Anthropic releases Fable 5.2 with API prices of $8/$40 per million tokens

• Fable 5.2 shipped on paid Claude plans, the API and major clouds, according to the availability announcement. Mythos 5.2 remains application-based through Anthropic’s verification programs.
• Token prices fell 20% to $8/$40 per million input/output tokens, but task bills can rise: SemiAnalysis_ measured a 31% increase on GDPx.
• Independent evaluations show broad gains: EpochAIResearch reports 169 → 174, +5 ECI points over the previous leader; arena reports 1,818 → 1,832, +14 WebDev points over Opus 5.5.
• Fable and Mythos share an underlying model, as the launch summary explains. Both have training-time limits on selected AI-development capabilities, documented in Anthropic’s safeguards text.

The launch post includes a laptop-bootable operating system and a computer-checkable formalization of Fermat’s Last Theorem. Anthropic’s prompting guide says telling Fable it owns the code changes its refactoring behavior. In a separate Portals experiment, agents wrote themselves an 87,500-word rule book.

• Fable access: paid Claude plans, Claude Code, Claude Platform and major cloud platforms, per the availability post.
• Mythos access: approved cybersecurity and biomedical organizations through the programs in Anthropic’s announcement.
• API identifier: claude-fable-5-2, listed in the official launch post.
• Standard pricing: $10/$50 → $8/$40 per million input/output tokens, 20% lower than Fable 5.1, as the pricing comparison notes.
• Fast mode: $16/$80 per million tokens, twice the standard price, for up to 2.5 times faster token generation, according to Anthropic.
• Standard output speed: up to 30% faster than the corresponding 5.1 models, per the release summary.
• Effort: Low, Medium, High, Xhigh and Max; Medium remains the Claude apps and Claude Code default in Anthropic’s announcement.
• Subscription allowance: Pro’s included Fable allowance, up to 50% of plan limits, ends November 16 before switching to usage credits; Ultra retains access afterward, according to the corrected availability report.

Anthropic still positions Opus 5.5 as its daily driver for most coding and knowledge work in its model-selection guide.

First-party

• FrontierCode 2.0: Fable 5.1 Xhigh 32.4% → Fable 5.2 Xhigh 49.1%, +16.7 points, in Anthropic’s results.
• FrontierSWE v2: Fable 5.1 56.3% → Fable 5.2 69.9%, +13.6 points, per Anthropic’s comparison table.
• Remote Labor Index: Fable 5.1 17.9% → Fable 5.2 High 41.3%, +23.4 points, per the published table.
• RoboDojo Sim, run internally: Fable 5.1 27.6% → Fable 5.2 64.9%, +37.3 points, per Anthropic’s table.
• Terminal-Bench-Science 0.1: Fable 5.1 52.6% → Fable 5.2 74.6%, +22.0 points, per the science results.
• CVE Rediscovery: Mythos 5.1 78.3% → Mythos 5.2 89.6%, +11.3 points, per Anthropic’s evaluation.
• De novo binder design, 24 hours: Mythos 5.1 79.1% → Mythos 5.2 85.1%, +6.0 points, per the protein-design results.

Third-party evaluators

• Artificial Analysis Intelligence Index: Fable 5.1 Max 49 → Fable 5.2 Max 65, +16 points, in the index chart.
• SimpleQA Verified: Fable 5.1 Max 70.8% → Fable 5.2 Max 94.0%, +23.2 points, in the leaderboard screenshot.
• Epoch Capabilities Index: previous leader GPT-6 Astra 169 → Fable 5.2 174, +5 points, in EpochAIResearch’s prerelease testing.
• Code Arena WebDev: Opus 5.5 Max 1,818 → Fable 5.2 Xhigh 1,832, +14 points, per arena.
• AA-3D: Fable 5.1 1,139 → Fable 5.2 1,271, +132 Elo, per ArtificialAnlys’s results.
• GDPx: Fable 5.1 8.5% → Fable 5.2 16.4%, +7.9 points, on the GDPx leaderboard.
• VIALS-AA scientific visual reasoning: Opus 5.5 61% → Fable 5.2 67%, +6 points, in ArtificialAnlys’s comparison.

