claude-sonnet-5
Meant to excel at
-
Reviewers found it strong for writing and building code at its tier, working well in long agent loops and writing tests first.
Practitioner report coderabbit.ai -
On routine diffs one reviewer could not tell its output from Opus 4.8, and it was much faster.
Practitioner report claudeai.guide
Watch out for
-
Over-builds small requests, rewrites its plan mid-task and is slower than Sonnet 4.6.
Practitioner report coderabbit.ai -
Drifts off-plan on multi-hour autonomous sessions and repo-wide refactors, where Opus held up better in one reviewer's paired tests.
Practitioner report claudeai.guide
Benchmarks
Agentic coding percentile 0.28
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 63.2 | 0.58 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-06-30 |
| terminal-bench 2.1 | 80.4 | 0.27 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-06-30 |
| terminal-bench 4.0 | 10.3 | 0.00 | Vendor's own site anthropic.com | 2026-09-28 |
Computer use percentile 1.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| osworld-verified | 81.2 | 1.00 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-06-30 |
Knowledge percentile 0.77
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| humanity's last exam | 57.4 | 0.77 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-06-30 |
Overall percentile 0.67
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gdpval-aa v2 | 1618 | 0.67 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-06-30 |
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
- +
On one browser-tool research prompt at extra effort, Sonnet 5 planned, stayed on the API-provider angle, got current pricing right and separated confirmed from unverified facts without babysitting. Verdict rests on a single task and the site earns affiliate fees.
Task: research | 2026-07-10 older than 90 days
Practitioner report unite.ai - +/-
After one to two weeks as default model in Claude Code, the reviewer could not distinguish Sonnet 5 from Opus 4.8 on everyday TypeScript and Python feature and debugging diffs, and it was much faster; Opus still won on multi-hour sessions and repo-wide refactors. Subjective, no benchmarks.
Task: feature | 2026-07-06 older than 90 days
Practitioner report claudeai.guide - +/-
CodeRabbit found Sonnet 5 posted cleaner, more precise review comments than Sonnet 4.6 (about 38-40 percent precision vs 29) but caught fewer known bugs, and high effort roughly doubled cost for little gain. Benchmark size was not disclosed, and the publisher sells AI code review.
Task: code-review | 2026-06-30 older than 90 days
Practitioner report coderabbit.ai