Kimi K3.1 Pro vs Claude Fable 5.2: Benchmarks, Price, Open Weights

TL;DR: Kimi K3.1 vs Fable 5.2

Last updated November 9, 2026, the day Claude Fable 5.2 launched.

Claude Fable 5.2 is the stronger model on the three benchmarks where both vendors publish a score on the same version, and Kimi K3.1 Pro costs a quarter as much per input and output token at Moonshot's list price. Anthropic launched Fable 5.2 on November 9, 2026 at $8/M input and $40/M output, with a 1M-token context and closed weights. Moonshot's Kimi K3.1 Pro is a 2.8T-parameter MoE (104B active) with a 1M context, open weights on Hugging Face, and a $2/$10 list price. On Humanity's Last Exam with tools, Anthropic reports 74.2 for Fable 5.2 and Moonshot reports 70.0 for K3.1 Pro. On Terminal-Bench-Science 0.1 it is 74.6 to 38.1. In Moonshot's launch table, K3.1 Pro beats Fable 5.1, not 5.2, on 28 of 40 rows. No published run tests both models in the same harness. Pick Fable 5.2 for the hardest long-horizon tasks. Pick K3.1 Pro for volume, open weights and output-heavy agent loops. On Morph it runs as morph-kimik31 at $2.50/$14.00 per million tokens.

For the previous generation, see the original Kimi K3 vs Claude Fable 5 comparison.

The 30-Second Verdict

Fable 5.2 is ahead wherever a like-for-like number exists. The open question is how far ahead, and whether the gap is worth 4x Moonshot's per-token price and closed weights for your workload.

• Hardest long-horizon work? Fable 5.2. Anthropic reports 74.6 on Terminal-Bench-Science 0.1, up from Fable 5.1's 52.6. Moonshot reports 38.1 for K3.1 Pro on the same version.
• Paying for output tokens at volume? K3.1 Pro. $10/M output at Moonshot and $14.00/M on Morph, against $40/M. Both models think on every request, and reasoning bills as output.
• Cache-heavy agent loops? Closer than the sticker says. Fable 5.2 cache hits cost $0.20/M, the same as Moonshot's K3.1 Pro cached rate. The gap lives on output and fresh input.
• Need weights, on-prem, or zero data retention? K3.1 Pro. Anthropic says Fable 5.2 requires 30-day data retention unless it expressly authorizes zero data retention.
• Security research? Neither out of the box. Fable 5.2's classifiers hand blocked cyber requests to Opus 4.8. Moonshot removed cyber data from K3.1 Pro's training.
• Long context? A tie at 1M tokens. Fable 5.2 caps output at 128K.

Kimi K3.1 Pro vs Claude Fable 5.2, Spec-by-Spec

Both models take 1M tokens of context, think on every request, and accept images. They split on weights, price, data retention, and how much each vendor discloses about the model itself.

Kimi K3.1 Pro is a post-trained version of Kimi K3 with, in Moonshot's words, "substantially larger post-training compute." The one architecture change: Moonshot replaced K3's full-attention Gated MLA layers with Gated DeepSeek Sparse Attention layers for more efficient long-context modeling. The full K3.1 Pro breakdown is on the Kimi K3.1 page.

Fable 5.2 replaces Fable 5.1 at a lower price, $8/$40 down from $10/$50. Anthropic says Fable and Mythos "are the same underlying model"; Mythos 5.2 is the version with restricted access. Anthropic's docs tell most teams to start elsewhere: "Use Claude Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short."

Kimi K3.1 vs Fable 5.2 Benchmarks: What Is Actually Comparable

Every score in this section is vendor-reported unless it is marked as Artificial Analysis. Moonshot ran K3.1 Pro at max effort and ran its coding benchmarks in Claude Code. Anthropic reports each model at its highest available effort. Neither launch post reports a SWE-bench score, so there is no SWE-bench comparison to make.

Same benchmark, same version: K3.1 Pro vs Fable 5.2

Three benchmarks appear in both vendors' launch tables at the same version. On Humanity's Last Exam with tools, the Fable 5.1, Opus 5.5 and GPT-6 Astra figures match across both tables, so the two rows line up.

