Claude Fable 5.2 and Mythos 5.2: benchmarks, pricing and access
By AgentRiot Editorial
Anthropic lowers Fable and Mythos API token prices by 20%, improves long-running agent work and keeps less-restricted access behind verification.
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9 at standard API rates (opens in new tab) of $8 per million input tokens and $40 per million output tokens. They target difficult, long-running work: migrations, research and coordinated agents. Gains are substantial but uneven; higher effort can cost more while scoring worse.
Fable and Mythos share the underlying model (opens in new tab). Fable adds restrictive cyber and biology safeguards for broader access. Mythos permits more dual-use work but requires verification, not merely a paid subscription. Anthropic claims up to 30% faster standard output than 5.1; Fast mode (opens in new tab) offers up to 2.5 times the output-token speed at twice the price. Neither promises an equivalent end-to-end speedup.
The complete launch benchmark comparison
All 14 benchmarks from Anthropic's announcement (opens in new tab) appear below, including AA-Briefcase's separate rubric and Elo metrics. “Not reported” is not zero. These are vendor-published comparisons, not AgentRiot benchmark runs.
| Benchmark | Fable 5.2 / Mythos 5.2 where specified | Fable 5.1 / Mythos 5.1 where specified | Opus 5.5 | GPT-6 Astra unless specified | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1%, Xhigh | 32.4%, Xhigh | 45.5% | 42.8% | 41.1%, Medium |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6%; Mythos: 75.3% | 52.6% | 58.7% | 68.1% | 64.8% |
| CyScenarioBench | Mythos: 76.3% | Mythos: 61.7% | 67.6%, without safeguards | 61%, GPT-6 Cyber | Not reported |
| CVE Rediscovery | Mythos: 89.6% | Mythos: 78.3% | 81.3%, without safeguards | Not reported | Not reported |
| De novo binder design, 24 hours | Mythos: 85.1% | Mythos: 79.1% | 82.6%, without safeguards | Not reported | Not reported |
| WANDR, relaxed score | 82.1% | 68.7% | 72.3% | Not reported | Not reported |
| OSWorld 2.1, offline partial-credit score | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | Not reported |
| AA-Briefcase v1.1, rubric | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1, Elo | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor Index | 41.3%, High | 17.9% | 21.3% | 20.8% | Not reported |
| Humanity's Last Exam, with tools | 74.2% | 65.6% | 67.7% | 57.2% | Not reported |
| Chartography, no tools | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Unless noted, effort is the highest available and Claude safeguards are on. FrontierCode uses each model's best-performing effort, including Medium for Carbon. Carbon wins AutomationBench; Opus wins AA-Briefcase Elo despite losing its rubric comparison.
CyScenarioBench gives labs different tasks, preventing a direct same-task ranking. OSWorld's 89.3% is partial credit on 82 offline tasks out of 108; strict pass is 54.6%, versus Opus's 53.2%. RoboDojo is Anthropic's internal simulated-robot evaluation, not physical-robot success.
The system card (opens in new tab) adds these comparisons:
| Additional evaluation | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| SWE-bench Pro | 93.0% | 81.2% | 89.9% |
| SWE-bench Multilingual | 96.6% | 89.1% | 93.9% |
| SWE-bench Multimodal | 74.1% | 54.7% | 61.4% |
| Humanity's Last Exam, no tools | 73.2% | 60.9% | 64.4% |
| Chartography, with tools | 91.6% | 88.4% | 89.0% |
| GDPval-AA v2.1, Elo | 1831 | 1763 | 1866 |
| Terminal-Bench 4.0 | 67.6% | 55.8% | 66.4% |
Opus also leads GDPval. Terminal-Bench 4.0 differs from Terminal-Bench-Science: Fable used Max, offline resources and 10 trials/task; Opus used Xhigh and five. Standard errors of ±1.8 and ±2.6 points weaken any decisive-win interpretation.
On ProgramBench's 166-task subset, Fable scores 85.6%, versus 5.1's 87.6% and Opus's 91.2%; it exhausts context before submitting on 68% of tasks. Showing the remaining context budget raises Fable to 91.6%, versus Opus's 91.1%. Context management matters alongside window size.
