Kimi K3.1 Pro vs Claude: Pricing, Benchmarks, and Which One to Use (2026)

TL;DR

Last updated November 9, 2026, the day Anthropic released Claude Fable 5.2. Kimi K3.1 Pro benchmark numbers are Moonshot-reported unless marked Artificial Analysis. Claude benchmark numbers are Anthropic-reported unless marked. List prices were re-read from Moonshot's and Anthropic's pricing docs that day.

Kimi K3.1 vs Claude depends on which Claude. Kimi K3.1 Pro is Moonshot AI's open-weight model: a 2.8T-parameter mixture-of-experts with 104B active, a 1M-token context and image input, on Moonshot's API since October 23, 2026 at $2 per million input tokens and $10 per million output. Anthropic's current lineup is Claude Fable 5.2 ($8/$40, launched November 9, 2026), Opus 5.5 ($4/$20), Sonnet 5.5 ($2/$10) and Haiku 5.5. K3.1 Pro lists at half the price of Opus 5.5, a quarter of Fable 5.2 and the same as Sonnet 5.5. In Moonshot's own table, with K3.1 Pro run inside Claude Code, it scores 65.7 on Terminal-Bench 4.0 against Opus 5.5's 66.4. Artificial Analysis, testing independently, scores K3.1 Pro 46 on Intelligence Index v5 against 58 for Opus 5.5, and measured it at 67% on Terminal-Bench 4.0, the highest score of any model in its October 23 results. Claude keeps the lead on computer use, office-task Elo and exploit benchmarks.

The Two Lineups

Kimi K3.1 Pro is one model with a faster API variant. Claude is four models, and the comparison changes with the rung. Against Fable 5.2, K3.1 Pro is the challenger at a quarter of the price. Against Opus 5.5, it costs half. Against Sonnet 5.5, the list price is identical and Sonnet reads cache for less. Against Haiku 5.5, K3.1 Pro is the expensive one.

Sources: Moonshot platform pricing (K3.1 Pro and Highspeed); Anthropic pricing and models docs (Claude); Morph's rate card (morph-kimik31). Moonshot bills K3.1 Pro flat with no tiering by context length. Anthropic bills the full 1M window at standard rates on every current model except Haiku 5.5, which switches to the higher rate above 100K prompt tokens. Claude models cap output at 128K tokens; Moonshot's API defaults K3.1 Pro to 131,072 output tokens and allows up to 1,048,576.

Benchmarks and Verification

The fairest head-to-head available is Moonshot's own release table. It puts Kimi K3.1 Pro against Claude Opus 5.5 and Claude Fable 5.1, with the Claude columns labeled "max, w/ fallback". Most of K3.1 Pro's coding rows were run inside Claude Code, Anthropic's own agent harness. Every cell below is vendor-reported by Moonshot. Fable 5.1 is now a legacy model; Fable 5.2 replaced it on November 9, 2026.

Source: Moonshot's Kimi K3.1 Pro model card. K3.1 Pro ran at max effort, temperature 1.0, top-p 0.95. For Terminal-Bench 4.0 the Claude cells are the maximum values reported in model blog posts or on the official leaderboard. Publishers disagree on some shared cells: Anthropic's Opus 5.5 system card lists Fable 5.1 at 55.8 on Terminal-Bench 4.0, and Opus 5.5 at 1846 on GDPval-AA v2.1 and 1822 on AA-Briefcase v1.1. No SWE-bench Verified score exists for K3.1 Pro or for the 5.5-generation Claude models, so none appears here.

What independent testing says

Artificial Analysis runs K3.1 Pro and the Claude lineup under the same conditions. Its numbers split along the same line as Moonshot's table: K3.1 Pro is the strongest open-weights model and competitive on agentic coding, and Opus 5.5 still leads the broad aggregate.

On launch day, October 23, 2026, Artificial Analysis placed K3.1 Pro 4th on its Intelligence Index at 55, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Index v5, released October 27, scores K3.1 Pro at 46, the top open-weights model, with Opus 5.5 leading at 58. Scores from the two versions are not comparable, but the gap to Opus 5.5 is 3 points on the old index and 12 on v5. AA also measured K3.1 Pro at 67% on Terminal-Bench 4.0, the highest score of any model in its launch results and ahead of Sonnet 5.5's 64%, and ranked it 3rd on GDPval-AA at 1,846 Elo behind only Opus 5.5 (1,866) and Muse Wonder Full (1,857). Its hallucination rate on AA-Omniscience fell to 44% from Kimi K3's 53%.