Fable 5.2 leads all three. On Humanity's Last Exam with tools, K3.1 Pro's 4.4-point lead over Fable 5.1 turns into a 4.2-point deficit against Fable 5.2. On Terminal-Bench-Science 0.1, Fable 5.2 nearly doubles K3.1 Pro's score. Two caveats. Fable 5.1's Terminal-Bench-Science number differs by 12.6 points between the two tables (40.0 vs 52.6), which shows how much run conditions move these scores. And Elo ratings are snapshots taken at different times.

Moonshot's table: Kimi K3.1 Pro vs Fable 5.1

Across the 40 rows where Moonshot's table has a score for both models, K3.1 Pro is ahead of Fable 5.1 on 28. Its clearest wins are knowledge-work and automation agents: Automation Bench-AA 73.3 vs 59.4, GDPval-AA 1846 vs 1754 Elo, OfficeQA Pro v2 64.4 vs 54.5, and Moonshot's internal hallucination test at 8.3 vs 23.3. Fable 5.1 keeps ProgramBench (53.5 vs 31.5), PostTrain Bench (44.6 vs 36.5), Humanity's Last Exam without tools (57.6 vs 51.1), Frontier Finance (60.8 vs 58.6), OSWorld and both cyber rows. Read it as a vendor table: Moonshot chose the rows, several are internal, and for many rows the competitor numbers come from published reports, leaderboards or Artificial Analysis rather than Moonshot's own runs.

Anthropic's table: Fable 5.2 vs Fable 5.1

This is the table that matters for reading Moonshot's. Fable 5.2 gains 22 points over Fable 5.1 on Terminal-Bench-Science 0.1 and 16.7 on FrontierCode 2.0, so a model that beat Fable 5.1 by a few points has not necessarily beaten Fable 5.2. Two rows look shared but are not. Anthropic's OSWorld row uses the offline set and puts Fable 5.1 at 83.5, while Moonshot's OSWorld V2.1 row puts it at 80.7. Anthropic's AutomationBench has Fable 5.1 at 31.4; Moonshot's Automation Bench-AA has it at 59.4. Different variants, so we keep them in separate tables.

Independent numbers

• Artificial Analysis Intelligence Index v5 (October 27, 2026): K3.1 Pro scores 46, the top open-weights model. Claude Opus 5.5 (max) leads at 58. The v5 article predates Fable 5.2.
• Artificial Analysis Intelligence Index, pre-v5 (October 23, 2026): K3.1 Pro launched 4th at 55. Scores on the two index versions are not comparable.
• Terminal-Bench 4.0: Artificial Analysis's own run measured K3.1 Pro at 67%, the highest of any model at the time, ahead of Claude Sonnet 5.5 (64%). Moonshot reports 65.7 in Claude Code. Anthropic's Fable 5.2 launch post has no Terminal-Bench 4.0 row, and Anthropic's Opus 5.5 system card puts Fable 5.1 at 55.8, against 57.9 in Moonshot's table.
• GDPval-AA: K3.1 Pro placed 3rd at 1,846 Elo, behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857).

Every Fable 5.2 number on this page is Anthropic-reported. It launched the day this page was updated, so the independent runs above all predate it.

Kimi K3.1 vs Fable 5.2 Pricing: per Token and per Task

At list price, Fable 5.2 costs 4x Moonshot's K3.1 Pro rate on input ($8 vs $2) and on output ($40 vs $10). Against Morph's K3.1 Pro rates it is 3.2x on input and 2.9x on output. Morph's per-token price sits above Moonshot's own list; the reasons to run K3.1 Pro on Morph are speed, the flat plan and standby pricing, covered below.

Cached input is a tie. Anthropic prices Fable 5.2 cache hits at 0.025x base input, $0.20/M, the same as Moonshot's K3.1 Pro cached rate and below Morph's $0.29/M. Anthropic bills cache writes at $10/M for a 5-minute cache and $16/M for a 1-hour cache; Moonshot bills them at $2.00/M and $4.00/M. In a long agent loop where most input is cache hits, the input side of the bill is close to even. The gap lives on fresh input and on output.

