Fable 5.2 vs Fable 5.1: Price, Speed, Benchmarks & Which to Use

‍

Anthropic released Fable 5.2 on November 9, 2026. It is as cheap, as fast, and as good as any row in the benchmarks Anthropic has published in its table. Anthropic has released three Fable versions in quick succession: Fable 5 in June, Fable 5.1 in September, and Fable 5.2 in November, giving you a successor before your next budget cycle. For a smooth migration, Fable 5.2 is the best pick from the Claude series.

This guide compares Fable 5.2 and Fable 5.1, including price, speed, benchmarks, effort, and protection.

Fable 5.2 vs Fable 5.1 at a glance

With the widest gaps in robotics, remote labor, and agentic science, Fable 5.2 outperforms the other variants on price, speed, reasoning, agentic coding, and benchmarks.

These are all Anthropic’s own figures, produced with safeguards enabled and the model’s highest available effort unless otherwise noted. Treat them as vendor claims until independently reproduced. Source: Anthropic

Price and speed: Fable 5.2 costs 20% less and writes up to 30% faster

Fable 5.2 costs 20% less than Fable 5.1 per token. It costs $8/million input tokens and $40/million output tokens, whereas $10/million and $50/million, respectively. Cache reads cost $0.20 per million tokens. Cache writes cost $10 per million tokens.

Fable 5.2 is also around 30% faster than Fable 5.1. Testers also found it communicates more clearly. In long agent runs, time matters most because each step depends on the previous one.

Fast mode is optional. Generates tokens up to 2.5 times faster. Input tokens cost $16/million, and output tokens cost $80/million. Only pay for it if latency is more important than cost.

The token price is not the total bill. A per-task cost varies with the effort level chosen. One independent analysis measured Fable 5.2 at $3.10 per task at low effort, $3.80 at medium, $9.08 at high, $14.83 at extra high, and $26.54 at max.

It also determined that Fable 5.2's cost at High effort is less than Fable 5.1's Max effort, and that the Intelligence Index cost for Fable 5.2 is 62 compared to 49 for Fable 5.1. Take these as third-party figures and not Anthropic's.

A basic illustration of the token math. Assume a job consumes 2M input tokens and produces 200k output tokens. On Fable 5.1, that costs $30 (2 × $10 plus 0.2 × $50). On Fable 5.2, it costs $24 (2 × $8 plus 0.2 × $40). That's the same 20% savings before any effort changes the token count.

Benchmarks: Fable 5.2 leads Fable 5.1 in every published row

In Anthropic’s benchmark table, Fable 5.2 scores higher than Fable 5.1 in every row. The cyber and protein-design rows compare the Mythos versions of the same model.

Source: Anthropic. Scores use each model's highest available effort unless noted, with safeguards on.

The higher the percentage, the more correct answers the model achieved on the test. The control indicator for the robots increased by 37.3 points. The number of people who can read charts increased by 28.6 points. The relative scores for remote knowledge work increased by 23.4 points. The reading scores of scientific research agents increased by 22.0 points. The smallest increase in this table was for computer use at 5.8 points.

Search and reasoning scores also increased. Fable 5.2 achieved an 82.1% score on Perplexity's WANDR benchmark, which evaluates agents responsible for large data collection tasks, matching Fable 5.1's 68.7%. It achieved 74.2% accuracy on Humanity's Last Exam tools, while Fable 5.1 achieved 65.6% accuracy.

It also improved hardware design. Fable 5.2 scored 83.5 in BenchCAD 2.1, which measures how well each model's designs match reference gear and circuits, when run at maximum effort. Fable 5.1 reached 44.4.

Be careful when interpreting these numbers. The tests were run by Anthropic, and each benchmark measures a very specific skill. A high score on one does not mean you will have the same score on your workload.

Also Read: Fable 5.2 benchmarks for what each test measures and how it compares with Opus 5.5 and GPT-6 Astra.

Effort levels: Retest your Fable 5.1 settings

Fable 5.2 has a different effort behavior than Fable 5.1; do not copy the old behavior. According to Cursor's docs, the two models differ in how they handle effort levels.

Fable 5.2 features five difficulty settings: low, medium, high, extra high, max. Always thinks before answering and determines thinking time; there is no thinking toggle. For the most difficult and longest tasks, use Save extra high and max.

Medium is the default in Claude apps and Claude Code. Also, at Medium, Fable 5.2 can complete harder tasks with fewer tokens than 5.1. So, Medium is a good place to begin.

Effort directly correlates with actual improvement. On FrontierCode 2.0, Fable 5.2 scored 45.3% at high effort. Opus 5.5 scored 45.5% at its best. The cost per task was about half that of Fable 5.2. According to Cursor, the model has a 300k context window and a maximum of 1M tokens.

