Anthropic started its IPO roadshow and released Claude Fable 5.2 and Mythos 5.2, which Anthropic describes as a step above Opus 5.5 in almost all evaluated capabilities. Fable 5.2 retook the world No.1 spot from Gemini 4 Carbon after about ten days at the top. Token prices fell 20% to $8 per million input tokens and $40 per million output tokens, yet measured cost per professional task rose. For business buyers this split between cheaper tokens and pricier tasks matters more than the leaderboard.
Fable 5.2 benchmarks, pricing and access terms
Artificial Analysis puts Fable 5.2 first on its Intelligence Index v5 at 65 points, with 80% on HLE-Diamond, 67 on AA-Omniscience v1.1 and 73% on Terminal-Bench-Science. Epoch recorded a Capabilities Index of 174 against 169 for GPT-6 Astra, while SemiAnalysis GDPx shows 16.4% for professional tasks. Pricing is $8 and $40 per million input and output tokens, with cache writes at $10 and $16 and cache reads at $0.20. Five effort levels are offered: low, medium, high, xhigh and max, with medium as default in Claude apps and Claude Code.
The cost pattern comes from how the model works: Fable 5.2 runs far more checks than its predecessors. Where Opus 5.5 and Fable 5.1 tried about 20 test cases, version 5.2 sometimes writes hundreds or thousands and launches subagents for independent verification. That approach lifted GDPx to first place but raised cost per task to $88.15 from $67.16, a 31% increase despite lower token prices. At max effort, cost per task reaches about $26.5, roughly 2.6 times Opus 5.5. On FrontierCode 2.0, Fable 5.2 at High matched Opus 5.5 at Max for about half the cost per task.
The release coincides with the IPO roadshow and a crowded frontier-model race involving Gemini 4 Carbon, GPT-6 Astra, Sonnet 5.5 and Kimi K3.1 Pro. Fable leads on Arena Code WebDev at 1832 and on AA-3D at 1271 Elo, but trails on some slices such as Terminal-Bench 4.0 and AutomationBench-AA. Anthropic also applied deliberate limits around recursive self-improvement, and GDPx shows no gain in finance, legal, quant trading and hardware design. Distribution runs through GitHub Copilot, VS Code, Perplexity Computer and Search, Hermes Agent, OpenRouter and Arena. The timing links a product milestone to investor marketing.
What Fable 5.2 changes for AI adoption
For companies that run coding, science and engineering workloads, the practical gain is higher task completion at controlled effort settings. CursorBench reports 60% at max and 56.9% at default against 49.2% for Fable 5.1, with fewer steps and tool calls in Copilot and VS Code. Reported demonstrations include formalizing the proof of Fermat's Last Theorem with 1,000 agents in 30 hours, modeling 28 of 34 gene-therapy capsids, and building an OS from an empty folder over eight days. Small teams can stay on medium or high effort, while large organizations can standardize high as the balance of quality and spend.
The limits change procurement and deployment checks. Fable 5.2 requires data retention by default for safety classifiers, with zero retention only for eligible enterprise customers. Subscription access is capped: Ultra and Team Premium users receive resets by tenure until March 8, Pro and Team Standard get a one-week trial at half the weekly limit, and Fable can consume only 50% of a weekly limit. Multiple evaluations show high outperforming xhigh and max, so higher spend does not guarantee better results. Buyers should test effort levels from default, track tokens per task, and confirm retention terms before rollout.
A useful confirmation signal will be sustained independent cost-per-task data for high versus max effort through the first quarter of deployment. If high keeps matching prior max-level quality at lower spend across coding and science benchmarks, the efficiency claim holds for budgets. If max remains costly without stable gains, vendors will need clearer guidance on when extra reasoning pays. The next GDPx and Artificial Analysis updates will show which pattern lasts.
