Sonnet 5: Opus quality at «half the price»… with two hidden catches
Anthropic launches Sonnet 5, a new version of its mid-tier family, promising to close the gap with Opus while keeping Sonnet's price: performance close to Opus 4.8 on agentic benchmarks, 1 million tokens of context and the new default model in Claude Code for Pro users.
But watch out for the «Opus quality for 60% less» headlines. First catch: the launch price ($2 input / $10 output per million tokens) is subsidized only until August 31st; after that, $3 and $15. Many comparisons out there are being made against the discounted price.
Second catch: the new tokenizer consumes 40% more tokens in English and 33% more in Spanish, according to Simon Willison. And your codebase is in English even if you prompt in Spanish, so the price difference is not that big anymore. They pee on us and tell us it's raining.
The café's conclusion: there is no magic wand. Run evals on your own codebase and compare price per task, not per token. Meanwhile, OpenAI answers with GPT Terra, Luna and Sol in preview. Watch in the video.

All the news we covered