Also Read: Claude Fable 5.1 high vs low effort for how effort changed cost and results on the earlier model

Safeguards: Fable 5.2 blocks about 55% fewer requests

Fable 5.2's safeguard power remains the same as Fable 5.1, except it can block fewer benign requests. The cyber and bio protections are as strong as ever, and there's less "noise," according to Anthropic.

Claude Code blocks around 55% overall, the average across early tests on claude.ai, Claude Code and Cowork. Cyber stocks dropped 60%. Biology blocks dropped by some 40%. The classifiers still block high-risk cyber work like penetration testing and exploit generation, as well as most research biology tasks. If a cyber query is blocked, it will be sent to Opus 4.8. If the biology query is blocked, it is sent to Opus 5.

The same underlying model is used for Fable and Mythos. Mythos 5.2 lacks these protections, and Anthropic restricts its use to vetted organizations through its Cyber and Life Sciences Verification Programs. The CVE Rediscovery test gave Mythos 5.2 a score of 89.6%, compared to 78.3% for Mythos 5.1.

In addition, Anthropic states that it found it safer to run tests on Mythos 5.2. In an adversarial sandbox test, it used credentials or files outside the sandbox in 3.1% of cases. Mythos 5.1 did so in 21% of cases. Anthropic says both models express suspicion more often than recent models. It makes practical use of the tests difficult.

Fable 5.2 introduces better protection against distillation, which occurs when one model is trained to learn from another model's output. Anthropic claims its methods reduce the data "distillers" harvest by 90 to 99%.

Real-world results: Long agent tasks show the biggest gap

Fable 5.2 is much better suited to dealing with long, multi-agent work that was less well done in Fable 5.1. There are samples everywhere, as seen in Anthropic's launch post.

1. Fermat's Last Theorem. A team of 1,000 agents in Fable 5.2 formalized the proof in 30 hours. It took Anthropic 11 days to formalize the first proof.
2. An operating system. One engineer queried Fable 5.2 to create a prototype OS from an empty folder. It managed more than 1000 subagents for eight days. The outcome was booting on a laptop.
3. Faster physics code. According to Discovered Materials, Fable 5.2 made Quantum ESPRESSO's phonon code 36 times faster. It succeeded on its 13 benchmark cases, which had taken 3.3 hours but were reduced to less than 4 minutes in the new release, 3 to 4 times faster than Fable's speedup in 5.1.

The results shown are exemplary, not average. They reflect ceiling performance, not averages.

Upgrade checklist: Five checks before you move from Fable 5.1

Run five tests before switching, since the models are different. Each check focuses on one of the differences mentioned above.

• Update the model ID: Fable 5.2 uses claude-fable-5-2.
• Reset your effort level: You should NOT save your Fable 5.1 setting. Start at medium, and only increase it when results are poor.
• Measure cost per task: Look at the total cost of the completed task, not just the token price.
• Review blocked requests. Although fewer blocks are likely, the classifiers still block high-risk cyber and biology work.
• Conduct a single long run. Execute an agent task that takes several hours and then compare to your Fable 5.1 baseline.

Which model to choose: Fable 5.2 for new work

Download Fable 5.2 unless you have tested and know why you need to remain on Fable 5.1. Everywhere I looked for a lead for Fable 5.1, the list of sources didn't show one. Anthropic rates Fable as its most ambitious, open-ended technical work.

Also inquire if the Fable model is necessary at all. For most coding and knowledge tasks, Anthropic recommends starting with Claude Opus 5.5. It recommends upgrading to Fable or Mythos or downgrading to Sonnet 5.5 or Haiku 5.5, as appropriate to the work. Fable 5.2 can be used as an Advisor, a smaller model, or an Orchestrator that delegates simpler tasks to smaller models. To build that split yourself, an AI agent builder lets you describe the lead agent, the helpers, and where a person reviews.

Availability is broad. Fable 5.2 is available on Amazon Web Services, Google Cloud and Microsoft Azure. The model ID for the API is claude-fable-5-2 (Anthropic). During the initial launch, Pro plans and regular Team seats were also eligible for Fable for up to 50% of the weekly limits as part of a promotion. Review the latest terms before making plans.

Conclusion

Fable 5.2 is the stronger choice for most new work. It is less expensive, faster and higher in every row Anthropic published. Before changing effort settings, re-test settings and set efforts to the task level. Use Fable when it matters; otherwise, stick to Opus 5.5.

Fable 5.2 is now live on Emergent! Have an app idea? Use Fable 5.2 on Emergent to turn your concepts into functional products. Simply describe your application in a text prompt, and let Emergent’s AI agents handle the development of complete web and mobile applications. Start building with Fable 5.2 today!