Model cards

claude-sonnet-5.5

Vendor anthropic Tier standard Price per million tokens, in / out $2 / $10 Released 2026-09-28 Context 1,000,000 Max output 128,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Everyday well-scoped work such as bug fixing and polished documents, slides and spreadsheets; over 30 percent faster than Sonnet 5.

  Vendor's own site anthropic.com

• Close to Opus 5.5 on most benchmarks and ahead on Terminal-Bench 4.0 (70.6 versus 66.4 percent).

  Vendor's own site anthropic.com

• Clear writing and good design sense for interfaces and slide decks, at $2 and $10 per million tokens.

  Vendor's own site anthropic.com

• Third on Vals AI's Excel modeling benchmark at 75.7 percent, at $8.79 per test versus $11.02 for Opus 5.5.

  Independent evaluator vals.ai

• Design and Go engine code close to Opus 5.5 quality at 30 to 51 percent lower cost per task.

  Independent evaluator thomas-wiegold.com

• Zero regressions across 100 runs on six replayed production monorepo changes, with Opus 5.5 regressing at higher effort.

  Practitioner report paddo.dev

• Raised bug coverage over Sonnet 5 at similar precision, about 60 percent cheaper per review.

  Independent evaluator coderabbit.ai

• Legal testers rated it best on extraction and document analysis at the lowest cost and above-average speed, statistically tied overall with pricier models (vendor-run test).

  Practitioner report litera.com

Watch out for

• Opus 5.5 remains clearly stronger on complex, open-ended work needing sustained judgment.

  Vendor's own site anthropic.com

• Scores lower at Max effort than at Xhigh; in two cases Cognition examined, it ran Claude Code's review skill across many subagents, causing a timeout or out-of-scope edits.

  Vendor's own site anthropic.com

• Filled in test expectations from memory and got some wrong; suited to human-reviewed work, not unattended loops.

  Independent evaluator thomas-wiegold.com

• Between-tool progress text arrives empty by default, so streaming agents appear to go quiet.

  Practitioner report brinvik.com

• Not an Opus 5.5 replacement for high-risk review; 6/13 hard bugs vs Opus 5.5 Standard's 8/13.

  Independent evaluator coderabbit.ai

• Trailed on general legal knowledge and drafting and cited source documents less often (vendor-run legal test).

  Practitioner report litera.com

• Migration pitfalls: thinking blocks must be preserved in tool loops, and parsers or UIs that only read text can break or look frozen; forced tool choice and legacy parameters may fail.

  Practitioner report pub.towardsai.net

• Artificial Analysis lists very high time to first answer token (about 430 seconds at max) and about 193K output tokens per index task, so max effort can cost more than Opus 5.5.

  Independent evaluator deeplearning.ai

Benchmarks

Coding percentile 0.91

Agentic coding percentile 0.62

Knowledge percentile 0.63

Overall percentile 0.76

Vision percentile 0.33

Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.

Practitioner reports

• -

  As the main model, Sonnet 5.5 delegated a lookup to an Explore subagent that stopped at a re-export; Haiku 5.5 followed it to the definition.

  Compared with claude-haiku-5.5 | 2026-10-09

  Practitioner report developersdigest.tech
• +

  Legal AI vendor's former-lawyer team found it the highest overall result they had tested, strongest on extraction and document analysis, weaker on drafting and general legal knowledge (vendor-run, sponsored).

  Compared with claude-opus-5.5 | 2026-10-06

  Practitioner report litera.com
• +/-

  In a Go chess move-generator test it matched Opus 5.5 on an independent perft check and ran slightly faster, but its own tests had two failures (counts filled from memory) and it rejected only 5 of 7 invalid FENs versus all 7 for Opus 5.5.

  Task: feature | Compared with claude-opus-5.5 | 2026-09-30

  Practitioner report thomas-wiegold.com
• +/-

  Reviewer recommends it for well-scoped everyday work where a human reviews output, and Opus 5.5 for unattended agent loops; it cost 30-51% less per task than Opus in his tests but 25% more than Sonnet 5 on a design task (over twice the output tokens).

