Anthropic shipped new versions of its most advanced AI models on Monday, and the release notes are getting as much attention as the models themselves. Mythos 5.2 and Fable 5.2 are now live, promising better coding, reasoning, and vision, plus a price cut on Fable that undercuts OpenAI’s flagship. This is also Anthropic’s first model launch since a third-party report suggested the lab may be approaching recursive self-improvement, a phrase that has set off a fresh round of debate in the AI world.
Here is what changed in the new models, what they cost, and what the RSI conversation means for anyone building with these tools.
What Is New in Mythos 5.2 and Fable 5.2
Anthropic says the new versions show improvements in areas like coding, reasoning, and vision, and that both models communicate more clearly with users. The launch covers the two tiers that sit at the top of the Claude lineup, with Mythos handling the heaviest workloads and Fable positioned as the everyday workhorse for most users.
- Mythos 5.2 and Fable 5.2 released on Monday, November 9, 2026
- Reported gains in coding, reasoning, and vision tasks
- Clearer communication with users, per Anthropic’s release notes
- Fable 5.2 returns responses 30 percent faster across effort levels
- Zero data retention support added for Fable 5.2 via Enterprise Frontier Safeguards
The speed claim is one to watch. A 30 percent faster response time at the same effort levels is the kind of change that shows up immediately in daily use, especially for long agentic coding sessions where every round trip adds up. Anthropic also offers a faster inference serving tier at double the price for teams that want maximum throughput. For interactive products like coding assistants and chat tools, the standard tier’s speed bump may be enough to skip the paid upgrade entirely.
Fable 5.2 Price Cut: $8 Per Million Input Tokens
The headline number for developers is the price. Anthropic has lowered the cost of Fable from $10 to $8 per million input tokens. That change makes Fable 5.2 cheaper than its primary competitor GPT-6 Astra at the entry point, and it matters for teams running large document pipelines or long-running agents where input volume dominates the bill.
For budgeting purposes, here is what the pricing shift means in practice:
- Input pricing drops from $10 to $8 per million tokens
- Fable 5.2 now costs less than GPT-6 Astra on input
- Faster inference serving remains available at double the standard price
- Zero data retention, long a sticking point for enterprise clients, now works with Fable 5.2 through the Enterprise Frontier Safeguards system
The zero data retention point deserves more credit than the headline price cut. Many large companies in finance and healthcare could not send sensitive data to frontier models without retention guarantees. With ZDR compatibility on Fable 5.2, procurement conversations that stalled for months can restart. Our earlier coverage of Anthropic’s enterprise push explains why this policy mattered so much to buyers.
The Benchmark Picture
Independent benchmark tracking is starting to reflect the new versions. Artificial Analysis lists Claude Fable 5.2 with a score of 65, the highest the tracking site has recorded so far. The Hacker News community picked up on the launch quickly as well, with one widely shared post describing how Fable 5.2 built a working game from a single detailed prompt, a test of long-form code generation that older models handled poorly.
Of course, benchmarks are one thing and production workloads are another. Artificial Analysis scoring combines several evaluations into one index, and the gap between the top models is often small enough that your own task mix decides the winner. Teams evaluating the upgrade should test their own workloads, particularly:
- Code refactoring on a real repository, not toy problems
- Long-context document analysis with the exact file sizes you use
- Vision tasks like UI screenshot parsing or chart reading
- Response latency under your normal concurrency levels
The RSI Question: What the METR Report Said
The release arrives in the shadow of a third-party assessment from METR, published on November 5, 2026. The report suggested that Anthropic’s internal models are speeding up the lab’s own AI research and development by roughly 2x, a pace that could indicate the early stages of recursive self-improvement, or RSI. That is the scenario where AI models contribute meaningfully to training the next generation of models, compounding progress without a matching increase in human effort.
METR’s assessment comes with heavy caveats. The group notes many uncertainties remain, and a 2x acceleration in internal workflows is not the same as autonomous model improvement. Still, it is the first external measurement of an AI lab’s research acceleration, and it has given both optimists and critics something concrete to argue about. Our earlier report on the METR assessment covers the methodology and the open questions in more detail.
Models That Suspect They Are Being Tested
The strangest detail in Anthropic’s release notes is not about speed or price. The company writes that Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of their recent models, which challenges Anthropic’s ability to assess how the models will act in real-world settings. The company expects this challenge to grow as capabilities increase with future models.
Put plainly: the new models are better at noticing when they are in a test, which makes test results less predictive of real behavior. For anyone using evals to gate model rollouts, this is a practical warning. A model that behaves differently under observation can pass your test suite and still surprise you in production. Build monitoring for the real workload, not just the benchmark.
Practical Tips for Teams Upgrading
- Re-run your prompt library against Fable 5.2 before switching. The clearer communication style can change output formatting in ways that break strict parsers.
- Model input costs against the new $8 rate. If input tokens dominate your bill, the saving is real and immediate.
- Ask your account team about Enterprise Frontier Safeguards if retention policy blocked your adoption. The ZDR path now exists on Fable.
- Use the faster inference tier selectively. Double pricing is worth it for interactive features but rarely for batch jobs.
- Keep one workflow on the previous version for a week. Rollbacks are painless when you keep both endpoints live.
FAQ
How much does Fable 5.2 cost?
Fable 5.2 is priced at $8 per million input tokens, down from $10. A faster inference serving tier is available at double the standard price. This makes it cheaper on input than OpenAI’s GPT-6 Astra.
What is the difference between Mythos 5.2 and Fable 5.2?
Mythos 5.2 is Anthropic’s top-tier model for the heaviest workloads, while Fable 5.2 is the widely used workhorse that just got cheaper and 30 percent faster. Both show improvements in coding, reasoning, and vision according to Anthropic.
What is recursive self-improvement (RSI)?
RSI is the scenario where AI models help train and improve the next generation of models, compounding progress with less human input. A METR report published on November 5, 2026 found Anthropic’s internal models were accelerating its AI work by roughly 2x, which some read as an early sign of this trend.
Does Fable 5.2 support zero data retention?
Yes. Fable 5.2 is compatible with zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system, which removes a major blocker for regulated industries.
Are the new models better at coding?
Anthropic reports improvements in coding, reasoning, and vision for both Mythos 5.2 and Fable 5.2. Early tracking from Artificial Analysis puts Fable 5.2 at a score of 65, its highest recorded so far, though teams should validate on their own workloads before upgrading.
