AI
Anthropic's Claude Sonnet 5.5 Delivers Opus-level performance at half the cost
Sonnet 5.5 runs 30% faster and cheaper than its predecessor while matching most benchmarks from the pricier Opus model.
The New Stack
Anthropic unveiled Claude Sonnet 5.5 this week as the second member of its new 5.5 model family, following Opus 5.5's launch the prior week. The update brings meaningful speed and cost improvements to Anthropic's midrange offering, positioning it as a practical alternative to the flagship Opus model for many real-world workflows.
Speed and cost gains#
Sonnet 5.5 processes requests over 30% faster than Sonnet 5 while reducing per-task costs by up to 30%. Pricing remains steady at $2 per million input tokens and $10 per million output tokens—matching OpenAI's GPT-6 Sol—yet token efficiency gains mean actual operational expenses should fall. The speed improvement makes this Anthropic's fastest Sonnet variant to date.
The company points to three core strengths for the updated model:
- Well-defined everyday tasks requiring straightforward reasoning
- Bug fixing and code refinement
- Polished document, slide, and spreadsheet creation
Early reviewers highlighted its ability to enhance user interface polish and generate presentation decks that need minimal human revision. Like Opus 5.5, Sonnet 5.5 also exhibits clearer, more natural-sounding prose compared to prior-generation models.
Benchmark performance approaches Opus#
On several capability tests, Sonnet 5.5 comes remarkably close to—or occasionally exceeds—Opus 5.5 despite its lower price point. The biggest leap appears in coding tasks: on Terminal-Bench 4.0, it scores 70.6%, far ahead of Sonnet 5's 10.3% and edging past Opus 5.5's 66.4%. (The Sonnet 5 result was dampened by timeout and token-limit issues.) On CursorBench, drawn from real Cursor sessions, the gap narrows to about two points.
For knowledge work, Sonnet 5.5 achieves 1,844 on Artificial Analysis' GDPval-AA benchmark—just two points shy of Opus 5.5 and roughly 400 points ahead of Sonnet 5. It also outperforms OpenAI's GPT-6 Sol (1,487). In informal tests involving long-horizon reasoning and image interpretation, Sonnet 5.5 became the first Sonnet model to solve Pokémon Red using only screenshots.
Anthropic notes that at lower or medium effort settings, Sonnet 5.5 can match Sonnet 5's peak scores on several benchmarks while costing roughly a tenth as much per task. However, Opus 5.5 still maintains an edge for "complex, open-ended work requiring sustained judgment."