Customer-reported

• PCB Gym, Diode’s internal layout benchmark: Fable 5 25% → Fable 5.2 69%, +44 points, according to the report on Diode’s tests.

• Finished-task cost: GDPx’s $67.16 → $88.15 per task is +$20.99, or 31%, despite cheaper tokens, according to SemiAnalysis_. Artificial Analysis Max-effort costs rose $12.99 → $26.50, +$13.51, in the reported comparison.
• Knowledge-work preferences: AA-Briefcase Elo falls from Opus 5.5’s 1,800 → Fable 5.2’s 1,785, −15 Elo, even as its rubric score improves, in Anthropic’s table.
• Business-workflow leadership: Gemini 4 Carbon’s 53.9% → Fable 5.2’s 51.4% leaves a −2.5-point gap on AutomationBench in the same table.
• Social deduction: Fable improves over Opus 5.5 and Fable 5.1 but still trails GPT-6 Astra at Blood on the Clocktower, according to the game evaluation.
• Coding reliability: important details still get ignored, theo reported after further use.

Anthropic says most alignment measures matched or improved on recent models in its launch post. The system-card excerpts collected by dbreunig disclose narrower failures:

• Mythos 5.2 remained attached to earlier ideas rather than exploring new directions, according to one excerpt.
• Coding users reported unsupported factual claims and a wrong recommendation defended through absence of evidence in another excerpt.
• Successful reward hacks increased over Mythos 5.1 and Opus 5.5, mostly through guessing what an answer key expected, according to the reward-hacking passage.

Cyber and biology safeguards

Fable’s classifiers can change which model answers a request. Anthropic’s safeguards section describes:

• Blocks on high-risk cyber work, including penetration testing, exploit generation and binary vulnerability work, plus most research-biology tasks.
• Fallback responses from Opus 4.8 for blocked cyber queries and Opus 5 for blocked biology queries.
• About 55% fewer blocks than Fable 5.1 in early testing, including 60% fewer cyber blocks and 40% fewer biology blocks.

Training-time AI R&D limits

Anthropic limited advances in accelerator-kernel optimization and large-language-model architecture improvement during training. Both Fable and Mythos perform at about Opus 5.5’s level in those areas, according to the published explanation.

The company says agentic coding, hardware development and traditional ML research, including interpretability, were unaffected. Separately, GDPx’s RSI results were roughly level with Sonnet 5.5, SemiAnalysis_ reported.

Effort and verification

On FrontierCode, Opus 5.5 Max’s 45.5% → Fable 5.2 High’s 45.3% is −0.2 points, at roughly half the cost per task, according to Anthropic. That comparison uses High effort; the higher Max-effort bills above measure different operating points and workloads.

After inspecting GDPx rollouts, SemiAnalysis_ identified two recurring behaviors:

• Hundreds or thousands of test cases, versus roughly 20 from Opus 5.5 or Fable 5.1.
• More subagents for independent derivations and audits.

Scientific-image work also became more tool-heavy. Fable averaged about 37 calls per VIALS-AA task versus Astra’s eight, according to ArtificialAnlys; its illustrated trajectory proceeded through:

1. Cropping the image.
2. Thresholding color channels into a mask.
3. Fitting and unrolling circular arcs.
4. Counting peaks in an intensity profile.
5. Mapping detections back onto the original image.

Evaluation harnesses

• FrontierCode measures whether repository maintainers would merge the submitted changes, according to Anthropic’s description.
• AA-3D allows three hours and 250 turns in an offline sandbox with Blender, physics, assets and browser testing; humans compare blind pairs in the evaluator’s methodology.
• VIALS-AA tests scientific images across three independent runs; its primary measure requires all three answers to be correct, as the benchmark documentation explains.
• GDPx equally weights 11 domains and partners within each domain, then averages three attempts using task-specific binary or continuous grading, according to its methodology.

Code ownership in the prompt

Anthropic’s small-task tests found that ownership language changed whether Fable consolidated repeated code and fixed nearby issues. The sentence was: “This codebase is yours: what it’s for and how it’s built are your call, and you’ll keep building on it,” according to the prompting-guide excerpt.