Kimi K3.1 vs Claude Opus 5.5

Opus 5.5 is the Claude most teams are deciding against. Anthropic's docs say to start there: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." It is also the model Moonshot chose to measure against. At list price Opus 5.5 costs twice as much as K3.1 Pro on input and output; cache reads cost the same $0.20 per million on both.

Terminal-Bench 4.0 is a tie. K3.1 Pro scores 65.7 and Opus 5.5 scores 66.4, and Anthropic reports a standard error of 2.6 points on its own Opus 5.5 run. The rest of the coding rows split. K3.1 Pro leads on SWE Marathon (51.7 vs 46.3), DeepSWE (75.5 vs 74.2) and Moonshot's internal long-horizon Code Bench (61.2 vs 58.5). Opus 5.5 leads on ProgramBench (65.0 vs 31.5, the widest gap in the table, where a task only counts when it scores 95% or better) and PostTrainBench (43.8 vs 36.5).

Outside coding, Opus 5.5 holds the knowledge-work, computer-use and closed-book rows: AA-Briefcase (1803 vs 1678 Elo), APEX-Agents (73.5 vs 69.4), OSWorld (81.8 vs 79.6) and HLE without tools (59.8 vs 51.1). K3.1 Pro wins the tool-assisted research rows, HLE with tools (70.0 vs 67.7) and BrowseComp (92.1 vs 90.8), and reports roughly half the hallucination rate on Moonshot's internal test (8.3 vs 15.0).

Kimi K3.1 vs Claude Sonnet 5.5

Sonnet 5.5 is the Claude that K3.1 Pro actually competes with on price. Both list at $2 per million input tokens and $10 per million output. Sonnet 5.5 reads cache at $0.10 per million, half of Moonshot's $0.20, so on cache-heavy agent sessions Sonnet is the cheaper of the two. Moonshot's table does not include Sonnet 5.5, so every number below comes from its own publisher.

The two publisher-reported Terminal-Bench 4.0 numbers point in opposite directions: Anthropic has Sonnet 5.5 at 70.6, Moonshot has K3.1 Pro at 65.7. Artificial Analysis is the only source that ran both in one harness, and it measured K3.1 Pro 3 points ahead. On AA's older index the two sat one point apart. The deciding factors are practical: Sonnet 5.5 for cache-heavy sessions and first-party support, K3.1 Pro for open weights, returned reasoning and the long-horizon coding rows.

Where Fable 5.2 and Haiku 5.5 Fit

Fable 5.2 launched November 9, 2026, at $8 input and $40 output per million tokens, down from Fable 5.1's $10/$50, with cache reads at $0.20. Anthropic positions it for when "your evals on Claude Opus 5.5 at higher effort still fall short." No K3.1 Pro vs Fable 5.2 head-to-head exists yet; Moonshot's table uses Fable 5.1. One row allows a cross-check. On Humanity's Last Exam with tools, Moonshot and Anthropic publish the same numbers for Fable 5.1 (65.6) and Opus 5.5 (67.7), and Anthropic reports Fable 5.2 at 74.2 against K3.1 Pro's 70.0. The Fable 5.1 column above understates the current Fable. For the full matchup, see Kimi K3.1 vs Claude Fable 5.2.

Haiku 5.5 is a different class of model. Anthropic positions it "for high-volume, latency-sensitive tasks such as classification, extraction, and routing," at $0.10 input and $0.50 output for prompts up to 100K tokens. K3.1 Pro's list price is 20x that. If you want an open model for Haiku-shaped work, compare DeepSeek V4.1 Flash, which Morph serves as morph-dsv41flash at $0.12 input and $0.80 output per million tokens.

Pricing: The Worked Example

Sticker rates mislead in both directions. K3.1 Pro's always-on, max-default reasoning inflates its output tokens, and Claude's cheap cache reads deflate its input. Take one long-context agent turn: 500K tokens of input (a repo slice plus a long trace) and 20K tokens of output.

Nominal token counts at list price, cache-write charges excluded. Cells are rounded to the cent; totals are computed before rounding. Claude rates from Anthropic's pricing docs; Moonshot rates from Moonshot's pricing docs; Morph rates from Morph's rate card. A 500K prompt bills Haiku 5.5 at its over-100K rate.

