Anthropic's Claude 5.2 models undercut GPT-6 Astra by 20% on price

Anthropic had a difficult autumn. In September it shipped Claude Fable 5.1 and Mythos 5.1, and in October the smaller Claude Haiku 5.5. OpenAI's GPT-6 Astra landed in early September, and Codex kept taking enterprise coding work on price. Then Anthropic's updated S-1 filing with the US Securities and Exchange Commission showed third-quarter revenue growth slowing and a $567 million operating loss.

The 5.2 generation reads like an answer to that filing. The models are faster, the token prices are lower, and the marketing leads with a benchmark where Anthropic can claim first place.

Two models, one base

Fable 5.2 and Mythos 5.2 share the same base architecture. The difference is in the guardrails and in who is allowed to run them.

Fable 5.2 is the general-purpose model. Anthropic positions it for coding, multi-agent reasoning and long technical tasks, and it is available through AWS, Google Cloud and Microsoft Azure. Mythos 5.2 runs with relaxed safety restrictions and is restricted to vetted organisations through dedicated cybersecurity and life-sciences verification programmes. If you are a normal API customer in Europe, Fable is the one you can actually buy.

Anthropic says output generation on the 5.2 models is up to 30% faster than on the 5.1 generation. It also says more than 90% of its own merged internal code is now written by AI. That second number comes from the vendor and has no independent audit behind it.

Benchmarks

On FrontierSWE v2, a test of agentic software engineering where the model has to work through a repository rather than answer a single question, Anthropic reports 69.9% for Fable 5.2. GPT-6 Astra is listed at 65.5%, Gemini 4 Carbon at 61.2%, and the previous Fable 5.1 at 56.3%. The jump over its own predecessor is the largest gap in that list.

The cybersecurity figure is narrower. Anthropic says Mythos 5.2 rediscovered 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on its CVE Rediscovery test. Fable 5.2 scores 64.9% on RoboDojo, a robotics control benchmark.

The demonstration Anthropic is pushing hardest involves 1,000 coordinated Fable 5.2 agents that formalised the proof of Fermat's Last Theorem in 30 hours. The theorem itself was proven by Andrew Wiles in 1994, so the claim concerns machine-checkable formalisation, not a new proof. Anthropic also lists three working electronic hardware devices designed from a single paragraph prompt and a production migration from Python to Rust that cut boot time roughly 26-fold. All of these are company demos. None has been reproduced publicly.

What it costs

Here is the published per-million-token pricing, with EUR converted at about $1.16 per euro (the rate at the time of writing, so treat the euro column as approximate):

The cached-input line is where the discount is biggest. Anthropic charges $0.20 per million cached tokens against $1.00 for Astra, an 80% cut. For an agent that resends the same long system prompt and tool definitions on every turn, that is where most of the bill sits.

Take a workload of 2 million input tokens and 500,000 output tokens a month, with 1.5 million of the input served from cache. On Fable 5.2 that comes to $0.30 for cached input, $4.00 for fresh input and $20.00 for output, so $24.30, roughly €20.95. On GPT-6 Astra it is $1.50 plus $5.00 plus $25.00, so $31.50, roughly €27.16. The real gap on this shape of traffic is closer to 23% than 20%.

Fast mode doubles both line items. At $16 and $80 per million, it is 60% more expensive on output than Astra's standard rate, so it only pays for itself in latency-sensitive pipelines.

The cheaper tiers are elsewhere

The 20% figure compares Anthropic's flagship with OpenAI's flagship. It says nothing about the rest of the market. OpenAI's GPT-6.1 Sol sits at one-fifth of Astra's price, $2 input and $10 output. xAI's Grok 4.7, released in September, is $2 and $6. Meta's Muse Ember is listed at $0.10 per million input tokens. DeepSeek-V4.1-Flash and Zhipu's GLM-5.3-Flash are open weights.

If your task is classification, extraction or routing, Fable 5.2 is not the model you should be pricing against. Anthropic is competing for agentic coding and long-horizon reasoning work, where the benchmark gap matters more than the per-token rate.

Availability in Europe

Fable 5.2 is reachable from EU regions of AWS, Google Cloud and Azure, which is how most European companies will consume it. Data residency depends on the region and contract you configure with the cloud provider, not on Anthropic's marketing page, so verify it in your own agreement before moving production traffic.

Billing is in US dollars, which leaves EU buyers with exchange-rate exposure and reverse-charge VAT handling. The AI Act angle is more concrete. Since 2 August 2026, Article 50 transparency obligations are binding, including detectable marking and disclosure of AI-generated content, and the EU AI Office has direct enforcement powers over general-purpose model providers. If you deploy Fable 5.2 in an EU-facing product that produces synthetic text at scale, the disclosure duty sits with you as deployer. The AI Omnibus passed in mid-2026 pushed the Annex III high-risk deadline to 2 December 2027, so the compliance clock for high-risk uses is longer than originally planned.

On the open-weight side, Mistral Large 4 shipped its weights on 3 November under the Mistral Ecosystem Open-Weight License. For teams that cannot send data outside the EU, that remains the simpler path, even if it does not match Fable 5.2 on FrontierSWE v2.

The IPO context

Anthropic's third quarter brought $17 billion in revenue, up from $1.4 billion a year earlier, alongside that $567 million operating loss. The company is reportedly targeting a raise of $69.3 billion to $100 billion at a valuation above $1 trillion. A 20% price cut on the flagship, three months after the previous generation launched, is what growth pressure looks like when the S-1 is already filed.

For anyone running models in production, the practical question is narrower. The prices are public, the benchmarks are self-reported, and the fastest way to settle it is to run your own eval on your own traffic. That is the approach we use on the AI Arena rig for local models, and it applies just as well to a $40-per-million API.