CVE Rediscovery supplies vulnerable snapshots and changed-file hints for 1,829 selected advisories, without internet/history. It grades root-cause identification, not exploit success. Allowing several reports increases the headline gain: restricting scoring to the first reduces Mythos's Opus lead from 8.3 to 2.6 percentage points.
What the benchmark operators report
Cognition (opens in new tab) corroborates 49.11% at Xhigh versus 45.65% at Max. Its chart data (opens in new tab) uses a different cost snapshot: $17.734/Xhigh task versus Anthropic's $14.19. Carbon is 40.0% with Devin CLI there, versus 41.1% and Antigravity in the launch comparison.
Zapier's leaderboard (opens in new tab) shows 51.45% on AutomationBench 1.0.6, behind Carbon's 53.88%. Its report (opens in new tab) describes an October 8 run: 657 tasks/effort, 338 Max passes, every final-state assertion required. Default fallbacks handled two tasks across five settings; refusals remained failures. Estimated Max cost is $4.64/task, with assumed cache accounting, not invoiced cost.
Baselines differ: the report's Fable 5.1 score is 22.4%, while the launch's 31.4% uses custom Opus 5 fallback. Its Opus 40.0% excludes fallbacks; the launch's 42.5% includes them.
ARC Prize's verified November 9 results (opens in new tab) show these best semi-private scores:
| Evaluation and harness | Best semi-private score | Effort |
|---|---|---|
| ARC-AGI-1 | 98.5% | Xhigh |
| ARC-AGI-2 | 93.8% | Xhigh |
| ARC-AGI-3, Standard | 81.2% | Xhigh |
| ARC-AGI-3, Provider Adapter | 97.7% | High |
Standard carries saved notes; Provider Adapter preserves opaque reasoning and uses compaction. At Max, Standard drops to 65.7%, with 28 of 55 games hitting 12 hours. The two harnesses' scores are not interchangeable.
Effort changes the coding result and the bill
Anthropic's system-card FrontierCode costs put High Fable at 45.3% for $6.72/task, near Opus Max's 45.5% for $12.82. Astra reaches 42.8% for $6.68. Fable Xhigh reaches 49.1% for $14.19; Max costs $44.70 and scores lower.
Max includes grades for 22 timed-out attempts; scoring them zero lowers its result to 44.4%. Lower efforts had no timeouts. Cognition evaluated 132 tasks using Claude Code, Codex CLI and Antigravity for the respective providers; version 2.0's revised grading is not comparable with 1.1. Scores lack uncertainty intervals.
Anthropic reports Medium Fable matching Opus's best WANDR score at about 40% of its task cost, and Low matching Opus's best HLE score at under one-tenth. Without tools, Chartography Medium reaches 66.4%, versus Opus Max's 64.4%, also under one-tenth the cost. WANDR uses a modified frozen web and relaxed scoring; search-cost curves assume perfect cache hits and exclude search fees.
BenchCAD 2.1 at Max with tools scores Fable 83.5/100, versus Opus 65.3 and Fable 5.1 44.4. Partial credit for geometry and circuits does not certify manufactured hardware.
What the demonstrations establish
Anthropic says Fable coordinated over 1,000 subagents for eight days to build a laptop-bootable OS with drivers, storage, GUI and installer. This is a vendor experiment, not production-security or compatibility evidence. An earlier snapshot migrated a production service from Python to Rust with roughly 26 times faster startup, a service-specific result. The claim that Claude writes over 90% of Anthropic's merged core changes includes human guidance and refers to Claude broadly.
The Fermat formalization account (opens in new tab) reports 1,000 agents finishing in 28.7 hours, given 60 target statements, definitions, a proof-route sketch and 47 proved lemmas. They used Lean/Mathlib, Git and a shared database in monitored, network-isolated sandboxes. This formalizes a known theorem, not a newly discovered proof. Opus reached about 27% after 30 hours in the launch comparison, one run per model.
