BlockBeats reports that Anthropic has released a guide on Claude model selection and cost optimization

AT A GLANCE

According to media reports, Anthropic recommends choosing models based on the task, and states that low-intensity execution with escalation after failure can reduce testing costs.

Article

Anthropic Releases Claude Cost-Saving Guide: Start with Low Intensity, Escalate Only on Failure, Save 41% and Improve Accuracy

Dongcha Beating AI News, Anthropic has released a Claude model selection guide, advising developers not to default to the strongest Fable 5.2. Most programming and knowledge work should start with Opus 5.5's Medium reasoning intensity. Feature development and daily automation with clear requirements can use the cheaper Sonnet 5.5; simple tasks such as classification and information extraction can be left to Haiku 5.5. For large-scale code migration, complex scientific research, or tasks that Opus struggles to complete, consider Fable 5.2.

Anthropic believes that an effective way to reduce costs is to let the AI check its own results. For example, run automated tests when programming, and make corrections when errors are found, rather than using the highest reasoning intensity from the start.

Officially, on the same subset of the SWE-bench Pro programming test, Opus 5.5 was first run with Low reasoning intensity, and only failed tasks were retried with High. The result was a pass rate of about 97%, with an average cost of 0.17 USD per problem; using High directly for everything yielded a pass rate of 95.3% and a cost of 0.29 USD per problem. The former reduces cost by about 41%, while the pass rate is slightly higher.

There is also a ceiling to the cost-effectiveness of simply increasing reasoning intensity. Raising Opus 5.5 from High to Xhigh improved its SWE-bench Pro score by only about 1.4 percentage points, while the cost became 2.5 times the original. Anthropic suggests that if a model still cannot solve difficult problems after reaching High, one can try a stronger model at a lower reasoning intensity.

For example, in the FrontierCode 2.0 programming evaluation, Fable 5.2 using High achieved 45.3%, nearly matching Opus 5.5's 45.5% when using Max, while the average cost per problem was only about half of the latter. Although Fable's API Token unit price is twice that of Opus, on such high-difficulty tasks the total cost of completing the task may actually be lower.

Original link https://m.theblockbeats.info/flash/376565

Key points

The original text states that Anthropic recommends most programming and knowledge work start with Opus 5.5's Medium reasoning intensity, rather than using Fable 5.2 by default.
The original text states that Sonnet 5.5 can be used for feature development with clear requirements and daily automation, while simple tasks such as classification and information extraction can be handled by Haiku 5.5.
In the test on the same subset of SWE-bench Pro as described in the original text, Opus 5.5 first used Low and then retried with High after failure, achieving a pass rate of about 97% at an average cost of 0.17 USD per task; directly using High yielded a pass rate of 95.3% at a cost of 0.29 USD per task, meaning the former approach reduced costs by about 41%.
The original text states that Opus 5.5, upgraded from High to Xhigh, achieved only about a 1.4 percentage point improvement on SWE-bench Pro, while costing 2.5 times as much.
The original text states that in FrontierCode 2.0, Fable 5.2 using High achieved 45.3%, while Opus 5.5 using Max achieved 45.5%, and the former's average cost per question was about half of the latter's.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

The idea introduced in this message is: do not directly use the strongest model or the highest reasoning intensity for all tasks. Instead, first execute at a lower intensity, check the results through automated testing, and only increase the intensity after a failure. The text also emphasizes that when comparing costs, one should look at the total cost of completing a task, not just the unit price of API Tokens. The above suggestions and data are all paraphrases of the provided media text.

Why it matters to readers

The test results in the text suggest that the choice of model and reasoning intensity is worth comparing alongside task completion cost; however, whether these results can be applied to other tasks cannot be confirmed from the materials.

You can first understand the process of "execute first, check results, and add more if it fails," and do not interpret the 41% savings mentioned in the text as a fixed saving that all tasks can obtain.

Risks and unknowns

The original link to the Anthropic guide was not provided, so the guide, model names, and evaluation data cannot be independently confirmed.
The materials do not provide the complete evaluation method, subset sizes, or cost calculation details.
It remains unclear whether the test results can be replicated in real-world tasks and whether they are applicable to other tasks.

BlockBeats reports that Anthropic has released a guide on Claude model selection and cost optimization

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center