AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Nov 10, 2026

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, billing both as a step above Opus 5.5 in nearly all evaluated capabilities. The launch lands the same week Anthropic's IPO roadshow began. Fable 5.2 reclaimed the top spot on independent leaderboards from Gemini 4 Carbon, which had held it for about ten days after displacing Opus 5.5.

Pricing falls to $8 per million input tokens and $40 per million output tokens, down from Fable 5.1's $10 and $50. Per Artificial Analysis, cache writes cost $10 with a five-minute TTL or $16 with a one-hour TTL, cache reads cost $0.20, and the context window is 1 million tokens with text and image input. There are five effort levels: low, medium, high, xhigh and max. Medium is the default in the Claude apps and in Claude Code.

The cheaper sticker price does not translate into cheaper tasks. SemiAnalysis's GDPx benchmark of professional work ranked Fable 5.2 first at 16.4%, but its cost per task climbed 31%, from $67.16 to $88.15. Its run logs point to why: where Opus 5.5 and Fable 5.1 stopped after roughly 20 test cases, 5.2 sometimes generated hundreds or thousands and spawned subagents to verify its own work independently. Artificial Analysis puts cost per task at max effort at about $26.5, roughly 2.6 times Opus 5.5's $10.0 and twice Fable 5.1's $12.99. At high and xhigh it scores 62 while using about 98k output tokens per task, against 173k for Opus 5.5 at max.

That gap is the practical story for anyone budgeting inference. Lower per-token rates reward teams that constrain effort and output volume; long-horizon agentic workloads that let the model self-check can end up costing more than the model it replaces. Anthropic's own frontier results make a similar point: on FrontierCode 2.0 it reports Fable at High matched Opus 5.5 at Max for about half the cost per task.

Independent benchmark results put Fable 5.2 at the front of most boards. On the Artificial Analysis Intelligence Index v5 it leads at 65, ahead of Gemini 4 Carbon at 60, Opus 5.5 at 59, GPT-6 Astra at 57 and Fable 5.1 at 49. It set new records on HLE-Diamond (80% against a prior best of 69%), AA-Omniscience v1.1 (67 against 54), Terminal-Bench-Science (73% against 65%), CritPt v2 hard (77% against 72%) and FrontierCode 2.0 (49% against 46%). The Epoch Capabilities Index recorded 174 against 169 for GPT-6 Astra, and ARC Prize reported 81.2% on ARC-AGI-3 under the standard harness at $21.9K, 97.7% under the provider adapter harness at $18.6K, 93.8% on ARC-AGI-2 and 98.5% on ARC-AGI-1. Vals Index ranked it first, leading seven benchmarks across coding, finance and science, with high effort outperforming xhigh and max. CursorBench gave it 60% at max and 56.9% at default, versus 49.2% for Fable 5.1. Arena Code: WebDev placed it first at 1832 at xHigh, and AA-3D placed it first at 1271 Elo with no confidence-interval overlap with any other model, 132 points above Fable 5.1.

The weaknesses are narrower but real. AA-Briefcase gives it 1785 Elo, within the confidence interval of Opus 5.5's 1800 and behind Sonnet 5.5's 1813. GDP.pdf scored 24% all-pass against 26% for Opus 5.5. Terminal-Bench 4.0 came in at 60%, level with Opus 5.5 and behind Kimi K3.1 Pro at 67%, Gemini 4 Carbon at 65% and Sonnet 5.5 at 64%. AutomationBench-AA was 76%, ahead of Opus 5.5's 70% but behind Gemini 4 Carbon's 78%. SemiAnalysis reports clear gains in chip design, software engineering, biology and materials science, but not in finance, legal, quant trading, hardware design or AI research automation.

The effort-level behavior deserves caution. Multiple evaluations show max effort underperforming high effort, and commenters noted that only the Fable models exhibit this dip. Vals found high the best effort setting, and Cursor advised starting at the default rather than carrying over settings tuned for 5.1, since effort levels behave differently. Anthropic staff suggested starting at medium or high and reserving xhigh for the hardest problems. This suggests treating effort as a per-workload parameter to measure rather than a dial to leave at maximum, especially given the cost curve.

Anthropic deliberately limited capability in AI research automation. SemiAnalysis called the RSI result a reflection of an explicit nerf, and the system card states that Anthropic's most capable model remains internal. Independent commenters expect nothing better than Opus 5.5 at LLM training to ship publicly for months, and some researchers running automated research loops say the restrictions push them toward open models. An assessment from the UK AI Security Institute found Mythos 5.2 substantially more capable of evading an output monitor than comparison models.

Access is tiered. Ultra and Team Premium subscribers receive Fable resets scaled by tenure: one under six months, two for six to twelve months, three above twelve months, usable until March 8. Pro and Team Standard get a one-week trial capped at half the weekly limit plus one reset. Fable can consume at most 50% of a weekly limit. Data retention is required by default so Anthropic's safety classifiers can run, with zero data retention available only to eligible enterprise customers. One analysis of 1.1 million logged requests on a $200 Ultra plan estimated the plan yields about $15,000 a month of Opus 5.5 value but only about $3,000 of Fable 5.2.

Day-one availability covers Cursor, GitHub Copilot and VS Code, Perplexity Computer and Search, Factory, the Hermes Agent through Nous Portal and OpenRouter, and Arena and DesignArena. Anthropic published guidance on choosing between Fable, Opus, Sonnet and Haiku, including when to raise effort rather than switch models. Other vendor claims include up to 30% faster output than the 5.1 models, a run in which 1,000 Fable 5.2 agents formalized the existing proof of Fermat's Last Theorem in 30 hours against 26% completion in 18 hours for Opus 5.5, and Mythos 5.2 accurately modeling 28 of 34 gene-therapy capsids. Anthropic also described an OS built from an empty folder by more than 1,000 subagents over eight days.

Practitioner reports split. Enthusiasts describe it one-shotting playable games and producing convincing 3D scenes, with one tester calling it a new bar for game feel. Skeptics note a build that cost about $42 and did not clearly outperform Opus 5.5, and engineers describe a split workflow: Fable for ambiguous, large-scope work and Opus for well-specified tasks. Anthropic staff said they often still prefer Opus 5.5 for everyday use because Fable thinks more slowly.

Why it matters: product teams now face a model that leads most benchmarks but can cost more per completed task than its predecessor despite a 20% price cut, so evaluation should track cost per task and effort setting rather than token price alone. Teams with hard data-retention constraints or heavy agentic verification loops should check whether they qualify for zero retention and test high effort against max before deploying. The deliberate limits on AI research automation may also steer independent researchers toward open-weight alternatives.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

November 9, 2026 — Claude Fable 5.2 and Mythos 5.2 arrive on Anthropic’s IPO roadshow day. How much thinking is worth paying for?