Sonnet 5.5 deserves a task-level comparison: effort, repeated context and correction time can matter more than a headline speed improvement. Its cache price changed after launch.
Read the launch and the subsequent price change
The September 28 announcement positions Sonnet 5.5 for bounded everyday work, coding fixes and documents. Standard input/output prices are $2/$10 per million tokens. Launch cache reads were $0.20, but Anthropic lowered them to $0.10 on October 7. Its app default effort is Medium, while the platform default is High. Provider benchmark claims should remain attributed to their evaluation setup.
Effort is part of the experiment
Our suggested comparison begins with a small repository bug that has a clear failing behavior and a meaningful check. Run the same description at two effort settings, without changing the available files or tools. Save the resulting diff and test output, including failures. Judge whether the fix addresses the reported behavior and introduces unnecessary changes. A longer reasoning run is worthwhile only when its additional work changes the quality of the result.
Separate token prices from completed-task cost
For an uncached example of 100,000 input and 20,000 output tokens, the standard-rate arithmetic is $0.40. That is not a claim about how many tokens a real task needs. Our recommendation is to track repeated context, output, repair attempts and tool charges in separate columns. Caching is only a saving when eligible input is actually reused; do not price every prompt as a cache hit in a forecast.
A polished document still needs substantive review
For a policy memo, give the model a short evidence packet and ask it to identify unsupported statements before improving the prose. Review quotations, dates, totals and recommendations separately from typography. Our acceptance criterion is an accurate, usable artifact with a visible evidence trail. A spreadsheet that looks professional but changes a denominator is a failure. Keep the source packet beside the final output so a colleague can audit the transformation.
Where to draw the escalation line
We recommend escalating tasks with unresolved requirements, conflicting evidence or expensive consequences rather than choosing a model entirely by perceived confidence. For coding, require tests tied to the behavior; for documents, require an accountable reviewer. No Sonnet benchmark was run for this article. The GPT-6.1 Sol cost guide offers a separate arithmetic worksheet, but matching token prices does not establish equal task quality or interchangeable safety behavior. Keep the original failing example in the acceptance packet so the reviewer can see the exact behavior the patch must change.
Frequently asked questions
Which cache-read rate should this article use?
The October 7 update lists $0.10 per million tokens; the launch page contains the earlier $0.20 rate.
Does higher effort always give a better result?
Measure the same task at both settings. Extra time is useful only when it improves the outcome enough to justify its cost.
What is the important coding metric?
Track correct, reviewed fixes with relevant checks, including retries and review time, rather than lines of code produced.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Introducing Claude Sonnet 5.5 ↗Anthropic
- 2
Help us keep this useful. Check contact availability →