Output is where thinking lands. Both models think on every request, and reasoning tokens bill as output. Artificial Analysis measured K3.1 Pro at 55K output tokens per Intelligence Index task and $1.98 per task. Anthropic says Fable 5.2 at Medium effort "can accomplish challenging tasks with fewer tokens than Fable 5.1," but no one has published Fable 5.2 tokens per task yet. Until someone does, per-task cost is something you measure on your own traces.

Cost per solved task. Moonshot's Terminal-Bench 4.0 cost chart, built on tbench.ai data, puts K3.1 Pro at $5.25 per task for 65.7% at max effort and $4.16 per task for 52.8% at high effort. The same chart puts Opus 5.5 at $14.30 per task for 64.85%. Fable 5.2 is not on it.

Subscriptions. If you reach Fable through a Claude plan, Claude Max ($100/month) includes Fable for up to half your weekly limit. Morph's Ultra plan is $200/month flat for unlimited Kimi K3.1 Pro and DeepSeek V4.1 Flash with no token meter.

Open Weights vs Closed Weights

The Kimi K3.1 Pro weights are on Hugging Face at moonshotai/Kimi-K3.1-Pro with no access gate. Moonshot had promised them by November 4, 2026; the repo was created November 2. The checkpoint holds 2,780,110,363,392 parameters with MXFP4 weights and MXFP8 activations from quantization-aware training, and Moonshot recommends vLLM, SGLang or TokenSpeed to run it. Fable 5.2 has closed weights. You reach it through the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or Claude Platform on AWS.

What the Kimi K3.1 Pro License allows

The base grant is MIT-style: use, copy, modify, merge, publish, distribute, sublicense and sell, plus the right "to run, deploy, fine-tune, or otherwise modify the Software." Two clauses kick in at scale:

• Model-as-a-Service above $20M. If you run a business that gives third parties inference or fine-tuning access and your aggregate revenue, with affiliates, exceeds $20M over any consecutive 12 months, you need a separate agreement with Moonshot before commercial use.
• Attribution at 100M users. Products with more than 100M monthly active users or more than $20M in monthly revenue must display "Kimi K3.1-Pro" prominently in the user interface.

The license exempts internal use and use through Moonshot's official products or certified inference partners. Read the LICENSE file before you build a business on it.

What closed weights cost you

Anthropic's Fable 5.2 model page says it "requires 30-day data retention and isn't available under zero data retention unless expressly authorized by Anthropic." The launch post adds that eligible customers can continue with zero data retention before implementing Anthropic's safety monitoring. Moonshot offers zero data retention to K3.1 Pro API customers, and self-hosting keeps prompts on your own hardware. Anthropic's docs list Fable 5.2's retirement as not sooner than November 9, 2027. Published weights have no retirement date.

Agent-Harness Differences That Bite in Production

These change how you build the loop around the model, which matters more day to day than any single benchmark row.

• K3.1 Pro breaks quietly if your harness drops reasoning. Moonshot requires "the complete assistant message returned by the API to be passed back to messages as-is," including reasoning_content and tool_calls. A framework that keeps only content breaks that requirement. Fix that before you benchmark K3.1 Pro.
• Default effort is not benchmark effort. Fable 5.2's API defaults to high effort, and Claude apps and Claude Code default to Medium. Anthropic's table reports each model at its highest available effort, such as xhigh on FrontierCode 2.0. K3.1 Pro defaults to max on Moonshot's API, the same setting as its benchmarks, and offers low, high and max. Moonshot fixes temperature=1.0 and top_p=0.95 on its API, so omit them.
• Fable 5.2 can answer with a different model. Its safety classifiers block high-risk cyber activity, including penetration testing, exploit generation and binary-based vulnerability work, plus most research-biology tasks. Blocked cyber queries get an answer from Opus 4.8 and biology queries from Opus 5. Anthropic also adds transparent fallbacks to Opus 5 for a small set of topics, including kernel development on certain nonstandard accelerators. Moonshot's table labels its Claude numbers "w/ fallback".
• Both run in Claude Code. Moonshot ran K3.1 Pro's Terminal-Bench, DeepSWE and ProgramBench numbers in Claude Code. Moonshot exposes an Anthropic-compatible endpoint at https://api.moonshot.ai/anthropic, and Morph serves the Anthropic Messages API at https://api.morphllm.com, so swapping Fable 5.2 for K3.1 Pro is an environment-variable change.
• K3.1 Pro tool and output controls. Moonshot's API supports tool_choice="required", strict JSON Schema output, dynamic tool loading through a system message, and partial mode for prefix continuation. Vision input takes base64 or a file ID, not a public image URL.

