Claude Fable 5.2 Is Here: 20% Cheaper, 30% Faster — and It Built an Operating System in 8 Days
Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on Monday 9 November 2026, and the headline numbers are hard to ignore: the new model costs 20% less than its predecessor, runs up to 30% faster, and scores 49.1% on FrontierCode 2.0, ahead of every rival Anthropic lists. Claude Fable 5.2 is Anthropic's most capable model for long, difficult technical work, built for long technical work rather than everyday chat. Here is what changed, what the benchmarks really say, and who should actually use it.
What Fable 5.2 and Mythos 5.2 actually are
Fable 5.2 and Mythos 5.2 are the same underlying model, split into two versions with different safety rules. Fable 5.2 is the generally available version. It carries cybersecurity and biology safeguards, and it is the one developers and businesses will use through the Claude Platform, Claude apps, Claude Code, and the cloud platforms AWS, Google Cloud and Microsoft Azure and Foundry. Mythos 5.2 is the same model with fewer restrictions, and it goes only to vetted organisations inside Anthropic's Cyber and Life Sciences Verification Programs. Ordinary users cannot sign up for Mythos, and Anthropic has not opened any public route to it.
For most readers, Fable 5.2 is the relevant product. Anthropic still recommends Claude Opus 5.5 as the default daily driver for most people, and Fable 5.2 sits above it, as a specialist tool for jobs that run for hours or days.
The price cut in numbers
The pricing change is the most concrete part of the launch. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. That is a 20% cut on both sides, and it arrives at the same time as the performance gains, which is unusual. Normally a new model costs more; here Anthropic is charging less for a better model, which puts direct pressure on every rival flagship.
There is also an optional fast mode, priced at $16 per million input tokens and $80 per million output tokens, which Anthropic says generates tokens up to 2.5 times faster. Cache reads cost $0.20 per million tokens. TechTimes' write-up of the announcement also notes a 1-million-token context window with up to 128,000 output tokens, which fits the model's design for very long tasks.
| Model and mode | Input price per million tokens | Output price per million tokens |
|---|---|---|
| Claude Fable 5.1 | $10 | $50 |
| Claude Fable 5.2 | $8 | $40 |
| Claude Fable 5.2 fast mode | $16 | $80 |
| Cache reads (Fable 5.2) | $0.20 | — |
For teams running long agent sessions, the cut compounds: a multi-day run costs a fifth less, and output arrives up to 30% faster. The comparison that matters most for buyers is in our Gemini Carbon vs Claude Opus 5.5 vs GPT-6 Astra post, where Fable 5.2 now changes the value side of the equation. Anthropic has cut prices before, including the Claude Haiku 5.5 price reduction, but this is the first time the price cut has landed on its flagship reasoning model.
Benchmarks: what the numbers say, and whose numbers they are
Every benchmark figure in this section comes from Anthropic's own published table, and that needs saying plainly. No independent benchmark body has verified these results yet, and Anthropic chose which rivals and which tests to show. On FrontierCode 2.0, a test of hard coding and research tasks, Fable 5.2 scores 49.1%, ahead of Opus 5.5 at 45.5%, GPT-6 Astra at 42.8%, Gemini 4 Carbon at 41.1% and Fable 5.1 at 32.4%.
The lead holds across most of the table. On FrontierSWE v2, a software-engineering benchmark, Fable 5.2 scores 69.9% against Opus 5.5's 62.3%. On RoboDojo Sim, a robotics simulation test, it jumps from 27.6% for Fable 5.1 to 64.9%. Anthropic does not win every row, though: on AutomationBench, Gemini 4 Carbon scores 53.9% against Fable 5.2's 51.4%.
| Benchmark | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 62.3% | — | — |
| RoboDojo Sim | 64.9% | — | — | — |
| AutomationBench | 51.4% | — | — | 53.9% |
The pattern across the table is consistent: the biggest gains are on long, multi-step work, not on quick single answers. That matches the model's design goal. Readers who want the wider context can compare these figures with the Gemini 4 Carbon launch and the GPT-6.1 Terra price story.
The 1,000-agent experiments
Alongside the benchmarks, Anthropic described a series of internal experiments in which roughly 1,000 Fable 5.2 agents worked together on single projects. These were Anthropic-designed tests, not independent proof, and they should be read as demonstrations of what long-horizon agent work can look like, not as the model inventing things on its own. The most striking result is a working prototype operating system built from an empty folder in eight days. It boots on a laptop, with a login screen, file storage and windows.
The second result is more remarkable in a different way. The same kind of agent swarm formalised Fermat's Last Theorem in the Lean proof language in about 30 hours. The first formalisation of that proof took 11 days of human effort, but the agents were formalising a known proof, not discovering new mathematics from scratch. The team also reports solving two open black-hole physics problems and designing three working electronic devices with the hardware firm Diode: a speakerphone, a voice-assistant hub and a handheld virtual pet. What these experiments really show is stamina: a thousand agents coordinating for eight days without falling apart is a different capability from answering a single hard question well.
Safeguards and the Mythos question
The launch also brings a change to how the model's safeguards behave. Anthropic says false-positive blocks are down by about 55% compared with Fable 5.1 in early testing, with cyber blocks down around 60% and biology blocks down around 40%. Those are Anthropic's early-testing figures, not independently replicated results. Blocked cyber queries may be routed to Opus 4.8, and blocked biology queries to Opus 5.
Anthropic also says Fable 5.2 carries its strongest-ever anti-distillation protections, aimed at stopping other labs from copying the model's behaviour by training on its outputs. The Mythos version, with fewer restrictions, stays behind the vetting wall for exactly this reason. Anyone promising a public Mythos sign-up is not telling the truth, and readers should treat such claims as scams.
Who should use Fable 5.2, and how to get it
For most people, the honest answer is Opus 5.5, which Anthropic still recommends as the default. Fable 5.2 is for the hardest work: long coding projects, research that spans days, autonomous agents and scientific computing. If a job runs for a week, the 20% price cut, the 30% speed gain and the benchmark lead all start to matter, and Fable 5.2 is the right choice.
Access is straightforward. The API model id is claude-fable-5-2, and it is available on the Claude Platform, in Claude apps and Claude Code, and through AWS, Google Cloud and Microsoft Azure and Foundry. Claude Duet users will find the same Anthropic product family logic at work: different models for different jobs, with the flagship reserved for the hardest ones.
Key Takeaways
- Claude Fable 5.2 launched 9 November 2026, costs $8/$40 per million tokens, 20% less than Fable 5.1, and runs up to 30% faster.
- On Anthropic's own benchmark table it scores 49.1% on FrontierCode 2.0, ahead of Opus 5.5 (45.5%), GPT-6 Astra (42.8%) and Gemini 4 Carbon (41.1%) — but no independent body has verified these figures yet.
- Fable and Mythos are the same model; Mythos has fewer restrictions and goes only to vetted organisations, not general users.
- About 1,000 Fable 5.2 agents built a prototype OS in 8 days and formalised Fermat's Last Theorem in about 30 hours in Anthropic's own internal experiments.
- Anthropic still recommends Opus 5.5 as the daily default; Fable 5.2 is for the hardest, longest-running technical work.
Conclusion
Claude Fable 5.2 is a genuine step forward on the numbers Anthropic has published: cheaper, faster and better on long tasks than its predecessor, with a benchmark lead over its named rivals on most of the table. The caveats matter just as much. The benchmarks are Anthropic's own, the agent experiments were Anthropic-designed, and the Mythos version is not available to the public. For developers and researchers, it is a stronger model for less money, and a sign the frontier price war is now fought on value, not just capability.
Comments
Post a Comment