FULL STORY

Claude Fable 5.2: From Leak to Launch and Benchmarks

After early leaks, Anthropic officially launched Claude Fable 5.2 and Mythos 5.2 on Nov 10. Reviews poured in same day: Shumer called it the best model yet despite strict guardrails, Pro users got limited free access, and it topped the AA Intelligence Index amid cost concerns.

2026-11-10 ~ 2026-11-10 · 6 episodes · 104 posts

Episode 1 · Claude Fable 5.2 Benchmarks Leak, Sparking Hoax Debate (2026-11-10, 18 posts)

Claude Fable 5.2 launched on November 10, and official benchmark comparison tables shared by multiple users (markk, accourtt, MoneyAd3945, etc.) show the new model leading Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon on nearly all of roughly 12-13 evaluations, with major gains over the previous Fable 5.1.

Confirmed

  • On the agentic coding benchmark FrontierCode 2.0: Fable 5.2 scores 49.1%, versus 32.4% for Fable 5.1, 45.5% for Opus 5.5, and 42.8% for GPT-6 Astra, ranking first
  • FrontierSWE v2 reaches 69.9%
  • Robotics scores jump from 27.6% to 64.9%
  • minchoi notes Fable 5.2 costs roughly half as much as Opus
  • The benchmark table spans comparisons across Fable 5.1, Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon

Why it matters

  • FrontierCode 2.0 measures whether a model's code changes would be merged by open-source project maintainers, making it an agentic coding metric close to real-world engineering ability; Fable 5.2 topping it at half the cost is directly relevant to anyone invested in coding automation
  • Benchmark data shared by multiple independent accounts corroborate each other, suggesting wide circulation, but all are reposts of the official table, with no deeper independent evaluation yet

Episode 2 · Anthropic Launches Claude Fable 5.2 and Mythos 5.2 (2026-11-10, 67 posts)

On November 10, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, positioning them as "collaborators" for tackling the most challenging technical problems. They beat Anthropic's own flagship Opus 5.5 on nearly every capability dimension in official evaluations (coding, reasoning, vision), with output up to 30% faster than 5.1 and API pricing cut by 20%. The release represents a concentrated push into high-intensity agentic work and frontier research, accompanied by demos including robotic control and biological structure prediction.

Confirmed

  • Pricing: Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, a 20% cut from the previous $10/$50; fast mode delivers 2.5x faster output at double the price, enabled via "speed": "fast" in the API and switchable in the Claude Code model picker
  • Performance: On the official FrontierCode 2.0 test, the High effort tier matches Opus 5.5 at Max effort at roughly half the per-task cost; the Xhigh tier sets the test's highest score; up to 30% faster than 5.1
  • Effort tiers: API defaults to high, the Claude app defaults to medium; xhigh is recommended for the hardest long-running agentic tasks
  • Availability and perks: Fable 5.2 is live on Claude Platform and Claude Code; long-standing Ultra and Team Premium subscribers can get 1-3 Fable usage resets depending on subscription length (under 6 months / 6-12 months / over 12 months), claimable before March 8
  • Anti-distillation: preserved thinking extends to Fable 5.2 on all API accounts; editing prior context may return a 400 error, and Anthropic recommends using compaction or context editing instead
  • The two-model relationship: Fable and Mythos share the same lineage. Fable, available to all users, ships with strong safety guardrails; cybersecurity and biomedical institutions can request Mythos through the Life Sciences and Cyber Verification programs
  • Robotics: Anthropic says Fable 5.2 outperforms other models and most dedicated robot control systems on the simulated dual-arm robot benchmark RoboDojo, completing tasks it was never explicitly trained on; Alex Albert relayed this demo
  • Science: Mythos 5.2 found a new method for predicting entire viral capsid structures, correctly modeling 28 of 34 capsids in tests, which could enable cheaper design of gene therapy vectors
  • Other official demos included orchestrating over a thousand subagents running for 8 straight days to write a bootable operating system from an empty folder

Not yet confirmed

  • Claims like "nearly全面 surpassing Opus 5.5" and "world's strongest model" come from Anthropic's own benchmarks and pre-release testing, with no independent third-party verification yet
  • X user @synthwavedd's claim that OpenAI will follow within days with "6.2 Astra" is rumor with no official backing

Why it matters

  • The 20% price cut plus speed gains directly change the cost-effectiveness calculus for high-intensity agent tasks; Anthropic recommends most everyday coding still start with Opus 5.5, reserving Fable/Mythos for the hardest open-ended problems, creating clear product tiering
  • Mythos's gated access for cybersecurity and biomedical institutions via verification programs, along with results like protein capsid structure prediction, shows frontier models moving deeper into high-stakes research, and its access governance deserves attention

47 more related posts →

Episode 3 · Shumer Reviews Fable 5.2: Smartest Model Ever, But Heavy Guardrails (2026-11-10, 3 posts)

Matt Shumer's review calls Fable 5.2 the best model ever—smart enough to trust with running most of his business—while criticizing its strict guardrails and noting the max effort mode scored lower.

Episode 4 · Anthropic offers Claude Pro users free limited trial of Fable 5.2 (2026-11-10, 2 posts)

Anthropic is offering Claude Pro subscribers a limited free trial of Fable 5.2, counting at most 50% toward weekly usage until November 16, along with a bonus usage reset.

Episode 5 · Fable 5.2 Wins Praise for Open-Ended Tasks (2026-11-10, 2 posts)

Anthropic's Alex Albert and Perplexity CEO Arav Srinivas both praise Claude Fable 5.2, saying it clearly outperforms Opus 5.5 on large open-ended tasks and can run for hours on whole projects.

Episode 6 · Claude Fable 5.2 Tops Artificial Analysis Intelligence Index at 65, but Per-Task Cost Draws Skepticism (2026-11-10, 12 posts)

On November 10, Artificial Analysis launched an evaluation page for Anthropic's Claude Fable 5.2 and updated its Intelligence Index leaderboard: the model scored 65 on Intelligence Index v5, ranking first among all 117 models, clearly ahead of second-place Gemini 4 Carbon (60) and Claude Opus 5.5 (59). X users @Angaisb, @haider1, and several Reddit users reshared screenshots of the leaderboard, and the results spread quickly through the community.

Confirmed

  • Artificial Analysis officially (@ArtificialAnlys) published the Claude Fable 5.2 evaluation page, confirming its score of 65 on Intelligence Index v5, first among 117 models
  • The leaderboard is Intelligence Index v5.0.1, comprising 9 evaluations; Reddit user @AbbreviationsBest858 posted the full ranking, which includes benchmarks such as AA-Briefcase v1.1, AutomationBench-AA, and Terminal-Bench
  • The next-ranked models are Gemini 4 Carbon (60) and Claude Opus 5.5 (59), an order confirmed by screenshots from Reddit users @likeastar20 and @haider1; GPT-6 series models are also on the list but with lower scores
  • The family includes 5 models and supports text+image input

Why it matters

  • The Intelligence Index is a widely cited community benchmark for overall model capability; Fable 5.2 taking the top spot by a clear margin means Anthropic's lead in the latest round of model competition is now backed by third-party evaluation
  • The leaderboard also shows the relative positions of leading models such as Gemini 4 Carbon, Claude Opus 5.5, and GPT-6, offering a reference for the current frontier-model landscape