Moonshot says it removed cyber-related data from both training stages of the open-weight release. In its own table K3.1 Pro scores 74.7 on ExploitBench and 20.7 on ExploitGym, against Fable 5.1's 83.0 and 30.4. Security teams that need Anthropic's model unblocked can apply for Claude Mythos 5.2 through the Cyber Verification Program.

Which One to Pick, by Workload

Many teams will run both: K3.1 Pro as the default for bulk build work, Fable 5.2 for the tasks that keep failing downstream. Fable 5.2 earns its slot when the cheaper model would retry its way past a 4x price gap. If the real alternative is the rest of the Claude lineup, Anthropic's default recommendation is Opus 5.5 at $4/$20; see Kimi K3.1 vs Claude for that comparison and GLM-5.3 vs Kimi K3.1 for the open-weight alternative.

Running Kimi K3.1 Pro on Morph

Morph serves Kimi K3.1 Pro as morph-kimik31 at $2.50/M input, $0.29/M cached input and $14.00/M output with a 1M context. The old ID morph-kimik3 still works and now serves the K3.1 weights; use morph-kimik31 in new code.

• Speed. About 115 tokens per second of output and about 1.5 s to first token, from the public hourly charts at ultra.morphllm.com (median of hourly medians, 24 hours ending November 9, 2026).
• Standby. Send service_tier: "standby" and K3.1 Pro bills at 50% off.
• Ultra. $200/month flat for unlimited K3.1 Pro and DeepSeek V4.1 Flash. Each subscription gets 24,290 uncached input tokens and 6,162 output tokens per minute on K3.1 Pro, and the plan is capped at 1,000 subscriptions.
• One key. The same key calls K3.1 Pro, DeepSeek V4.1 Flash (morph-dsv41flash) and GLM-5.3 (morph-glm53-744b).

Because Morph serves the Anthropic Messages API, Claude Code runs K3.1 Pro with four environment variables:

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 let the Morph CLI set up your coding agent:

curl -fsSL https://www.morphllm.com/install.sh | bash && morph install

The same key works on the OpenAI-compatible endpoint. Codex CLI connects with a provider whose base_url is https://api.morphllm.com/v1 and wire_api = "responses"; see the Codex provider configuration guide.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-kimik31",
    messages=[
        {"role": "user", "content": "Refactor this function to remove the nested loop."},
    ],
)
print(resp.choices[0].message.content)

For agent workloads, Morph documents three K3.1 Pro features: dynamic tool loading, Agent Runs tagged with a run_id, and session caching keyed by prompt_cache_key or an x-session-id header. Morph is also an OpenRouter provider for moonshotai/kimi-k3.1-pro. Create a key on the API keys page.

Kimi K3.1 Pro vs Claude Fable 5.2: The Trade-Offs

Frequently Asked Questions

Is Kimi K3.1 better than Claude Fable 5.2?

On the evidence available on November 9, 2026, no. Fable 5.2 leads on all three benchmarks where both vendors report the same version: Humanity's Last Exam with tools (74.2 vs 70.0), Terminal-Bench-Science 0.1 (74.6 vs 38.1) and AA-Briefcase v1.1 (1785 vs 1678 Elo). Each number comes from its own vendor, and no published run tests both models in the same harness. In Moonshot's launch table, Kimi K3.1 Pro beats Fable 5.1, the previous version, on 28 of 40 rows. K3.1 Pro's case is price ($2/$10 per million tokens at Moonshot vs $8/$40), open weights, and zero data retention on Moonshot's API.

