Anthropic rolled out Claude Fable 5.2 and Claude Mythos 5.2 on Monday, coinciding with the start of the company's IPO roadshow. The company said both models are "a step above Opus 5.5 in almost all the capabilities we evaluated." Independent evaluations place Fable 5.2 at or near the top of many leaderboards, but practical costs, deliberate capability limits and subscription gating have drawn close attention.
Performance and benchmarks
- Artificial Analysis Intelligence Index v5 ranks Claude Fable 5.2 (max) at 65 index points, with several new records on subbenchmarks (for example HLE-Diamond 80%, AA-Omniscience 67). Opus 5.5 (max) scores 59 and Gemini 4 Carbon (high) 60.
- SemiAnalysis GDPx: Fable 5.2 leads with 16.4%; the gain is attributed to breadth of test generation — Fable often produces hundreds or thousands of test cases and spawns subagents for independent verification, whereas Opus 5.5 and Fable 5.1 typically stopped around ~20 test cases.
- FrontierCode 2.0: Anthropic reports Fable 5.2 (High) matched Opus 5.5 (Max) on some tasks at about half the per-task cost; on Xhigh it was the top released model.
- Other independent metrics: Epoch Capabilities Index set a record at 174, Vals Index places Fable #1 across several domains, and CursorBench / Arena Code show leading placements. There are also weaker spots for Fable 5.2 — for example GDP.pdf all-pass 24% vs Opus 5.5 at 26%, and some Terminal-Bench measures where other models lead.
Pricing — list price versus per-task cost
- Token pricing: Fable 5.2 list price is $8 per 1M input tokens and $40 per 1M output tokens, a 20% cut from Fable 5.1's $10/$50.
- Real task cost: despite the cheaper per-token price, SemiAnalysis GDPx reports per-task cost rising from $67.16 to $88.15 (a 31% increase) because Fable 5.2 generates far more tests and subagent work.
- Efficiency and cost at different effort levels: at high and xhigh Fable 5.2 used roughly 98k output tokens per task in one measurement (vs Opus 5.5 max ~173k tokens), but at max reasoning the cost per task can spike — one figure cites about $26.5 per task at max,
2.6× the Opus 5.5 cost ($10) and2× Fable 5.1 ($12.99).
Effort levels and the "max" anomaly
- Anthropic provides five effort levels: low, medium (the default in Claude apps and Claude Code), high, xhigh and max.
- Multiple independent evaluations found that for Fable 5.2, max or xhigh sometimes underperform high: non-monotonic scaling appears in several views and users have warned against using max reasoning with Fable 5.2.
- The practical reason appears to be Fable's verification behavior: it spawns many subagents and produces large numbers of test cases to check results, which increases token use and cost.
Restrictions, subscriptions and data retention
- Access controls: Fable-specific "resets" and trial rules vary by subscription tier. Ultra and Team Premium subscribers receive Fable resets scaled by tenure (1 reset <6 months, 2 resets 6–12 months, 3 resets >12 months, usable until March 8). Pro and Team Standard get a one-week trial capped at half the weekly limit plus one reset.
- Quota cap: Fable can consume at most 50% of a weekly limit, so accounts using only Fable effectively lose half their weekly allowance for other models.
- Data retention: Fable 5.2 requires default data retention so Anthropic's safety classifiers can run; zero-data-retention is available only to eligible enterprise customers.
Rollout and integrations
On day one the models appeared across multiple integrations: Cursor, GitHub Copilot / VS Code (reported to have solved more tasks than Fable 5.1 while using fewer steps and tool calls in some setups), Perplexity Computer (Pro/Max) and Search (Max), Factory, Hermes Agent via Nous Portal and OpenRouter, Arena and DesignArena. Several practitioners set Fable 5.2 as their new default.
Large demos and case studies
Anthropic and community users demonstrated ambitious agentic runs: 1,000 Fable 5.2 agents formalized the existing proof of Fermat's Last Theorem in 30 hours in one reported setup; other demos include building an OS from an empty folder by 1,000+ subagents over eight days, and multi-day projects that consumed billions of tokens and tens of thousands of dollars in API spend. These runs illustrate capabilities but also very high resource use.
Community and practitioner reactions
- Best practices: Anthropic staff and experienced users recommend starting at medium or high effort, reserving xhigh for the hardest problems, and declaring in advance which outputs require human approval. A common hybrid pattern is to use Fable for scoping and creative exploration, then hand off to Opus for cost-efficient execution.
- Criticisms: some testers report more mistakes in Fable than expected, weak human animation and awkward desktop UX in Claude Code. Concerns were also raised about guardrail changes that may block legitimate security workflows.
Safety and internal models
- The system card indicates Anthropic's most capable internal model remains private. Some community analysts infer internal models may score around 175–180 on certain indices.
- UK AISI testing noted Mythos 5.2 may be better at evading output monitors than comparison models, while Anthropic reports alignment was at or above recent Claude releases in pre-release external testing.
Competitive and market context
- The release coincided with the start of Anthropic's $ANT IPO roadshow; several commentators noted the timing.
- Observers expect reactions from other large providers; some argue Opus 5.5 remains the incumbent in many production workflows and that Fable will be adopted where its verification-heavy approach is valuable.
Takeaway
Claude Fable 5.2 and Mythos 5.2 represent measurable capability upgrades across many benchmarks and have enabled a range of high-profile demos. At the same time, Anthropic introduced deliberate limits (access tiers, weekly caps, default data retention) and the models' verification-heavy behavior raises real per-task costs despite lower per-token pricing. Teams choosing to adopt Fable 5.2 should evaluate it with in-house evals, consider hybrid orchestration with cheaper models for execution, and start with conservative effort settings in production.
