CONFIRMED: Sonnet got faster, not simply cheaper
Anthropic launched Claude Sonnet 5.5 on September 28 as an update to its mid-tier model for everyday work. The central promise is simple: more than 30% faster output than Sonnet 5, at the same price of $2 per million input tokens and $10 per million output tokens. Cached input remains $0.20 per million tokens. The real savings, however, depend on how many tokens and tool calls each task requires.
The surprise is coding and cost per task
Sonnet 5.5 is positioned for bug fixing, documents, spreadsheets, presentations, and well-scoped tasks. On Terminal-Bench 4.0, Anthropic reported a 70.6% score, above Opus 5.5 at 66.4%, although that figure comes from the company itself. An independent Artificial Analysis run also put Sonnet ahead in its test, while adding an important caveat: at maximum effort, it used about 193,000 tokens per task, more than Opus 5.5. That can reverse the economics in workflows that let the model think for too long.
The technical contract changed with the model
The official documentation contains changes that matter more to agent maintainers than to people who only chat with the model. Adaptive thinking is on by default and effort controls depth. To turn off upfront thinking in tool workflows, applications must use between_tools because disabled returns an error. Forced tool use no longer works, thinking blocks are tied to the model and conversation, and text between calls may arrive inside thinking blocks. On the Claude API and Google Cloud, the older computer use toolset is also not accepted.
What it means for small Brazilian businesses
For a small or midsize company using an agent for coding, support, or document analysis, the release is attractive because it may deliver more iterations per minute without a list-price increase. The benefit is clearest on defined tasks, such as fixing a bug, reviewing a contract, or turning data into a report. But changing only the model ID is an incident waiting to happen. Streaming, thinking history, tool choice, computer use, and effort limits should be tested in staging, with cost per completed task measured in reais, not only tokens.
The editorial read
Sonnet 5.5 matters because it turns Anthropic’s mid-tier model into a serious candidate to replace Opus on routine work. It is not a blank check for automation: the benchmark is strong, maximum effort is expensive, and the migration includes real breaking changes. My view is to test it now on tightly scoped workflows, with controlled effort and a rollback ready. The promise is not a magical model. It is a faster agent when the surrounding system knows how to keep the task bounded.
Sources
Anthropic, official what's new and migration documentation
Decrypt, independent coverage and analysis
SiliconANGLE, launch coverage