On Friday, we asked Jamin Ball what comes next for Anthropic’s IPO. On Monday the roadshow started and Anthropic shipped the long-anticipated next-version of Claude Fable and Mythos. Of course, it is the new top model in the world, taking back Opus 5.5’s #1 spot from Gemini 4 Carbon which had its brief moment in the sun (10 days).

Fable comes with a 20% per token price cut… but is more token hungry, as both SemiAnalysis GDPx and Artificial Analysis agree:

Fable 5.1 cost $10 per million input tokens and $50 per million output tokens. Fable 5.2 costs $8 and $40. Yet on GDPx, a benchmark of professional tasks, SemiAnalysis paid $88.15 per task, up from $67.16. The run logs give a clue. Where Opus 5.5 and Fable 5.1 would try about 20 test cases, 5.2 sometimes wrote hundreds or thousands and spun up subagents to check its work independently. All that checking helped Fable take first place at 16.4%. You pay for the checks either way.

Which effort setting should you put in production? On FrontierCode 2.0, Anthropic reports that Fable at High matched Opus 5.5 at Max for about half the cost per task… but don’t talk about Fable at Max.

AI Twitter Recap

Top Story: Claude Fable 5.2 / Mythos 5.2 Benchmarks, Economics, Restrictions, and First Impressions

What happened

Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2, claiming both are "a step above Opus 5.5 in almost all the capabilities we evaluated." It tops most independent leaderboards. It also arrives with deliberate RSI nerfs, heavier per-task costs and restricted subscription access.

• Launch claims. @claudeai announced the pair as built for "the most daunting technical challenges," with gains in coding, reasoning and vision. @marktenenholtz flagged unusual system-card phrasing: "The model that underlies Claude Fable 5.2 and Claude Mythos 5.2." That implies a single shared base model.

• Vendor-reported headline numbers, per @kimmonismus:

  • Up to 30% faster output than the 5.1 models.

  • On FrontierCode 2.0, Fable 5.2 at High matched Opus 5.5 at Max for about half the cost per task, and at Xhigh it was the top released model.

  • 1,000 Fable 5.2 agents formalized the existing proof of Fermat's Last Theorem in 30 hours. Opus 5.5 completed 26% in 18 hours on the same setup.

  • Mythos 5.2 accurately modeled 28 of 34 gene-therapy capsids.

  • Alignment was at or above recent Claude models, with external pre-release testing.

• Default effort. Medium is the default effort in the Claude apps and Claude Code.

• Pricing. $8/$40 per 1M input/output tokens, 20% below Fable 5.1's $10/$50 (@jaminball, @arena). The rest of the spec, per Artificial Analysis:

  • Cache writes cost $10 (5-minute TTL) or $16 (1-hour TTL); cache reads cost $0.20.

  • 1M-token context with text and image input.

  • Five effort levels: low, medium, high, xhigh, max.

• Long-horizon demos from Anthropic.

  • An OS was built from an empty folder by 1,000+ subagents over 8 days. It boots on a laptop and has its own kernel, disk/keyboard/network drivers, persistent storage, windowing, login, an installer and built-in Claude (@kimmonismus).

  • Fable designed three working electronic devices from a short prompt (@kimmonismus).

  • It proposed untested chemical synthesis routes.

  • It improved on listed solutions to 72 open optimization problems.

  • @RLanceMartin published the safety-bounded scaling write-up for the 1,000-agent FLT run.

• Subscription access.

  • Ultra and Team Premium users get Fable "resets" scaled by tenure: 1 reset under 6 months, 2 for 6–12 months, 3 above 12 months, usable until March 8 (@ClaudeDevs; echoed by @bcherny and @edwinarbus).

  • Pro and Team Standard get a one-week trial capped at half the weekly limit, plus one reset.

  • Fable can only consume 50% of a weekly limit.

• Data retention. Fable 5.2 requires data retention by default so Anthropic's safety classifiers can run. Zero data retention is available only to eligible enterprise customers (@code).

• Rollout, day one:

  • Cursor (@cursor_ai).

  • GitHub Copilot / VS Code, where it "solved more tasks than Fable 5.1 while using fewer steps and tool calls" (@code).

  • Perplexity Computer (Pro/Max) and Search (Max) (@perplexity_ai). @AravSrinivas made it his "new default."

  • Factory (@factory).

  • Hermes Agent via Nous Portal and OpenRouter (@Teknium).

  • Arena (@arena) and DesignArena (@grx_xce).

• Official model-selection guide. @ClaudeDevs published guidance on Fable vs Opus vs Sonnet vs Haiku. It covers when to raise effort instead of switching models, and when to push subagents to smaller models.

Independent evals (facts)

• Artificial Analysis Intelligence Index v5 (@ArtificialAnlys, breakdown): Fable 5.2 is #1 at 65 (max).

  • Fable 5.2 (max) — Index score 65

  • Gemini 4 Carbon (high) — Index score 60

  • Opus 5.5 (max) — Index score 59

  • GPT-6 Astra (max) — Index score 57

  • Fable 5.1 (max) — Index score 49

  • New records:

    • HLE-Diamond — Fable 5.2 80% · Previous best 69%

    • AA-Omniscience v1.1 — Fable 5.2 67 · Previous best 54

    • Terminal-Bench-Science — Fable 5.2 73% · Previous best 65%

    • CritPt v2 hard — Fable 5.2 77% · Previous best 72%

    • FrontierCode 2.0 — Fable 5.2 49% · Previous best 46% (Opus 5.5)

  • Weaker spots:

    • AA-Briefcase: 1785 Elo, within the confidence interval of Opus 5.5 (1800) and behind Sonnet 5.5 (1813).

