source: simon willison: what's new in claude sonnet 5
level: technical
anthropic released claude sonnet 5 today, positioning it as a model with performance close to opus 4.8 but at lower prices. the system card notes that sonnet 5 is significantly less capable at cyber tasks than mythos 5, which helped avoid us government release restrictions. the model has a 1 million token context window and 128,000 maximum output tokens, matching the tool and platform features of sonnet 4.6.
a key api change is the removal of sampling parameters temperature, top_p, and top_k. adaptive thinking is on by default unless explicitly disabled. pricing remains $3 per million input tokens and $15 per million output tokens, with an introductory discount to $2 and $10 until august 31st. however, a new tokenizer produces approximately 30% more tokens for the same english text compared to sonnet 4.6, effectively raising costs.
tests with the claude token counter tool show the new tokenizer increases token counts by about 1.4x for english, 1.33x for spanish, 1.28x for python code, and nearly no change for simplified mandarin. this means the real cost per english input is roughly 30% higher than the listed price. the model's image understanding remains unremarkable, with sonnet 5 mistaking a pelican for a goose.
why it matters: developers using claude sonnet 5 for english-language tasks will face a hidden 30% cost increase due to the new tokenizer, requiring budget adjustments and careful evaluation of the model's price-performance trade-off.