Claude Fable 5.2: Anthropic's Long-Horizon Model Gets Cheaper and Pulls Away

Anthropic released Claude Fable 5.2 on November 9, 2026: the generally available, safeguarded twin of Mythos 5.2. List price is $8 / $40 per million input / output tokens (20% below Fable 5.1), with a 1M context, 128K max output, and an optional fast mode at $16 / $80. Outputs are up to 30% faster than Fable 5.1.

The case for Fable 5.2 is long-horizon work. On Anthropic's self-reported numbers, Remote Labor Index nearly doubles Opus 5.5 (41.3% vs 21.3%), with large gaps on FrontierSWE v2, Terminal-Bench- Science, AutomationBench, and RoboDojo Sim. On bounded coding the lead is thin or negative, and the model still costs twice Opus 5.5.

Switch from Fable 5.1 now. Move up from Opus 5.5 only for multi-hour agents and the hardest open-ended work, which is what Anthropic itself recommends. Security and biology work belongs on Mythos 5.2.

At a Glance

See Anthropic's model overview and system card for the primary sources behind this page.

Fable vs Mythos

Anthropic says plainly: Claude Fable and Claude Mythos are the same underlying model. Fable 5.2 carries strong cybersecurity and biology safeguards so it can ship to all users. Mythos 5.2 shares Fable 5.2's specifications and pricing and is available only to organizations verified through Anthropic's Cyber and Life Sciences Verification Programs.

Classifiers block high-risk cyber activity (penetration testing, exploit generation, binary-based vulnerability discovery) and most research-biology tasks. Blocked requests are answered by Opus 4.8 (cyber) or Opus 5 (biology). In Claude apps this is automatic; on the API developers opt in to fallbacks. Anthropic reports about 55% fewer blocks than Fable 5.1 in early testing (about 60% fewer in cybersecurity and 40% fewer in biology), while saying the safeguards remain as robust as Fable 5.1's.

Anthropic reports Mythos 5.2 at 89.6% on its internal CVE Rediscovery evaluation (Mythos 5.1 78.3%, Opus 5.5 81.3% without safeguards) and notes Fable 5.2's safeguards "will block most cybersecurity- related tasks." For the rest of the Mythos picture, see our Mythos 5.2 launch post.

Anthropic's own routing advice: start with Claude Opus 5.5 ("our strongest daily driver") for most coding and knowledge tasks, then move up to Fable or Mythos, or down to Sonnet 5.5 or Haiku 5.5. Docs say move to Fable 5.2 when your evals on a Claude Opus model at higher effort still do not meet your requirements. Fable 5.2 can also act as an Advisor or Orchestrator over smaller models.

What's New

Five practical changes versus Fable 5.1.

Price and speed

List price drops from $10 / $50 to $8 / $40 per million tokens (20% lower). Cache reads fall from $0.25 to $0.20. Outputs are up to 30% faster. A new optional fast mode reaches up to 2.5x output tokens per second at $16 / $80.

Long-horizon capability

The large gains sit on long-horizon rows. Remote Labor Index moves from 17.9% to 41.3%. FrontierCode 2.0 moves from 32.4% to 49.1% at best effort. The benchmark section covers the rest.

Fewer safeguard blocks and stronger anti-distillation

About 55% fewer blocks than Fable 5.1 in early testing, with Anthropic calling this release "the strongest protections against distillation of any model we've released," reducing data harvested by distillers by 90% to 99%. Training gains in accelerator kernel optimization and LLM architecture research were limited to about Opus 5.5 level, with a narrow set of such requests falling back to Opus 5.

Breaking changes

Four API breaks matter for migrations (see the migration guide): editing earlier turns now invalidates thinking blocks on every account (400 unless prefix_mismatch_behavior: "drop_block" with the thinking-binding-controls-2026-08-01 beta header); only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking blocks; the older computer_20251124 tool returns a 400 everywhere except Amazon Bedrock (use computer_toolset_20260801); and with the advisor tool, a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor. Prompt lines asking for less thinking no longer suppress thinking at medium and above. Long tasks may end a turn early at low and medium effort. At xhigh and max it may spawn subagents to review its work, which adds cost.

Showcase (Anthropic claims)

Anthropic says a team of 1,000 agents formalized Fermat's Last Theorem in Lean in 30 hours with minimal human intervention, using one-third of the action tokens of its first formalization (which took 11 days) and producing a proof seven times shorter; Opus 5.5 in the same setup completed 26% in 18 hours. The system card caveat: this used "an earlier snapshot of the model" and built "on definitions and smaller proven results from an earlier formalization," and each model got a single run.

One engineer had it build a prototype operating system from an empty folder: over 1,000 subagents, eight days, boots on a laptop. It also designed three working electronic devices (voice-assistant hub, USB speakerphone, handheld virtual pet) from a one-paragraph prompt, with Diode manufacturing the boards and no electrical modifications needed. Uber's Andrew Mager: "It works until the job is actually complete instead of stopping at a plausible checkpoint."

Benchmarks

Anthropic reports Fable 5.2, Fable 5.1 and Opus 5.5 with safeguards on, at each model's highest available effort unless noted. Most rows are max effort; FrontierCode 2.0 uses each model's best effort (xhigh for both Fable models); Remote Labor Index uses default settings (high effort); ARC-AGI-3's headline is at high effort. Several rows were run by outside benchmark owners (Cognition, Proximal, Zapier, Scale AI, Cursor, ARC Prize Foundation). Comparator scores come from the same Anthropic tables and can differ from those models' own launch posts. In Anthropic's Table 8.1.A, Fable 5.2 scored highest on 14 of 16 rows; the two misses were GDPval-AA and AA-Briefcase Elo, both won by Opus 5.5. All scores below are self-reported by Anthropic.

