 1M tokens |
| Max output | 131,072 by default, up to 1,048,576 | 128,000 |
| Open weights | Promised “by Nov 4, 2026”; not released yet | No, closed |
| Licence | Not published | Closed |
| Inputs | Text, image and video in; text out | Text, image, files and PDFs in; text out |
| Released | 2026-10-23 | 2026-09-22 |
Sources for the AA rows: the AA comparison page, Intelligence Index v5 article and Kimi launch article.
Price and cost
Opus 5.5 charges twice the Kimi rate for input and output tokens. Cache reads cost the same, at $0.20 per 1M tokens.
- Kimi API: unlocks after a top-up of at least $1 (Kimi docs).
- Opus 5.5 elsewhere: Bedrock, Azure, Vertex and Claude Platform on AWS list the same $4/$20/$0.20 standard rates. Regional and EU endpoints cost $4.40/$22/$0.22 (OpenRouter).
- Kimi elsewhere: no Bedrock, Vertex or Azure offer for K3.1 Pro was found. Bedrock carries the older Kimi K3.
- Effort and cost: Opus 5.5 scores 35 on the Index at low effort and 58 at max, at about 10× the cost per task (AA).
Benchmarks: read with care
AA and the vendors disagree on Terminal-Bench. AA has Kimi ahead (67% vs 60%), while the vendor numbers have Opus ahead (66.4% vs 65.7%) (VentureBeat).
- Kimi’s own runs: Moonshot’s table mixes its own runs with scores reported by others, on different harnesses. All its Kimi numbers use reasoning effort “max”, temperature 1.0 (Kimi).
- Moonshot’s science result: its own Terminal-Bench-Science score is 38.1 for Kimi against 63.3 for Opus (VentureBeat).
- Anthropic’s caveat: it says benchmark margins are less reliable, and that Opus 5.5 often suspects it is being evaluated (Anthropic).
Kimi K3.1 Pro
- Open weights: promised “by Nov 4, 2026” and not yet published. AA still lists the model as proprietary (Kimi).
- Licence: the K3.1 Pro licence is unpublished. Its predecessor’s “Kimi K3 License” allows use, modification and sale, with extra terms for large operators (Hugging Face). Whether K3.1 Pro uses the same terms is unknown.
- Size: 2.8T total and 104B active parameters (AA). K3 ships MXFP4 weights and runs on vLLM and SGLang (Hugging Face). A hardware spec was not found.
- Reasoning: it always thinks, with effort set to low, high or max (Kimi docs).
- Agent features: tool_choice, dynamic tool loading, strict JSON schema, and 3D and CAD scripting (Blender, FreeCAD). Its web search is “not recommended” for now.
- Where to get it: Kimi.ai web and mobile, Kimi Work (Windows, Apple-silicon Mac), Kimi Code and the API. Enterprise access is off by default (Kimi).
- Data location: its privacy policy points to Singapore servers (VentureBeat).
- Weak spots: slow (41 tokens/s) and verbose (110M tokens on the Index vs a median of 63M) (AA). Its hallucination rate on AA-Omniscience is 44% (AA).
- Cyber data: Moonshot says it removed cyber-related training data (internal exploit score 81.0% vs 90.5% unfiltered) (Kimi).
- Docs quality: the quickstart still has “[xx.x%]” placeholders and a 6.85% hallucination figure, against 8.3% in the blog.
Claude Opus 5.5
- Weights: closed; it cannot be self-hosted, and its size is undisclosed (AA).
- Reasoning: adaptive thinking only, as thinking-off was removed. Effort runs from low to max (Anthropic).
- Agent features: computer use, browser use, web search ($10 per 1,000), code execution and Managed Agents ($0.08 per session-hour) (Claude docs).
- Where to get it: Claude apps (Pro, Max, Team, Enterprise), Claude Platform, AWS Bedrock, Google Cloud Vertex and Microsoft Foundry (Anthropic).
- Regions: global by default, US-only inference at 1.1×, and EU via Bedrock or Vertex (OpenRouter).
- Safeguards: cyber, biology and frontier-LLM requests can silently fall back to Opus 4.8 or Opus 5, or be blocked. Opus 5.5 is not yet in the Cyber Verification Program (Claude Help Center).
- Cost and verbosity: the highest cost per AA task ($10.04) and 220M tokens on the Index (AA).
Both models offer zero data retention.
Alleged influencer brief (unverified)
This is an allegation, not an established fact. An X user posted screenshots of what he says is a Kimi K3.1 Pro creator brief. It reportedly said “Please DO NOT put on the paid partnership label” and asked for pre-approval of copy.
The document is unauthenticated. The issuer, the scale and whether any paid posts ran are unknown. Moonshot had made no public statement as of October 27 (VentureBeat, Pekingnology).
Who each suits
Kimi K3.1 Pro suits teams that want:
- Lower token prices, at half the Opus rate
- Terminal and agentic coding work, where AA has it ahead
- An open-weight option, but only once the weights and licence land
- Extra checks on facts, given the 44% hallucination rate
Claude Opus 5.5 suits teams that want:
- The highest overall AA v5 score
- Strength on hard knowledge and science tests (HLE-Diamond, Omniscience, CritPt, FrontierCode)
- Faster output, at 98 tokens/s against 41
- A managed model with broad cloud access, if they accept twice the token price and its safeguard fallbacks
Facts as of Oct 28, 2026
Sources
- Artificial Analysis: Intelligence Index v5
- Artificial Analysis: Kimi K3.1 Pro launch benchmarks
- Artificial Analysis: Kimi K3.1 Pro
- Artificial Analysis: Claude Opus 5.5
- Artificial Analysis: Kimi K3.1 Pro vs Claude Opus 5.5
- Artificial Analysis: Kimi K3.1 Pro vs Claude Opus 5.5 (High effort)
- Kimi blog: Kimi K3.1 Pro
- Kimi: K3.1 Pro pricing
- Kimi Platform: K3.1 Pro quickstart
- Hugging Face: moonshotai/Kimi-K3
- Hugging Face: Kimi K3 licence
- Anthropic: Introducing Claude Opus 5.5
- Anthropic: Claude Opus
- Claude docs: Pricing
- Claude Help Center: why Claude switched models in your conversation with Opus 5 or Opus 5.5
- OpenRouter: Claude Opus 5.5
- ARC Prize: Claude Opus 5.5 results
- Arena: text leaderboard
- VentureBeat: Moonshot's Kimi K3.1 Pro flirts with open-weights leadership amid paid-influencer scandal allegations
- Pekingnology: Kimi's global moment, and a very Chinese scandal