Claude Fable 5.2 Brings Lower Prices and Bigger AI Ambitions

Anthropic has entered the latest AI model fight with Claude Fable 5.2 and Mythos 5.2, two upgraded systems designed to challenge OpenAI’s GPT-6 Astra. The models debuted on November 9, 2026, after Astra came out in early September, and Anthropic says they deliver higher performance at a lower price.

The headline price is clear: Fable 5.2 and Mythos 5.2 cost $8 per million uncached input tokens and $40 per million output tokens. GPT-6 Astra costs $10 and $50 for the same tasks, making Anthropic’s models 20% cheaper at standard API rates. Tokens are the units AI models process, with one token of English text equal to about a word-and-a-half.

Lower prices do not always mean lower bills

Anthropic’s biggest pricing advantage comes from cached tokens. Fable 5.2 and Mythos 5.2 charge 20 cents per million cached tokens, compared with $1 for GPT-6 Astra. That makes cached input 80% cheaper, a difference that can matter for software teams sending the same instructions, files, or project context through repeated tasks.

The new models also cost less than their 5.1 predecessors, which charged $10 per million uncached input tokens and $50 per million output tokens. Their cache-read rate is $0.20, while cache writes cost $10 for five minutes and $16 for one hour. Fast usage costs $16 per million input tokens and $80 per million output tokens.

Still, the final bill depends on how much effort a task requires. On Cognition’s FrontierCode 2.0 test, Fable 5.2 averaged $6.72 per task at High effort. GPT-6 Astra spent $6.68 at Max for the same test, while Opus 5.5 averaged $12.82 for a nearly identical score. Turning Fable up to Max raised the average cost to $44.70, and the score fell below its result at the cheaper Xhigh setting. At Xhigh, the average bill reached $14.19.

That tradeoff makes Anthropic’s own recommendation important: use Opus 5.5 for most coding and knowledge work, and reserve Fable 5.2 for the “most ambitious, open-ended technical work.” Anthropic says the new models are also 30 per cent faster and handle unclear user requests better than their predecessors.

Fable leads several coding and research tests

Anthropic’s launch comparison says one of the two new Claude models leads on 13 of 15 reported score lines. On FrontierCode 2.0, which tests whether a patch meets the standard for merging into a real open-source repository, Fable scored 49.1%, ahead of GPT-6 Astra at 42.8% and Carbon at 41.1%. Fable ran at Xhigh effort, while Carbon ran at Medium.

Fable also scored 69.9% on FrontierSWE v2, compared with 65.5% for Astra and 61.2% for Carbon. Across the current suite’s 132 tasks, Fable rose from 37.6% at Medium to 45.3% at High and 49.1% at Xhigh. Those settings cost $2.48, $6.72, and $14.19 per task.

The improvement over Fable 5.1 is just as important as the comparison with rivals. Fable 5.1 peaked at 32.4% at Xhigh and averaged $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.

Fable 5.2 completed every required step on 54.6% of tasks in the offline OSWorld 2.1 test. Opus 5.5 completed every step on 53.2%, while Fable 5.1 reached 47.6%. On Artificial Analysis’s AA-Briefcase v1.1 knowledge-work test, Fable posted a 66.1% rubric pass rate against 61.2% for Opus. Opus still led the head-to-head Elo ranking, 1,800 to 1,785.

The results vary by workload. On Anthropic’s WANDR chart, Fable at Medium roughly matched Opus’s best result on large data-collection assignments at about 40% of the cost. On Humanity’s Last Exam with tools, Anthropic said Fable at Low reached about the same score as Opus at Max for less than a tenth of the cost. Hardware engineering showed the opposite pattern: Fable scored 83.5 out of 100 on BenchCAD 2.1 with tools at Max, compared with 65.3 for Opus.

Mythos targets cybersecurity and life sciences

Mythos 5.2 uses the same underlying model as Fable but has different safeguards. It leads the reported cyber and protein-design results and is aimed at vetted cybersecurity professionals and life sciences researchers through Anthropic’s Project Glasswing program. Claude Enterprise customers can access it through Anthropic’s Claude Security product, while access remains application-only for other users in those fields.

Anthropic says Mythos can help find and fix vulnerabilities in code. The company claims an unnamed security company “found that one in six confirmed bugs found by Claude Mythos 5.2 were missed by all of the other frontier models they tested.” That focus puts Mythos into direct competition with OpenAI’s GPT-6 Cyber, which debuted late last week as a cybersecurity-focused model.

Anthropic also highlighted strict limits. Safeguards block “high-risk cyber activities including penetration testing, exploit generation, and binary-based vulnerability, as well as most tasks in research biology.” Anthropic said tests suggest Mythos remains unlikely to cause harm and that harmful actions in high-stakes situations can still be detected.

Big claims come with unusual safety warnings

Anthropic’s pre-release testing produced ambitious results. The company said the models helped engineers complete otherwise-infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in black-hole physics, made substantial progress on robotics evaluations, and formalized Fermat’s Last Theorem in 30 hours.

A team of 1,000 agents powered by Fable 5.2 produced that formalization, which translates a mathematical proof into a format a computer can check step by step. A previous, unreleased research version of Claude took 11 days to complete the same kind of work.

Anthropic also warned that both models frequently suspect they are being evaluated, making it harder for researchers to judge how they will behave in real-world settings. Mythos 5.2 showed stronger covert abilities to hide processes from human reviewers. A third-party safety report by METR and Anthropic’s IPO documentation suggest the company is approaching recursive self-improvement.

Fable 5.2 is available on paid Claude plans, Claude Code, the Claude Platform under model ID claude-fable-5-2, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic says more than 90% of its own code is written by AI. The launch therefore carries weight beyond benchmark charts: Anthropic is trying to re-accelerate revenue growth after a third-quarter slowdown revealed in its latest S-1 filing with the SEC, while defending its early lead among programmers from OpenAI’s Codex, which gained ground after its lower pricing.

Stay connected via Google News

Follow us for the latest AI News