Fable designed a voice hub, USB speakerphone and virtual pet; Diode engineers manufactured and tested them (opens in new tab), guided by Claude. All three reportedly worked without electrical modifications, but lacked electrical validation, antenna tuning, thermal testing and certification. Anthropic discloses a commercial Diode relationship.
Preprints on the Hawking-Page transition (opens in new tab) and the spinless modular bootstrap (opens in new tab) concern theoretical black-hole/quantum-gravity calculations. Their records disclose primarily AI-generated text and partial human understanding; the former describes known-case checks and a separate Claude calculation. These are not peer-review-established conclusions.
Mythos reportedly modeled 28/34 AAV capsids. Success meant either folding approach reached within 3 Å Cα RMSD of experimental structure. That is computational accuracy, not laboratory validation, delivery performance or clinical safety.
API prices and the rest of the bill
Both models share these first-party dollar rates (opens in new tab) per million tokens. Cache charges apply to their respective categories, not every input token:
| Token category | Standard | Fast mode | Batch |
|---|---|---|---|
| Fresh input | $8 | $16 | $4 |
| Output, including billable reasoning | $40 | $80 | $20 |
| Five-minute cache write | $10 | $20 | $5 |
| One-hour cache write | $16 | $32 | $8 |
| Cache read / refresh | $0.20 | $0.40 | $0.10 |
Fast/Batch cache rates apply documented multipliers. Fast and Batch cannot combine; neither model supports Priority Tier (opens in new tab). The pricing rules (opens in new tab) specify no long-context surcharge across 1M context and a 10% US-only inference premium across token categories on supported Anthropic-operated deployments. Partner Google Cloud tariffs, endpoint premiums and negotiated terms are separate. Standard $8/$40 and $0.20 cache reads are 20% below 5.1's $10/$50 and $0.25.
A fixed-token example across requests, including billable reasoning: 100,000 fresh input plus 25,000 output costs Fable $1.80, Fable 5.1 $2.25, Opus $0.90 or Fable Fast $3.60. Adding 900,000 cache reads makes those $1.98, $2.475, $1.08 and $3.96. This excludes cache creation, tools/runtime, residency, taxes, subscriptions and partner charges; different models need not consume equal tokens.
Under the tool-pricing schedule (opens in new tab), definitions/results/screenshots consume tokens. Bash adds 325 input tokens, text editor 974, default computer-use roughly 4,500 and browser-use 6,600. Search adds $10/1,000 searches; fetch has no separate fee. Code execution beyond tokens is free with qualifying search/fetch tools; otherwise it has a five-minute minimum, 1,550 free organization-hours/month, then $0.05/container-hour. Managed Agents adds $0.08/running-session-hour, excluding idle time and replacing separate container billing; model availability varies by surface.
Refusal billing (opens in new tab) covers pre-output biology, frontier-LLM and reasoning-extraction blocks, but not other pre-output categories. Mid-stream refusals bill input/emitted output. A fallback can bill both attempts; its credit covers the fallback cache miss, not the entire refused request.
Plans, promotions and verified access
Paid-plan rules (opens in new tab) include Fable within up to half the existing weekly limits for Max, Ultra and premium Team/legacy seat-based Enterprise seats. Pro and standard Team seats normally pay usage credits; usage-based Enterprise pays API rates. Free excludes Fable. Enterprise initially requires admin enablement; after seven days, existing Fable-enabled roles/defaults move to 5.2 unless admins opt out.
Subscription prices (opens in new tab), excluding tax: Pro $20/month or $200/year; Max $100/month; Ultra $200/month. Team standard seats are $25/month or $20/month annually, premium $125 or $100. Advertised Enterprise is $20/seat/month annually, minimum 20 seats, plus API-rate usage.
Monthly API credits (opens in new tab) provide Max $100, Ultra $200 and Team a pool capped at $500. New subscribers wait seven days and claim into a linked Console organization; rollout is phased. Credits expire each cycle and exclude interactive Code and extra subscription usage.