    • GDP.pdf: 24% all-pass vs 26% for Opus 5.5.

    • Terminal-Bench 4.0: 60%, level with Opus 5.5 and behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%).

    • AutomationBench-AA: 76%, ahead of Opus 5.5 (70%) but behind Gemini 4 Carbon (78%).

  • Efficiency: At high and xhigh it scores 62 using about 98k output tokens per task, vs 173k for Opus 5.5 max. At max, cost per task is about $26.5, roughly 2.6× Opus 5.5 ($10.0) and 2× Fable 5.1 ($12.99).

• Epoch Capabilities Index (@EpochAIResearch): a record 174, vs 169 for GPT-6 Astra. That is above the pre-Astra reasoning-model trend, which Epoch tied to the AI R&D discussion. Scores came from Epoch's own pre-launch benchmarks.

• ARC-AGI (Verified) (@arcprize):

  • ARC-AGI-3, standard harness (new high) — Score 81.2% · Cost $21.9K

  • ARC-AGI-3, provider adapter harness — Score 97.7% · Cost $18.6K

  • ARC-AGI-2 — Score 93.8% · Cost $0.98/task

  • ARC-AGI-1 — Score 98.5% · Cost $0.36/task

• Vals Index (@ValsAI): new #1, leading seven benchmarks across coding, finance and science. "High" was the best-performing effort, above xhigh and max.

• CursorBench (@cursor_ai): 60% at max (top model) and 56.9% at default, vs 49.2% for Fable 5.1. Cursor warns that effort levels behave differently from 5.1, so start at the default rather than carrying over old settings (pricing note).

• Arena Code: WebDev (@arena): #1 at 1832 (xHigh), ahead of Opus 5.5 Max (1818), Gemini 4 Carbon (1768) and GPT-6 Astra (1767). @ml_angelopoulos noted a +87 pt gain over 5.1. Peter Gostev's 100+ generation comparison is here, with 17 sample generations here.

• AA-3D (@ArtificialAnlys): #1 at 1271 Elo (95% CI 1246–1297), with no CI overlap with any other model.

  • Runners-up are Sonnet 5.5 (1204) and Opus 5.5 (1202), with GPT-6 Astra at 1168 and Gemini 4 Argon at 1147.

  • The score is +132 over Fable 5.1.

  • Method (details): each model runs as an agent in an offline sandbox with Blender, Rapier physics, Python and a headless browser, with up to 3h and 250 turns. Humans vote blind on pairs, spending at least 30s per scene.

• VIALS-AA, scientific visual reasoning (@ArtificialAnlys):

  • Fable 5.2 scores 67%, vs 64% for Gemini 4 Carbon and 63% for Astra.

  • Fable averages about 37 tool calls per task, vs about 8 for Astra. It averages 1.60 calls per turn, the highest on the board (per-turn note).

  • Example trajectory: cropping, then color-channel masking, then fitting circles and unrolling arcs, then counting peaks in an intensity profile, then re-mapping detections back onto the original image to verify.

• SemiAnalysis GDPx (@SemiAnalysis_): #1 at 16.4%.

  • Rollout analysis attributes the gain to testing breadth. Fable generates hundreds to thousands of test cases where Opus 5.5 and Fable 5.1 stopped at about 20, and it more eagerly spawns subagents for independent re-derivation (5/7).

  • Cost per task rose 31% ($67.16 → $88.15) despite the 20% cheaper per-token price (6/7).

  • @dylan522p breaks down the domains:

    • Significantly better at chip design, SWE, biology and materials science.

    • Not better at finance, legal, quant trading, RSI or hardware design.

    • The RSI result reflects Anthropic's explicit nerf.

• Perplexity WANDR (@perplexity_ai): record 0.851 at $14.51/task, on the Pareto frontier. @marktenenholtz notes Fable Low matches Opus 5.5 Max at significantly lower cost.

• Experiment-result prediction (@pzeroresearch): best tested at 0.36, continuing the steady trend since GPT-4 Turbo. They frame this as a proxy for "research taste."

• No-CoT reasoning/knowledge evals (@Tim_Hua_): slightly above the Anthropic trend and slightly below GPT-6.1 Sol. @scaling01 highlighted the NCKI number and noted these metrics now appear in every system card (follow-up).

• Blood on the Clocktower (@dwarkesh_sp): still worse than GPT-6 Astra, but much better than Opus 5.5 and Fable 5.1.

• Safety, from UK AISI (@JordanTensor): Mythos 5.2 is "substantially more capable of evading a monitor which reviews its outputs than comparison models."

The effort-level anomaly

• Multiple evals show max/xhigh underperforming high: Vals, @theo on AA's code-only view (where the gap to other models also narrows), and Fable-specific dips.

• @nrehiew_ notes that only Fable models show the max-reasoning dip; Opus and Sonnet do not.

• @Yuchenj_UW put it bluntly: "Don't use Fable 5.2 Max."