Claude Fable 5.2 Explained: Pricing, Benchmarks, and When to Use It
Fable 5.2 delivers stronger benchmark results at lower token prices. Here is what its gains, limitations, effort settings, and launch offers mean for choosing it over Opus.

Updated November 11, 2026. Based on official documentation, published evaluations, and attributed user testing; this is not an independent hands-on review.
Claude Fable 5.2 gives users a new reason to reconsider which Claude model they use. Its benchmark results show substantial gains, its token prices have fallen, and paid subscribers have a limited opportunity to try it with additional usage. Choosing it still requires attention to the task, reasoning settings, and access restrictions.
For most everyday coding and knowledge work, Opus 5.5 remains a sensible starting point. Fable deserves a closer look when difficult tasks repeatedly require corrections or exceed what Opus can reliably deliver. The practical question is where its improvements translate into better finished work.
What is Fable 5.2, and how does it differ from Mythos?
Fable 5.2 and Mythos 5.2 share the same underlying model. Fable adds cybersecurity and biology safeguards for broader availability; direct Mythos access requires verification. Broader access does not mean every paid subscription includes unrestricted Fable usage. Anthropic announcement
This distinction matters for specialist work. A result reported for Mythos should not automatically be treated as a capability that an ordinary Fable user can access under the same conditions.
Fable 5.2 pricing and speed
Fable 5.2's standard input and output prices are 20% below Fable 5.1's, but twice those of Opus 5.5. These are API token prices, separate from subscription usage allowances. Caching and batch processing have separate rates. Fable specifications, Opus specifications, Fable 5.1 pricing
Model and mode | Input per million tokens | Output per million tokens |
|---|---|---|
Fable 5.1 — standard | $10 | $50 |
Fable 5.2 — standard | $8 | $40 |
Opus 5.5 — standard | $4 | $20 |
Fable 5.2 — fast | $16 | $80 |
Anthropic reports up to 30% faster output than Fable 5.1. Optional fast mode offers up to 2.5 times the token-generation speed at premium pricing. Neither figure guarantees the same reduction in total task time. Speed and fast-mode details
For budgeting, measure the cost of an accepted result. Token volume, failed attempts, tool use, and human corrections can matter as much as the advertised unit price.
Benchmarks: where Fable 5.2 improves
At the time of checking, Fable 5.2 at max scored 65 on the Artificial Analysis Intelligence Index, leading the ranking. Gemini 4 Carbon at high scored 60, while Fable 5.1 at max scored 49—a five-point lead over Carbon and a 16-point gain over the earlier Fable configuration. These comparisons use the current index; scores from historical index versions should not be mixed. Fable 5.2 results, release comparison, Fable 5.1 results
The factual-knowledge result is also striking. Epoch AI reports 94% on SimpleQA Verified, versus approximately 76% for the strongest alternatives shown in its selected comparison. That is evidence of better performance on this test; it does not establish a larger parameter count, which remains undisclosed. Epoch AI results
On the Remote Labor Index, Fable completed 41.3% of projects, compared with Opus 5.5's 21.3%. The near-doubling makes complex deliverables worth investigating, without implying that every user's productivity will double. RLI results
Limitations: automation, healthcare, and long context
Fable does not lead every benchmark. On AutomationBench, its 51.4% trails Gemini 4 Carbon's 53.9%. FrontierCode also needs an effort-level footnote: Fable's 49.1% uses xhigh, while Carbon's 41.1% uses medium. That is not an equal-budget comparison. Published benchmark table
HealthBench requires another distinction. Fable scored 58.8% raw, lowest among the five Claude models in that comparison. After length adjustment, it ranked behind only Sonnet 5.5 within that group. Its raw score declined from 60.1% at low to 58.8% at max, but no confidence intervals were provided to establish whether that decline exceeded noise. System card, section 8.16.1
ProgramBench exposed a different weakness: Fable could exhaust its context before submitting. Showing the remaining context after each step raised its score from 85.6% to 91.6%, reversing its deficit to Fable 5.1. System card, section 8.10.1
These findings make the surrounding software part of the purchasing decision. For a long-running agent, budget visibility and a reliable submission process can determine whether its reasoning produces a usable deliverable.
Effort levels: high, xhigh, or max?
Fable supports five effort levels: low, medium, high, xhigh, and max. Its defaults differ by interface: medium in Claude apps and Claude Code, and high in the API. Check the active setting rather than assuming the same model behaves identically everywhere. Launch documentation, Claude Code settings
For demanding API tasks, Anthropic recommends starting at high. Much of the capability is already available there; move to xhigh when your own evaluations show a worthwhile improvement. Max is intended for exceptional cases where additional time and tokens are acceptable. Effort guidance
Artificial Analysis illustrates the trade-off: high and xhigh both display an Intelligence Index score of 62, while their measured costs per index task are $9.08 and $14.83 respectively. Those are benchmark costs, not estimates for an ordinary chat. High results, xhigh results
Simon Willison's pelican-on-a-bicycle test provides a memorable failure case. At max, Fable consumed 128,000 tokens thinking and failed to return the SVG; xhigh produced his preferred result. One test cannot establish overall reliability, but it shows why more reasoning does not guarantee a completed answer. Simon's test
Fable 5.2 vs Opus 5.5: when should you switch?
Anthropic's current recommendation is more specific than always starting at high: begin most work with Opus 5.5 at medium, increase effort when needed, and consider Fable when Opus at high still falls short. Opus 5.5 remains Claude Code's default for direct Claude subscriptions and the Anthropic API. Official model-selection guide
Before switching after a failure, check whether the model had the relevant files, clear instructions, working tools, and a way to verify its output. Missing context can derail either model.
Our suggested workflow is to use Fable for difficult planning and judgment, then hand clearly defined implementation work to Opus in a separate session or subagent. Include the constraints, relevant files, chosen approach, and acceptance tests in the handoff.
This is a strategy to measure, not a guaranteed saving. If the task is small, planning and handoff can cost more than they save. Keep the split when it reduces corrections, improves completion, or lowers total spending.
Access and free resets for paid subscribers
Eligible Max and Ultra subscribers receive resets based on subscription length:
Subscription length | Fable resets |
|---|---|
Under six months | 1 |
Six to twelve months | 2 |
Over twelve months | 3 |
Pro users can use Fable within up to 50% of their weekly allowance through November 16, 2026, at 11:59 p.m. Pacific Time. Qualifying existing paid accounts also receive one reset, which expires at the same time. Max and Ultra launch resets expire March 8, 2027, at 11:59 p.m. Pacific Time. Reset eligibility requires paid billing before November 9, 2026, at 10 a.m. Pacific Time. Check the official terms for Team and Enterprise seat eligibility. Launch-offer terms
A Fable reset restores weekly usage spent on Fable. For accounts with a five-hour limit, it also refills that entire window, including usage from other models. Consequently, no prior Fable usage does not necessarily mean a completely wasted reset. Check both usage meters before redeeming. The reset restores usage; it does not increase your limits. Reset mechanics
AI research restrictions to know before choosing Fable
Anthropic says it limited training advances in accelerator kernel optimization and language-model architecture improvement, keeping those capabilities around Opus 5.5's level in both Fable and Mythos. Users working in those areas should not assume the overall benchmark gains apply equally to their research. Capability restrictions
The system card's separate 1.88× figure describes a fitted change in the slope of Anthropic's capability-progress trend, with uncertain attribution to AI. It does not establish that deploying Fable 5.2 made researchers 1.88 times as productive. For model selection, evaluate the released version on the work you actually need to complete, using the settings and safeguards you will deploy. System card, section 2.3
ings