How much does Kimi K3.1 cost compared with Fable 5.2?

Claude Fable 5.2 lists at $8/M input, $0.20/M cache hits and $40/M output. Kimi K3.1 Pro lists at $2/M input, $0.20/M cached input and $10/M output at Moonshot: a quarter of Fable 5.2 on input and output, and the same on cached input. Morph serves K3.1 Pro as morph-kimik31 at $2.50/M input, $0.29/M cached and $14.00/M output, with 50% off on the standby tier and a flat $200/month Ultra plan.

Why does Moonshot compare Kimi K3.1 Pro with Fable 5.1 instead of Fable 5.2?

Timing. Moonshot launched K3.1 Pro on October 23, 2026, and Anthropic released Fable 5.2 on November 9, 2026. Moonshot's table uses Claude Fable 5.1 (max, with fallback). Anthropic's own table shows Fable 5.2 well ahead of 5.1, for example 74.6 vs 52.6 on Terminal-Bench-Science 0.1 and 74.2 vs 65.6 on Humanity's Last Exam with tools, so K3.1 Pro's wins over Fable 5.1 do not carry over to Fable 5.2.

Can I self-host Kimi K3.1 Pro? Does Fable 5.2 have open weights?

Kimi K3.1 Pro: yes. The weights are on Hugging Face at moonshotai/Kimi-K3.1-Pro (repo created November 2, 2026, not gated) under the Kimi K3.1 Pro License, an MIT-style grant with two scale clauses: Model-as-a-Service businesses above $20M in revenue over 12 months need a separate agreement with Moonshot, and products above 100M monthly active users or $20M monthly revenue must display the model name. The checkpoint is MXFP4 with quantization-aware training, and Moonshot recommends vLLM, SGLang or TokenSpeed. Claude Fable 5.2: no. The weights are closed, and you reach it through the Claude API or the cloud platforms Anthropic lists, including Amazon Bedrock, Google Cloud and Microsoft Foundry.

Can Claude Code use Kimi K3.1?

Yes. Moonshot ran K3.1 Pro's own coding benchmarks in Claude Code. On Morph, set ANTHROPIC_BASE_URL to https://api.morphllm.com, ANTHROPIC_AUTH_TOKEN to your Morph API key, ANTHROPIC_MODEL to morph-kimik31 and ANTHROPIC_SMALL_FAST_MODEL to morph-dsv41flash. Moonshot also runs an Anthropic-compatible endpoint at https://api.moonshot.ai/anthropic.

What changed from Claude Fable 5.1 to Fable 5.2?

The price fell from $10/$50 to $8/$40 per million tokens, and Anthropic says Fable 5.2 generates outputs up to 30% faster than Fable 5.1. On Anthropic's own table, FrontierCode 2.0 rose from 32.4 to 49.1 (both at xhigh effort), FrontierSWE v2 from 56.3 to 69.9, and Terminal-Bench-Science 0.1 from 52.6 to 74.6. Fable 5.2 has a 1M-token context and 128K max output. Fable 5.1 is now a legacy model.

Which is better for security work, Kimi K3.1 or Fable 5.2?

Neither by default. Fable 5.2 ships safety classifiers that block high-risk cyber activity, including penetration testing and exploit generation, and blocked requests get an answer from Opus 4.8 instead. Security professionals can apply for Claude Mythos 5.2, the same underlying model, through Anthropic's Cyber Verification Program. Moonshot removed cyber-related data from K3.1 Pro's training for the open-weight release and says its cyber capability remains behind frontier proprietary models that have been public for at least three months. In Moonshot's table, K3.1 Pro scores 74.7 on ExploitBench against Fable 5.1's 83.0.

How do Kimi K3.1 and Fable 5.2 compare on context window?

Both take 1M tokens. Fable 5.2 caps output at 128K tokens. Kimi K3.1 Pro on Moonshot's API defaults to 131,072 output tokens and allows up to 1,048,576. Moonshot prices K3.1 Pro flat across the window, and Anthropic includes the full 1M window at standard pricing for Fable 5.2. Morph serves morph-kimik31 with a 1M context.