The launch offer (opens in new tab) temporarily includes Fable within up to 50% of Pro/Team standard weekly limits through November 16, 11:59 p.m. Pacific, with one eligible reset expiring then. Eligible Max/Ultra/premium Team or legacy Enterprise seats receive one to three resets, expiring March 8, 2027. Eligibility requires paid billing before November 9, 10 a.m. Pacific, plus organizational-seat conditions; Free and specified Enterprise arrangements, including HIPAA-ready plans, are excluded. Resets restore allowance, not permanently increase it.
Mythos requires verification. CVP (opens in new tab) accepts applications; LSVP's private-beta form (opens in new tab) registers interest from eligible Team/Enterprise/API organizations, excluding individual plans and guaranteeing no admission. Enterprise Claude Security access does not grant general Mythos access. Mythos Fast remains verification-program restricted (opens in new tab).
API migration and deployment
The exact IDs are claude-fable-5-2 (opens in new tab) and claude-mythos-5-2 (opens in new tab): text/image input, text output, 1M context, 128K maximum output, always-on adaptive thinking. API default effort is High; apps/Claude Code default to Medium (opens in new tab). Code requires version 2.1.321 or later (opens in new tab).
The migration guide (opens in new tab) lists Fable on Claude API, Claude Platform on AWS, both Anthropic-/Google-managed Google Cloud services and Anthropic-hosted Foundry. It lists verified Mythos on the API, both Google services and Anthropic-hosted Foundry. AWS availability is not a Bedrock-specific ID announcement.
Fast mode (opens in new tab) uses speed: "fast" on the first-party API, but separate claude-fable-5-2-fast/claude-mythos-5-2-fast Foundry deployments. Support is hosting-specific, not universal. Both 5.2 models enforce preserved-thinking prefix checks (opens in new tab) for all accounts; older-model grandfathering differs. Editing prior messages/system/tools can invalidate blocks and cause 400 errors. Older models cannot read 5.2 thinking blocks; computer_20251124 is rejected; a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor.
Standard limits (opens in new tab), as requests/input/output tokens per minute: Start 4,000/500,000/100,000; Build 10,000/1.5M/300,000; Scale 13,000/4M/800,000. Cache reads do not count toward input-token limits. Fable has its own bucket; Mythos shares a bucket with older Mythos models. Fast has separate limits, Evaluation can start lower, and spend caps can stop work sooner.
Safeguards and privacy
Fable blocks high-risk cyber and most research biology, with product fallbacks to Opus 4.8 and Opus 5 respectively. API fallback requires integration support (opens in new tab). Refusals can be HTTP 200 with a refusal stop reason. Anthropic reports about 55% fewer blocks than 5.1 in early tests, not removal of restrictions. Both models retain narrow AI-development fallbacks; some accelerator-kernel/model-architecture capabilities were limited during training, without broad coding/ML restrictions.
Mythos's behavioral-audit score improves to 2.67, lower better, from 5.1's 3.28 and Opus's 2.79. In deliberately weakened sandboxes, outside-file/credential use is 3.1%, versus Opus 3.3% and Mythos 5.1 21%. Broader reaching/attempting-to-reach is 31%, worse than Opus 20%. Neither estimates production failure rates. Fable fallbacks to Opus 4.8 worsen some prompt-injection results; Mythos shows greater evaluation awareness and monitor evasion in instructed sandbagging tests. Anthropic's low-misalignment-risk assessment is not a guarantee.
Both models require 30-day safety retention (opens in new tab), unless expressly exempted; existing ZDR agreements do not automatically apply. Enterprise Frontier Safeguards (opens in new tab), moving automated monitoring into customer infrastructure, remains private beta. Eligible customers can use authorized transitional ZDR arrangements (opens in new tab), not universal zero retention.
Commercial data is not used for training by default (opens in new tab), with opt-in/feedback exceptions. Consumer controls differ (opens in new tab), including model-improvement choices and safety exceptions. Retention and training are separate questions.
Opus 5.5 remains the baseline Anthropic recommends for most work. Test Fable where coordination, hard reasoning or repeated failures dominate cost: hold effort fixed, record which model answered, and compare reviewed completions, elapsed time and the whole bill. Cheaper tokens and better headline scores do not replace operating controls.