Uncached, K3.1 Pro at Moonshot's list price costs $1.20, the same as Sonnet 5.5, half of Opus 5.5 ($2.40) and a quarter of Fable 5.2 ($4.80). With 90% of the input cached, Sonnet 5.5 pulls ahead at $0.35 against K3.1 Pro's $0.39, because it reads cache at half the rate. Morph charges more per token than Moonshot's list: $1.53 uncached and $0.54 cached on this turn, still below Opus 5.5. On Morph's standby tier (service_tier: "standby", 50% off) the same turns cost $0.77 and $0.27.

Cost per task on Terminal-Bench 4.0

Per-task cost is the number that settles agent budgets, because it folds in how many tokens each model burns. On Moonshot's chart K3.1 Pro scores slightly higher than Opus 5.5 for about 37% of the cost. The chart plots Opus 5.5's max-effort score (64.85); the table above uses its xhigh score (66.4), which Anthropic says is within noise. Dropping K3.1 Pro to high effort saves only about $1 per task and costs 13 points, so max is the right default for hard agent work. Artificial Analysis puts K3.1 Pro at $1.98 per Intelligence Index task.

Flat-rate plans: Claude Ultra vs Morph Ultra

If you pay for Claude Code by subscription, the plan matters more than the per-token math above. Anthropic and Morph both sell a monthly plan called Ultra: Claude Ultra at $200 and Morph Ultra at $200. They meter differently.

Claude Ultra caps total usage per week. Morph Ultra caps throughput per minute and never meters tokens, which suits agents that run all day. See Morph Ultra, and for the Claude side, Claude Code pricing and Anthropic API pricing.

Reasoning-Trace Transparency

When an agent takes a wrong turn on step 40 of a long run, the first question is why. The two vendors answer it differently.

Moonshot's API returns K3.1 Pro's reasoning in a reasoning_content field next to content. Thinking is always on and cannot be disabled; you choose low, high or max effort. The model card adds a requirement that matters for harness authors: pass the complete assistant message back into messages as-is, including reasoning_content and tool_calls. Harnesses that trim reasoning history between turns break that contract. With strict JSON output, parse content only, never reasoning_content.

Claude's current models hide the reasoning by default. On Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5, the thinking display setting defaults to "omitted", which returns thinking blocks with an empty thinking field. Set it to "summarized" and you get "a summary of Claude's full thinking process rather than the raw chain of thought." Anthropic's docs are explicit: "No display setting returns the raw chain of thought." If you audit agent decisions, debug long runs, or build interpretability tooling, K3.1 Pro gives you more to read. If you only need the answer, the difference does not matter.

Guardrails and Safety

Claude's safeguards show up even in coding benchmarks. Moonshot labels its Claude columns "w/ fallback", and Anthropic's system cards explain the term. On Terminal-Bench 4.0, requests flagged by Anthropic's safeguards were answered by a fallback model under the default server-side policy: 2.5% of requests, affecting 10% of trials, for Opus 5.5, and 1.2% of requests, affecting 1.5% of trials, for Sonnet 5.5. Anthropic also ships Fable 5.2 and Claude Mythos 5.2 as "the same underlying model," with Mythos 5.2 in limited availability.

Moonshot handled the open-weight risk in training instead. Its release post says it "identified and removed cyber-related data from both training stages" of K3.1 Pro, and that "K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months." Its own table bears that out: ExploitBench 74.7 against Opus 5.5's 91.0, and ExploitGym 20.7 against 34.5. For authorized security work, Claude is the stronger model. For teams whose risk is a false refusal mid-task, the fallback rate above is the number to test against your own workload. Moonshot offers Zero Data Retention to API customers.

Agent Harness and Ecosystem

Claude Code is Anthropic's first-party agent, included in every paid Claude plan, and the Claude models are tuned to drive it. Moonshot measured K3.1 Pro in that same harness: its Terminal-Bench 4.0, SWE Marathon, DeepSWE, ProgramBench and PostTrainBench numbers all come from K3.1 Pro running inside Claude Code. Moonshot serves an Anthropic-compatible endpoint at https://api.moonshot.ai/anthropic with streaming, tool use, image input, thinking and structured output, plus an OpenAI-compatible one at https://api.moonshot.ai/v1, and ships its own Kimi Code agent. Moving a Claude Code setup to K3.1 Pro is a base-URL change, not a harness migration.

