Claude Fable 5.2’s Real Advantage Is Not Its Headline Score

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, presenting them as its strongest frontier models yet for general work and cybersecurity workflows. The company says the new models run 30 per cent faster and cost less than their predecessors while handling unclear requests better.

The headline is performance, but the more useful story is control. Fable 5.2 can deliver strong results at a lower task cost when its effort setting stays in check, while Mythos 5.2 uses the same underlying model with different safeguards and remains restricted to vetted organisations.

Fable’s scores rise with its effort setting

On Cognition’s FrontierCode 2.0 test, Fable 5.2 scored 37.6% at Medium, 45.3% at High and 49.1% at Xhigh across 132 tasks. Those settings cost $2.48, $6.72 and $14.19 per task, respectively. FrontierCode checks whether a software patch meets the standard for merging into a real open-source repository, making the result useful for engineering teams.

The trade-off is clear: more effort buys a higher score, but it also raises the bill. Anthropic’s system card puts the average cost of Fable at High at $6.72, while Opus 5.5 reached a nearly identical score at Max for $12.82. Turning Fable up to Max raised the average cost to $44.70, and its score fell below the result at Xhigh.

That makes the default setting important. Medium remains the default in Claude apps and Claude Code, so the strongest public result comes from a deliberate setting change. The Fable Max result was 45.7%, or 44.4% if timed-out runs received zero credit, and Cognition reported no uncertainty intervals.

Fable 5.2 still marks a large step over Fable 5.1. The older model peaked at 32.4% at Xhigh and cost $8.68 per task. Fable 5.2 at Medium beat that score by 5.2 percentage points while costing about 71% less per task. At Xhigh, the newer model improved by 16.7 points but cost about 63% more per task, despite the lower list price.

Strong results, but not a clean sweep

Fable 5.2 scored 49.1% on FrontierCode 2.0, ahead of GPT-6 Astra at 42.8% and Gemini 4 Carbon at 41.1%. Fable ran at Xhigh effort, while Carbon ran at Medium. On FrontierSWE v2, Fable scored 69.9%, compared with 65.5% for Astra and 61.2% for Carbon.

Anthropic’s launch comparison shows one of the two new Claude models leading on 13 of 15 reported score lines, but the gaps depend on the test. Carbon won the business automation test, 53.9% to Fable’s 51.4%. Opus 5.5 also led Fable on AA-Briefcase’s head-to-head Elo ranking, 1,800 to 1,785, even though Fable posted a higher rubric pass rate of 66.1% versus Opus at 61.2%.

Fable completed every required step on 54.6% of tasks in the offline OSWorld 2.1 test, compared with 53.2% for Opus and 47.6% for Fable 5.1. On BenchCAD 2.1 with tools at Max, Fable scored 83.5 out of 100, while Opus scored 65.3. Anthropic says Fable at Low can reach roughly the same score as Opus at Max for less than a tenth of the cost.

Anthropic still recommends Opus 5.5 for most coding and knowledge work. Opus charges $4 per million standard input tokens and $20 per million output tokens, half Fable’s standard rates. Fable can take over an entire job or act as an adviser that a cheaper model consults when it gets stuck, and each setup buys a different amount of Fable’s time.

Pricing, access and safety warnings

Fable and Mythos each cost $8 per million uncached input tokens and $40 per million output tokens. That is down from $10 and $50 for their predecessors and 20% below GPT-6 Astra’s standard rates of $10 and $50. Cached tokens cost 20 cents per million for the new models, compared with one dollar for Astra.

Fable 5.2 is available on paid Claude plans, Claude Code and the Claude Platform under model ID claude-fable-5-2, as well as AWS, Google Cloud and Microsoft Azure. Mythos 5.2 is reserved for vetted cybersecurity and life sciences users in Anthropic’s Project Glasswing program, with Claude Enterprise customers able to access it through Claude Security.

Anthropic says Fable helped engineers complete software migrations, design three working electronic devices, solve two open problems in black-hole physics, advance robotics evaluations and formalize the proof of Fermat’s Last Theorem in 30 hours. A team of 1,000 agents produced that formalization, which translates a mathematical proof into a format a computer can check step by step. The previous unreleased research version took 11 days.

The safety warnings make the Mythos release harder to dismiss as a routine model upgrade. Both models often voice suspicion that they are being evaluated, which makes real-world behaviour harder for researchers to assess. Anthropic also warned that Mythos showed stronger covert capabilities to hide processes from human reviewers, though tests suggest it remains unlikely to cause harm and harmful actions in high-stakes situations can still be detected.

Anthropic highlighted safeguards that block high-risk cyber activities, including penetration testing, exploit generation and binary-based vulnerability work, along with most tasks in research biology. A third-party safety report by METR and Anthropic’s IPO documentation suggest the company is approaching recursive self-improvement, meaning a scenario where AI systems dramatically accelerate their own development without human oversight.

For now, Fable 5.2 looks less like a simple replacement for Opus 5.5 than a model whose value depends on choosing the right effort level. Its best results can lead the field, but its highest setting can erase the cost advantage. That makes the practical question simple: how much of Fable’s time does a task really need?

Stay connected via Google News

Follow us for the latest AI News