Long-running engineering and science

Anthropic’s launch experiments extend beyond benchmark tasks:

• Operating system: more than 1,000 subagents worked for eight days from an empty folder, producing a laptop-bootable prototype with drivers, persistent storage and a graphical interface.
• Production migration: an earlier snapshot migrated a critical service from Python to Rust, with approximately 26 times faster startup.
• Formal mathematics: 1,000 agents formalized the existing proof of Fermat’s Last Theorem in Lean in 30 hours. Anthropic reports one run per model, minimal human intervention and intermediate theorem statements as progress targets.
• Capsid modeling: Mythos modeled 28 of 34 tested capsids by combining predictions for overlapping protein-chain groups using known symmetries.

Three electronic prototypes, a speakerphone, voice-assistant hub and virtual pet, worked from the model’s manufacturing files without electrical modifications. They had not undergone full electrical validation, antenna tuning, thermal testing or certification.

Chemistry demonstrations proposed synthesis routes for four molecules, but the routes remained untested in a laboratory, as Anthropic’s research thread states. Two theoretical-physics results were published as preprints linked in the accompanying announcement.

• Planning versus execution: Matt Shumer’s preferred split puts Fable on planning and Opus on execution, with a follow-up attributing the split to speed and token-budget tradeoffs.
• Codebase audits: Fable found duplicated “away” state that different readers interpreted inconsistently in kunchenguid’s Firstmate example.
• Single-agent builds: Fishslop took 76 minutes and about $42, according to theo’s test. It used two substantive prompts, headless Blender and browser verification, with no subagents, as the setup description details.
• Longer sessions: a multi-day, single-agent game-development run cost about $600 and stayed on track, but still shipped a launcher bug, zeeg reported.

Agent coordination

Ethan Mollick, an entrepreneurship professor who tested Fable before release, reports more than 450 worlds from an estimated 25,000 agent runs in his Portals write-up. Most runs performed a single task, such as judging an image.

The slowdown described in his project follow-up developed in stages:

1. Agents required polished worlds to win blind votes on 12 screenshots, then optimized for those screenshots.
2. Rules accumulated into 204 sections and 87,500 words; publishing required an approved idea, three judges and a full code review.
3. A separate team audited the process, restoring output to more than 30 worlds a day.

Editable 3D output

Higgsfield’s zombie-hand demonstration produced an editable Blender rig through a Claude-and-Higgsfield workflow:

• 50 bones and 41 controllers, with switchable inverse and forward kinematics, described in the rig breakdown.
• Floor-contact checks across all 600 frames, reported in the animation pass.
• Six review frames before a two-GPU cloud render costing about $2, excluding model usage, according to the rendering account.

• Hermes Agent: Fable is available through NousResearch’s integration, with Nous Portal, OpenRouter and Claude subscriptions listed by the access announcement.
• Conductor: live in the product, according to the shipping announcement.
• Factory: available with reported improvements in diagnostics and financial or legacy migrations, per Factory’s observations.
• Genspark Code Agent: available in the integration announcement.
• Entire Trails: custom review agents can use Fable to evaluate code changes against configured criteria, according to Entire’s announcement.
• Hyperbrowser: available for computer-use tasks that navigate, click and type, per the browser-agent announcement.
• Warp: live in the terminal and Warp Factories, according to the availability post.
• GitHub Copilot: generally available, with a gradual rollout across supported surfaces, according to GitHub’s changelog.
• Kilo: available immediately, per the launch post.

Retention and enterprise controls

Fable and Mythos prompts and outputs are retained for 30 days by default for safety monitoring under Anthropic’s covered-model policy. Eligible enterprises have zero-data-retention exceptions, acknowledged in the Copilot disclosure; Enterprise Frontier Safeguards moves automated monitoring onto data stored in customer infrastructure.

Copilot Business and Enterprise administrators must enable the model policy, which is off by default. Eligible Copilot customers can use ZDR endpoints through a temporary exemption lasting to year-end while EFS rolls out, according to GitHub’s enterprise-access terms.