  Task: agentic-long-running | Compared with claude-opus-5.5 | 2026-09-30

  Practitioner report thomas-wiegold.com
• +/-

  Sonnet 5.5 was a close second to Opus on design and equal on Go chess correctness, but invented some test values and applied illegal moves.

  Compared with claude-opus-5.5 | 2026-09-30

  Practitioner report thomas-wiegold.com
• +/-

  In seven side-by-side tasks at high effort, Sonnet won 4 and Opus 3: Opus was better on a vague 90-day growth plan and motion-design work, Sonnet was close or better on landing page, explainer and dashboard builds at roughly half the cost.

  Task: writing-content | Compared with claude-opus-5.5 | 2026-09-29

  Practitioner report aifire.co
• -

  HN commenters (quoted in a review) call it weaker than Opus on open-ended thinking with fewer error/consistency checks, and say it is priced too close to Opus; a Reddit user calls it a solid subagent for Opus.

  Task: research | Compared with claude-opus-5.5 | 2026-09-29

  Practitioner report eesel.ai
• -

  An adversarial esoteric-language benchmark by one developer scored it 7.4% versus 17.8% for Sonnet 5, a regression on that niche task; the same review notes forced tool_choice and disabled thinking now return 400 errors, breaking migrations.

  Task: feature | Compared with claude-sonnet-5 | 2026-09-29

  Practitioner report eesel.ai
• +

  Sonnet 5.5 never broke a passing test at any effort level, with 10 to 13 clean runs; Max gained one edge case at higher cost than Opus.

  Compared with gpt-6-sol | 2026-09-29

  Practitioner report paddo.dev
• -

  After switching a CRM sales assistant to Sonnet 5.5, users saw only a spinner between tool calls until thinking display was changed.

  Compared with claude-sonnet-5 | 2026-09-29

  Practitioner report brinvik.com
• -

  Reported that Max effort was costly and slow, with one SVG test taking about 16 minutes and failing, plus an adversarial benchmark where it asked questions instead of continuing.

  Compared with claude-sonnet-5 | 2026-09-29

  Practitioner report eesel.ai
• +/-

  In a one-run-per-scenario test, Sonnet 5.5 beat Sonnet 5 on a usage-tracker bug hunt (7 vs 3 quality, cheaper and faster) but lost a high-effort refactor sketch (6 vs 7) and a routes audit. Heuristic scoring, unpublished prompts, and Kunya sells model access.

  Task: bug-fix | Compared with claude-sonnet-5 | 2026-09-29

  Practitioner report kunya.ai
• +

  In an update to a 20-runs-per-model replay of six real monorepo changes, Sonnet 5.5 had zero regressions across 100 runs, and medium effort was best for the Opus/Sol pair. One codebase, 3-4 runs per change, prompts written by someone who knew the fixes.

  Task: bug-fix | Compared with claude-opus-5.5 | 2026-09-29

  Practitioner report paddo.dev
• +/-

  Simon Willison's launch-day SVG test: low through high effort finished in 10-17 seconds, but max ran nearly 16 minutes, cost $1.28 and failed after using all 128K thinking tokens (as tabulated by a review).

  Task: feature | 2026-09-28

  Practitioner report eesel.ai
• +/-

  CodeRabbit measured Sonnet 5.5 catching 6 of 13 hard known bugs at 41 percent precision (Sonnet 5: 4 of 13), in about half the time and 60 percent lower cost, but with different misses and below Opus 5.5 (8 of 13). Thirteen cases is small; vendor sells code review.

  Task: code-review | Compared with claude-opus-5.5 | 2026-09-28

  Practitioner report coderabbit.ai
• -

  At max effort the pelican-on-a-bicycle SVG prompt burned 128K tokens ($1.28) with no output; at xhigh it produced a good SVG in 41 seconds for under 6 cents.

  Compared with claude-opus-5.5 | 2026-09-28

  Practitioner report simonwillison.net