Run Kimi K3.1 Pro inside Claude Code on Morph

Morph serves Kimi K3.1 Pro as morph-kimik31 over the Anthropic Messages API at https://api.morphllm.com and the OpenAI-compatible API at https://api.morphllm.com/v1. Point Claude Code at it with four environment variables, using DeepSeek V4.1 Flash for the small, fast model slot:

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-kimik31"
export ANTHROPIC_SMALL_FAST_MODEL="morph-dsv41flash"
claude

Or install the agent setup in one line: curl -fsSL https://www.morphllm.com/install.sh | bash && morph install. For Codex, add a provider with base URL https://api.morphllm.com/v1 and wire_api = "responses". Create a key at API keys; the same key serves Kimi K3.1, DeepSeek V4.1 Flash and GLM-5.3. Morph is also an OpenRouter provider for moonshotai/kimi-k3.1-pro. The old morph-kimik3 ID still works and now serves K3.1 weights, but use morph-kimik31.

Speed figures are approximate: about 115 output tokens per second and about 1.5 seconds to first token, the median of hourly medians over the 24 hours ending November 9, 2026, from the public charts at ultra.morphllm.com. Anthropic publishes relative latency only: Moderate for Fable 5.2 and Opus 5.5, Fast for Sonnet 5.5, Fastest for Haiku 5.5.

Morph also documents three features for long agent runs: dynamic tool loading, Agent Runs keyed by run_id, and session caching with prompt_cache_key or x-session-id.

Open weights and the Kimi K3.1 Pro License

Claude's weights are not published. K3.1 Pro's are: Moonshot posted moonshotai/Kimi-K3.1-Pro to Hugging Face on November 2, 2026, ungated, in native MXFP4 weights with MXFP8 activations from quantization-aware training. The Kimi K3.1 Pro License starts from an MIT-style grant to use, modify, fine-tune and redistribute, then adds two conditions. A company running a model-as-a-service business with more than $20 million in aggregate revenue over 12 months needs a separate agreement with Moonshot. A product with more than 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3.1-Pro" in its interface. Internal use is exempt.

Vision and Modality

Vision is close to a tie. Moonshot's model card lists text and image input for K3.1 Pro, with a 401M-parameter MoonViT-V2 vision encoder, and Moonshot's API also accepts video uploaded through its Files API. Anthropic states that "all current models support text and image input". Both output text only. One practical difference: Moonshot's API does not take public image URLs, so send images as base64 or an uploaded file ID.

Opus 5.5 reads images a few points better unassisted. Give both models a Python tool and K3.1 Pro catches up. If your agent reads screenshots and drives a desktop, OSWorld favors Claude by about 2 points.

Kimi K3.1 vs Claude: Pros and Cons

Verdict by Workload

Pick per workload. No single model wins every row.

• Hardest long-horizon work, budget secondary: Claude Fable 5.2 at $8/$40, which Anthropic recommends when Opus 5.5 at higher effort still falls short. See Kimi K3.1 vs Fable 5.2.
• Computer use, office tasks, security work, first-try correctness: Claude Opus 5.5 at $4/$20. It leads K3.1 Pro on OSWorld, AA-Briefcase, ProgramBench and ExploitBench in Moonshot's own table.
• Long-horizon coding loops that verify and retry: Kimi K3.1 Pro. It ties Opus 5.5 on Terminal-Bench 4.0 at about 37% of the per-task cost and leads on SWE Marathon and DeepSWE.
• Cache-heavy sessions at the $2/$10 tier: Claude Sonnet 5.5, for its $0.10 cache reads. Pick K3.1 Pro at the same price when open weights or visible reasoning matter.
• All-day Claude Code usage on a flat bill: K3.1 Pro on Morph Ultra at $200 per month with no token meter, or Claude Ultra at $200 with a weekly cap.
• Classification, extraction, routing: Claude Haiku 5.5, or DeepSeek V4.1 Flash for an open model.
• Everything at once: route by task. Keep Claude for the hard tier and send the high-volume tier to a cheaper model. See LLM routing.

For the full K3.1 Pro fact base, see Kimi K3.1 Pro: specs, benchmarks, pricing and the Kimi K3.1 API guide. For K3.1 Pro against the other open flagship, see GLM-5.3 vs Kimi K3.1. For the original K3 matchup against Fable 5 and Opus 4.8, see Kimi K3 vs Claude.

FAQ

Is Kimi K3.1 better than Claude?

