Fable 5.2 pricing: Is it worth more than Opus 5.5?

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens in November 2026. That is 20% less than Fable 5.1, but twice the standard input and output rates of Claude Opus 5.5. Fable leads Opus on some benchmarks; the price of a completed task depends on how much work each model does and which effort setting you use.

How much does Fable 5.2 cost?

The table below uses November 2026 standard API prices. Input tokens are the text and other content sent to the model; output tokens are what it generates.

Sources: Anthropic Fable pricing, Anthropic Opus 5.5, and Anthropic cache pricing.

Caching changes the comparison. A Fable 5.2 cache read costs 2.5% of its standard input price, and the cache read price is the same as Opus 5.5's. Our guide to the Fable 5.1 cache read price cut covers why repeated context can change an agent's bill.

Fable 5.2 batch requests cost $4 per million input tokens and $20 per million output tokens, the same as standard Opus 5.5. Fable 5.2 fast mode costs $16 and $80 respectively; Anthropic's documentation lists both rates.

Anthropic also offers US-only inference at 1.1 times the input and output token price.

Fable 5.2 has a 1 million token context window and can generate up to 128,000 output tokens, according to Anthropic's documentation. Anthropic says its 5.2 models generate output up to 30% faster than Fable 5.1.

There is also a data retention condition. Anthropic's model overview says Fable 5.2 requires 30-day data retention and is unavailable under zero data retention unless Anthropic expressly authorizes it.

Does Fable 5.2 beat Opus 5.5?

On Cognition's FrontierCode 2.0 benchmark, Anthropic reports that Fable 5.2 at xhigh effort scored 49.1%, against Opus 5.5's best score of 45.5%. Artificial Analysis also put Fable ahead on its Intelligence Index. Other tests put Opus ahead.

The FrontierCode scores are Anthropic-reported results on Cognition's benchmark, and the AA-Briefcase figures also come from Anthropic's benchmark table. The Intelligence Index is an independent test. A score can help you shortlist models, but only results and costs on your own tasks show whether Fable is worth the extra spend.

The size of the gain also varies by task. Anthropic's results put the Remote Labor Index at 41.3% for Fable 5.2 at high effort versus 21.3% for Opus 5.5. The FrontierCode gap is 3.6 percentage points.

Anthropic itself recommends starting most coding and knowledge tasks on Opus 5.5, then moving to Fable or Mythos as needed. We would use the same starting point: keep a set of harder tasks for the Fable comparison, alongside everyday tasks.

Why can a cheaper model cost more per task?

A lower token price can be offset if the model uses more tokens per task. Fable 5.2 can use more tokens on a task than Fable 5.1 or Opus 5.5, particularly at higher effort settings.

Take a job with 10 million input tokens and 1 million output tokens, with no cache reads, priced at the standard rates above.

Equal token counts make the arithmetic clean, but a real task does not have to use the same number of tokens on each model. Artificial Analysis measured 275 million generated tokens across its index for Fable 5.2 at max effort, compared with a median of 72.8 million for comparable models.

Artificial Analysis also reports what it cost to run each model through its Intelligence Index at max effort.

At max effort, the 20% token price cut did not produce a lower bill per task: Fable 5.2 cost about 2.6 times as much as Opus 5.5.

How does effort change FrontierCode cost?

At high effort, Anthropic reports that Fable 5.2 scores about the same as Opus 5.5's best FrontierCode result for roughly half the cost per task. At xhigh effort, Fable 5.2 reaches 49.1%.

Cognition describes the FrontierCode score as a weighted aggregate of rubric items, not a share of tasks solved.

Our guide to cost per task and token economics walks through the same decision for other workloads.

What do production rankings show?

List prices do not include the effect of each application's cache usage. On Requesty's published rankings, the production data from September 7 to November 9, 2026 shows the following results for Opus 5.5 and Fable 5.1.

Blended cost is the charged price per million tokens after cache discounts. Time to first token is the wait before a response starts; a 5xx error is a server error. These are observations across production requests, not a controlled test of the same tasks on both models.

Cache hits can bring the bill well below an estimate based on undiscounted list prices. Measure cache savings alongside task success when testing Fable 5.2. Our Opus 5.5 guide covers its API and pricing in more detail.

Which model should you use?

Run both models on the same set of tasks, record the effort setting and count a task as solved only after it passes your checks. Divide total spend, including failed attempts, by the number of tasks solved.

Worked example: 100 tasks at max effort

Take the Artificial Analysis max-effort costs: $26.54 per task for Fable 5.2 and $10.04 for Opus 5.5. Over 100 tasks, that is 100 x $26.54 = $2,654 for Fable and 100 x $10.04 = $1,004 for Opus, an extra $1,650.

If Fable solved 10 more tasks, each additional solve would cost $165 ($1,650 divided by 10). The 10 extra solves are an example, not a benchmark result.

Keep the failed tasks: they are the best candidates for deciding whether escalation to Fable pays for itself. Requesty's cost tracking records spend per call and model, which can help you total the cost of each task across the supported models you test.

Run both models through Requesty

Fable 5.2 is available on Requesty as anthropic/claude-fable-5.2 at Anthropic's list price plus Requesty's 5% fee, so you can run the same tasks on it and on anthropic/claude-opus-5-5. Because Fable 5.2 requires 30-day data retention, an organization that requires zero-data-retention models in Requesty's compliance settings cannot route to it.

Requesty's fallback policies retry a request that fails (for example, an error) on the next model in your chain. Escalating a completed but wrong answer to Fable is a separate check you run in your own application.