Blog

Compare Claude Fable 5.2 token prices, speed and task cost

Anthropic's November 9 launch gives AI learners a practical way to read token prices and speed claims without confusing them with the cost of successful work.

ยท 3 min read

A model can have a lower price per token without being the cheapest way to complete your task. It can also generate tokens faster without finishing an entire job proportionally sooner. Anthropic's November 9 launch of Claude Fable 5.2 and Mythos 5.2 is a useful case for artificial intelligence (AI) certification learners because the announcement gives separate figures for price, output speed and performance on particular tasks.

Read the two prices before comparing models

Anthropic lists standard developer pricing at $8 per million input tokens and $40 per million output tokens. Tokens are the units used to measure a model's input and generated output; the two sides have different rates. Its optional fast mode is listed separately at $16 per million input tokens and $80 per million output tokens. These are the announcement's developer token rates, not a description of the allowance included in a subscription.

For a deliberately simplified arithmetic example, assume one request has 10,000 billable input tokens and 2,000 billable output tokens at those listed rates. Standard mode gives an input subtotal of $0.08 and an output subtotal of $0.08, or $0.16 together. Using the same hypothetical counts at the fast-mode rates gives $0.32. This is a calculation under stated assumptions, not a measured request or an estimate of every charge for a complete application.

Output speed is not a whole-workflow stopwatch

The announcement says Fable 5.2 and Mythos 5.2 generate output up to 30% faster than their 5.1 predecessors. It separately says optional fast mode speeds token generation by up to 2.5 times. Keep the words up to and token generation attached to those claims. They do not state that every multi-step job, including its checks and reviews, finishes at the same improved rate.

For a learning exercise, draw a small workflow with an initial response, a tool action and a reviewer decision. Mark where output generation occurs and where it does not. Then ask which times you would need to measure before promising an end-to-end improvement. This is a proposed measurement exercise, not a report that we ran those models.

Compare accepted work, not just a rate card

Anthropic's evaluation discussion ties some cost comparisons to particular tasks and effort settings. For example, it reports that Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort for roughly half the cost per task on FrontierCode 2.0, a coding evaluation concerned with whether repository maintainers would accept changes. That is evidence reported for a particular test, not a universal price-performance guarantee.

The company also recommends starting with Opus 5.5 for most coding and knowledge work, then choosing a different model as the task requires. The launch therefore does not support a rule that the newest or most capable model should handle every request.

A comparison sheet for learners

Choose a small, representative set of fictional tasks and write an acceptance rule before looking at any model's answer. In a later authorized experiment, record the model and mode, billable input and output counts, elapsed time and whether the result met that rule. Record retries and reviewer corrections rather than ignoring unsuccessful attempts. This is a suggested evaluation design, not a claim about observed results.

Your explanation should answer three distinct questions: what a unit of model usage costs, how quickly output arrives and what it takes to obtain acceptable work. Use the official launch figures as the dated starting point, and keep any conclusions about a different workload conditional on evidence from that workload.

Browse the Nex Arc study guides

Sources