Long-horizon and agentic work

Self-reported by Anthropic. This is where the price is earned. Remote Labor Index nearly doubles Opus 5.5.

Remote Labor Index used default config (high effort, 200k context, 32k tokens per response); 99 of 240 freelance projects were judged at least as good as a human professional's. FrontierSWE v2: 34 ultra-long tasks, 20-hour budget per trial, max effort; highest or tied on 20 of 34. Terminal-Bench-Science standard error is about 4.3 points (Mythos 5.2 without safeguards 75.3%). Gemini 4 Carbon leads AutomationBench. BenchCAD without tools: 75.4 / 36.3 / 49.0.

Coding

Self-reported by Anthropic. On bounded coding Fable 5.2 is ahead of Opus 5.5 by a few points at most, and behind on ProgramBench. Coding alone does not justify 2x the price.

At max effort FrontierCode 2.0 scored 45.7%, and 44.4% if 22 of 660 timed-out attempts count as zero, which is below Opus 5.5. That is a single Cognition result with no uncertainty interval. Terminal-Bench 4.0 standard errors are about 1.8 (Fable 5.2) and 2.6 (Opus 5.5) points, so the gap to Opus is within noise; Mythos 5.2 without safeguards scored 70.5%. SWE-bench Multilingual 96.6% / 89.1% / 93.9%; Multimodal 74.1% / 54.7% / 61.4%. Cursor attributes part of CursorBench to prompt and harness improvements; at high effort Fable 5.2 scored 56.9%. On ProgramBench, Fable 5.2 ran out of its 1M context before submitting on 68% of tasks; when the harness showed remaining context, it scored 91.6% (Fable 5.1 86.8%, Opus 5.5 91.1%).

Reasoning, search, computer use, vision

Self-reported by Anthropic.

HLE without tools: 73.2% / 60.9% / 64.4%. ARC-AGI-3 at max effort is 65.66% on the Standard harness (verified by the ARC Prize Foundation); on the Provider Adapter harness Fable 5.2 scored 97.72% (high) vs GPT- 6 Astra 98.6% (max), so Astra leads on that harness. OSWorld strict pass: 54.6% / 47.6% / 53.2%. Chartography with tools: 91.6% / 88.4% / 89.0%.

Knowledge work

Self-reported by Anthropic. On pairwise-judged knowledge work Opus 5.5 is still preferred; Fable 5.2 wins on rubric pass rate and long- horizon delivery (Remote Labor Index).

Honest weak spots

ProgramBench regresses versus Fable 5.1. Opus 5.5 leads on GDPval-AA and AA-Briefcase Elo. HealthBench raw score (58.8%) sits below Fable 5.1 (63.9%) and Opus 5.5 (64.3%). Gemini 4 Carbon leads AutomationBench. GPT-6 Astra leads on the ARC-AGI-3 Provider Adapter harness.

Effort and Cost

Effort is the hidden lever. On FrontierCode 2.0 (system card section 8.4), high effort roughly matches Opus 5.5's best for about half the cost per task; xhigh is the peak; max costs about 3x xhigh and scores lower. Anthropic says most of the gain is already present at high, use xhigh for the hardest long-running work, and reserve max for measuring the ceiling. Medium is the default in Claude apps and Claude Code.

Other Anthropic efficiency claims: on Humanity's Last Exam with tools, Fable 5.2 at low effort matches Opus 5.5's top score for less than one-tenth the cost per task; on Chartography, medium effort (66.4%) beats Opus 5.5 at max (64.4%) for less than one-tenth the cost; on WANDR, medium effort matches Opus 5.5's top score for about 40% of the cost. Takeaway: per-token price is 2x Opus 5.5, but per-task cost can be lower at the right effort. Benchmark your own tasks.

Pricing and Availability

Prices below are from Anthropic's pricing page.

Fable cache reads cost the same $0.20 as Opus 5.5 despite 2x base price (0.025x on the Fable family). That matters for agentic loops dominated by cache reads. US-only inference (inference_geo: "us") adds 1.1x on Claude API, Claude Platform on AWS, Google Cloud Anthropic managed, and Foundry US Data Zone; Bedrock and Google-managed price regionally. Fast mode is available on Claude API, Claude Code, Google Cloud (both options), and Foundry Anthropic-hosted as claude-fable-5-2-fast; it is not listed for Bedrock and is not available on Batch.

Platform ids

Who Should Switch

• On Fable 5.1: switch. Cheaper, faster, better on most rows. Check the four breaking changes in the migration guide first, and remove prompt lines that try to suppress thinking.
• On Opus 5.5: stay for everyday coding and knowledge work. Move up for multi-hour agents, research workflows, CAD and hardware, robotics, and as an orchestrator or advisor over smaller models. Test at high effort before xhigh. See our Opus 5.5 post.
• Security or biology work: Fable will block much of it. Apply for Mythos 5.2 (see the Mythos post).
• Need zero data retention: not available by default.
• Run your own evals; every score here is self-reported.

Outlook

Watch independent reproductions of the long-horizon numbers, whether the Opus 5.5 gap holds on bounded tasks, how far fast mode reaches across platforms, and how the 2x price premium holds against Opus 5.5 once teams measure their own task costs. The honest uncertainty is that every headline score on this page is Anthropic's, often under harnesses the company or named partners chose, and several comparator numbers differ from those models' own launch posts.

Retirement is not sooner than November 9, 2027. Fable 5.1 remains active (retirement not sooner than September 1, 2027). Banner: Anthropic.