It depends on which Claude and which task. Against Opus 5.5, Moonshot's table (K3.1 Pro run inside Claude Code) has K3.1 Pro ahead on SWE Marathon (51.7 vs 46.3) and DeepSWE (75.5 vs 74.2), level on Terminal-Bench 4.0 (65.7 vs 66.4), and behind on ProgramBench (31.5 vs 65.0), OSWorld (79.6 vs 81.8) and ExploitBench (74.7 vs 91.0). Artificial Analysis scores K3.1 Pro 46 on Intelligence Index v5 against 58 for Opus 5.5, and ranks it the top open-weights model. Against Sonnet 5.5, at the same $2/$10 list price, AA measured K3.1 Pro ahead on Terminal-Bench 4.0 (67% vs 64%). Fable 5.2 launched November 9, 2026 and has no head-to-head yet.

How much cheaper is Kimi K3.1 than Claude?

Moonshot lists K3.1 Pro at $2 input, $10 output and $0.20 cached per million tokens: half of Opus 5.5 ($4/$20), a quarter of Fable 5.2 ($8/$40), and the same as Sonnet 5.5 ($2/$10), which reads cache at $0.10. On a 500K-input, 20K-output turn with no caching, K3.1 Pro costs $1.20 at Moonshot's list against $2.40 for Opus 5.5 and $4.80 for Fable 5.2. Morph serves it as morph-kimik31 at $2.50 input, $0.29 cached and $14.00 output, 50% off on standby, or flat on the $200/month Ultra plan.

Kimi K3.1 vs Claude Opus 5.5: which is better for coding?

They split the coding rows. K3.1 Pro leads on SWE Marathon (51.7 vs 46.3) and DeepSWE (75.5 vs 74.2); Opus 5.5 leads on ProgramBench (65.0 vs 31.5) and PostTrainBench (43.8 vs 36.5); Terminal-Bench 4.0 is 65.7 vs 66.4, inside the 2.6-point standard error Anthropic reports for Opus 5.5. Moonshot's cost chart, with tbench.ai data, puts K3.1 Pro at $5.25 per task against $14.30 for Opus 5.5. Use K3.1 Pro where retries are cheap, Opus 5.5 where the first attempt has to land.

Kimi K3.1 vs Claude Sonnet 5.5: which should I use?

Same $2/$10 list price. Sonnet 5.5 reads cache at $0.10 against $0.20, so a 500K turn with 450K cached costs $0.35 on Sonnet and $0.39 on K3.1 Pro. On Terminal-Bench 4.0, Anthropic reports Sonnet 5.5 at 70.6% and Moonshot reports K3.1 Pro at 65.7%; Artificial Analysis, the only source that ran both, measured K3.1 Pro at 67% and Sonnet 5.5 at 64%. K3.1 Pro adds open weights and returned reasoning; Sonnet 5.5 adds cheaper cache reads and first-party support.

Can I use Kimi K3.1 in Claude Code?

Yes. Moonshot ran most of its coding benchmarks with K3.1 Pro inside Claude Code, and its API has an Anthropic-compatible endpoint at https://api.moonshot.ai/anthropic. On Morph, set ANTHROPIC_BASE_URL to https://api.morphllm.com, ANTHROPIC_AUTH_TOKEN to your Morph key, ANTHROPIC_MODEL to morph-kimik31 and ANTHROPIC_SMALL_FAST_MODEL to morph-dsv41flash, then run claude.

Can Kimi K3.1 do vision like Claude?

Yes, for images. Moonshot's model card lists text and image input with a 401M-parameter MoonViT-V2 encoder, and Moonshot's API also accepts video. Anthropic says all current Claude models take text and image input. Opus 5.5 leads on MMMU-Pro (87.7 vs 84.1); with a Python tool K3.1 Pro edges it (90.0 vs 89.8).

Does Claude show its reasoning like Kimi K3.1?

No. Moonshot's API returns K3.1 Pro's reasoning in a reasoning_content field, and the model always thinks. On Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5, Claude's thinking display defaults to omitted. Opting in to summarized returns a summary, and Anthropic's docs state that no display setting returns the raw chain of thought.

Should I switch from Claude to Kimi K3.1?

Not wholesale. Route by task. Keep Opus 5.5 or Fable 5.2 for computer use, security work and first-try correctness, and Haiku 5.5 for cheap classification. Send long-horizon coding loops that can verify and retry to K3.1 Pro, at half of Opus 5.5's list price. If you already work in Claude Code, testing K3.1 Pro takes four environment variables. On Morph that costs $2.50 input and $14.00 output per million tokens, or a flat $200 per month on Ultra.