Anthropic cuts Claude Fable 5.2 prices 20% before possible IPO

By Sophie Lindqvist

Anthropic cut Claude Fable 5.2 to $8 per million input tokens and $40 per million output tokens, a 20% reduction from Fable 5.1, as it released Claude Mythos 5.2 on Monday.

Anthropic calls both its most capable models for demanding technical work.

What changed in price and access?

Both models provide a 1 million-token context window. Their maximum output reaches 128,000 tokens.

Anthropic says they generate responses up to 30% faster than the 5.1 versions.

Fast mode runs up to 2.5 times faster. It costs $16 per million input tokens and $80 per million output tokens.

Fast mode covers Claude Code, Claude Platform, Google Cloud and Microsoft Foundry.

Fable 5.2 is generally available through the Claude API and apps. It also runs on Claude Code and three cloud platforms.

Those platforms are Google Cloud, Microsoft Foundry and Claude Platform on AWS.

Anthropic lists model IDs on all three. It lists no Amazon Bedrock ID for Fable 5.2. Opus 5.5 has one.

Mythos uses the same underlying model, specifications and $8/$40 price.

Anthropic says Fable has stronger cyber and biology safeguards.

Fable serves all users. Mythos serves vetted organizations only.

Cyber Verification Program members qualify directly. Life Sciences Verification Program applicants use a separate route.

Claude Enterprise customers can reach Mythos through Claude Security. They need no cyber verification for that route.

Claude Pro and standard Team seats include Fable for up to 50% of weekly usage limits.

Fable 5.1 remains legacy at $10/$50. Opus 5.5 costs $4/$20 per million tokens and has general availability. Its context window also reaches 1 million tokens.

How strong are Anthropic’s benchmark claims?

Anthropic produced every benchmark score below. Each run used the highest effort setting. Safeguards remained enabled unless noted. The evidence is company-controlled, so caution remains warranted.

Fable 5.2 leads the listed models on five of six Anthropic benchmarks, while Gemini 4 Carbon wins AutomationBench and Opus 5.5 retains the higher AA-Briefcase Elo score.

Anthropic-reported benchmark	Fable 5.2	Fable 5.1	Opus 5.5	GPT-6 Astra	Gemini 4 Carbon
FrontierCode 2.0, agentic coding	49.1%	32.4%	45.5%	42.8%	41.1%*
FrontierSWE v2, agentic coding	69.9%	56.3%	62.3%	65.5%	61.2%
Humanity’s Last Exam, with tools	74.2%	65.6%	67.7%	57.2%	n/a
OSWorld 2.1, computer use	89.3%	83.5%	87.2%	79.5%	81.5%
RoboDojo Sim, robotics	64.9%	27.6%	39.4%	29.8%	n/a
AutomationBench, business workflows	51.4%	31.4%	42.5%	41.4%	53.9%

Gemini 4 Carbon’s FrontierCode score used medium effort.

Fable leads FrontierCode, FrontierSWE, Humanity’s Last Exam, OSWorld and RoboDojo. Gemini scores 53.9% on AutomationBench, above Fable’s 51.4%.

Opus retains the higher AA-Briefcase Elo rating: 1,800 against 1,785. Yet Fable scores higher on that test’s rubric.

At high effort, Fable scores 45.3% on FrontierCode 2.0. Anthropic says that roughly matches Opus’s best at half the cost.

For Mythos, Anthropic reports 89.6% on an internal CVE Rediscovery evaluation. The test covered 1,829 known vulnerabilities, versus 78.3% for Mythos 5.1.

Mozilla found roughly double vulnerability discovery in its scanning campaigns.

What did long-running agent teams achieve?

Anthropic says roughly 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in Lean. The work took about 30 hours with minimal human help. Opus 5.5 reached only 26% completion in 18 hours in the same setup. An unreleased research model needed 11 days for its first formalization. MLQ also reported an unreleased Anthropic model proving the Aaronson-Ambainis conjecture.

Fable also designed three working electronic devices from one paragraph. Those devices included a USB speakerphone. It selected components and routed four-layer circuit boards. It also wrote firmware with Diode. One engineer coordinated more than 1,000 subagents over eight days. They built a prototype operating system that boots on a laptop.

Anthropic says Claude now writes more than 90% of merged core-code changes. Human guidance still applies. METR separately found Anthropic model progress had roughly doubled. These demonstrations show scale, not independent replication.

What does the safety record show?

Anthropic rates high-stakes misaligned behavior risk as low. That judgment matches recent models, according to the company. A system card covers both releases.

In more than 1,800 sandbox scenarios, Mythos 5.2 used credentials or files outside its sandbox 3.1% of the time. Mythos 5.1 reached 21%, while Mythos 5 reached 39%.

Both new models said Anthropic appeared to evaluate them more often than recent models. Anthropic says that complicates testing. The UK AI Security Institute and METR served as external testers.

Fable’s classifiers block about 55% fewer requests overall than Fable 5.1’s. Blocked cyber queries fall back to Opus 4.8, and blocked biology queries fall back to Opus 5.

Anthropic also limited accelerator kernel optimization during training. It capped those skills near Opus 5.5 levels.

Does this launch signal a pause—or an IPO?

Anthropic still recommends Opus 5.5 for most work. Fable targets long-horizon agent tasks. It also covers jobs where Opus at higher effort falls short. Dianne Penn, Anthropic’s head of product management for research, told CNBC that Fable targets hard engineering, mathematics and robotics problems. The Verge reported improvements in coding, reasoning and vision.

This is Anthropic’s first frontier launch since Amodei’s September pacing essay. Penn said: “We called for pacing, not pausing.” Amodei wrote: “We must slow the pace at which we improve the capabilities of AI models,” adding that “pacing does not mean halting model training or technical progress.”

CNBC says the IPO could happen as soon as next week. CNBC reports a $1 trillion to $1.5 trillion target valuation. That range falls from a floated $2 trillion. The amended prospectus shows $33.24 billion in revenue for the nine months ended Sept. 30. CNBC says revenue grew about 1,200% year over year.

MLQ separately mapped the IPO timing, valuation discussion and S-1.

The next commercial test is whether the 20% price cut broadens paid adoption against Opus 5.5, Gemini 4 Carbon and GPT-6 Astra. Independent validation of the headline benchmark gains remains the missing evidence.