Anthropic models: Claude pricing, thinking budgets and cache accounting

Sonnet versus Opus is an acceptance-test decision

Use Sonnet as one candidate for regular review and drafting, then give Opus a task where deeper analysis could change the outcome. A useful code comparison includes a known defect and a harmless change. Ask both models to identify the defect, explain its consequence and avoid inventing problems with the harmless change. Counting confident comments is not enough; count verified findings, false positives and the time needed to check the proposed fix.

Fable's editorial use case here is evidence-heavy analysis and long-form writing. Test it on sources that disagree, requiring quotations and a separation between reported facts and the writer's inference. Haiku is the candidate to try for bounded extraction or quick transformations. For an invoice, the right behavior on a missing field is an explicit unknown, not a plausible invention. None of those suggested roles establishes a quality ranking among these four Claude generations.

Four Claude rates, one comparable workload

Anthropic's rate card lists base input/output prices per million tokens of $10/$50 for Fable 5, $5/$25 for Opus 5, $2/$10 for Sonnet 5 and $1/$5 for Haiku 4.5. With 8,000 ordinary input tokens and 2,000 output tokens, the arithmetic is $0.18, $0.09, $0.036 and $0.018 respectively. These examples exclude cache operations, extra tools and other billing categories; they compare a fixed token workload rather than an equal-quality finished answer.

A Claude consumer subscription and Anthropic API billing are different purchases. Connecting a developer key should not be budgeted as though every request comes out of the consumer plan's monthly allowance. Review the actual account entitlement and usage dashboard. If a team member already pays for Claude, decide whether the needed feature is its consumer workspace or a separately metered API workflow before canceling or adding anything. The product comparisons linked below address that migration question.

Claude cache reads are not ordinary input tokens

Anthropic reports cache_creation_input_tokens, cache_read_input_tokens and input_tokens separately. Its documented total input is the sum of those three categories. That is an important implementation distinction: treating input_tokens as though it already contains all cached input understates the request, while adding categories twice overstates it. For cost reconciliation, apply the appropriate rate to each category rather than multiplying the combined total by ordinary input pricing.

The standard five-minute cache write costs 1.25 times base input, a one-hour write costs twice base input, and cache reads for the four listed generations cost one tenth of base input. For Sonnet 5, writing a 20,000-token prefix at the five-minute rate costs $0.05; reading those tokens later costs $0.004, before the rest of the request. That example explains why reuse matters and why the first request is not the same price as a warmed repeat.

Eligibility also differs by Claude model. The current guide lists a 512-token minimum for Fable 5 and Opus 5, 1,024 for Sonnet 5, and 4,096 for Haiku 4.5. Below the minimum, a marked prompt can be processed without caching and without an error. Do not transfer a Sonnet cache assumption to Haiku merely because the messages look identical. Confirm cache fields in usage; this guide does not promise that every chat request configures explicit cache breakpoints.

Haiku's dated ID and thinking budget deserve separate checks

The Haiku identifier ends in 20251001. Preserve that complete identifier when reproducing an extraction result rather than referring vaguely to the Haiku family. Its configured context is 200,000 tokens, whereas the Fable, Opus and Sonnet entries here use one million. A workflow that fits the larger entries may need document selection or chunking before moving to this Haiku entry. The shared vision flag does not remove that capacity difference.

Our bundled Haiku definition uses budget-style reasoning; the other three Claude definitions use effort-level behavior. Check the setting available in the actual workspace rather than assuming a control transfers unchanged. When assessing a long analytical answer, inspect missing evidence and invented citations before rewarding verbosity. The per-model exercises below keep the old pages' useful testing detail together, where differences between Claude tiers can be considered without opening four near-identical reference pages.

Concrete Anthropic model checks

Suggested evaluation protocols, not executed benchmark results.