Anthropic’s Sonnet 5.5 arrives in a market that increasingly sells time to think. Developers can let a model work longer, recruit subagents and revisit its own output. The intuitive expectation is that more effort should buy better work. One detail in this release shows why that relationship can break.
Launched September 28, Sonnet 5.5 is positioned as the quicker, lower-cost partner to Opus 5.5 for well-defined work. The revealing result is not simply that it performs better than Sonnet 5. On one coding evaluation, its highest effort setting performs worse than the setting immediately below it. The extra work sometimes changes the job.
The release changes the amount of work per task
Anthropic says the model generates output more than 30% faster and can complete typical work at up to 30% lower cost than its predecessor. The company attributes the cost improvement to fewer tokens, despite unchanged base input and output prices at launch. Those are vendor measurements, not a forecast for every application.
This distinction matters because an agent’s bill contains its chosen process. It can inspect files repeatedly, launch reviews or retry commands. A lower price per token does not address unnecessary work; fewer necessary steps can. Evaluating a replacement model therefore requires measuring the whole task, including whether the final result still meets the same standard.
The highest setting can cross the task boundary
The FrontierCode footnote explains that Max effort trails Xhigh. The evaluation asks whether code changes could be merged without human edits and penalises out-of-scope work. In two cases examined by Cognition, additional subagent review contributed to a timeout or extra edits outside the requested task.
The observation is narrow but useful. It does not establish that review is generally harmful or that Max effort is always worse. It shows that a model can increase activity in a way the actual objective does not reward. Improving adjacent code is still an error when the contract is to make a bounded change and stop.
An effort control is also a policy control
An effort setting can change more than the length of an internal calculation. It can change which tools are used, how much context is gathered and whether the agent decides to invite another reviewer. Comparing settings is consequently an experiment on the surrounding workflow, not just on reasoning depth.
A team could test this by keeping repository, request and acceptance criteria fixed, then logging newly introduced edits and tool calls across settings. The useful question is which additional actions improve the deliverable. Counting only successful checks would miss the cost of investigating unrelated problems or turning a small repair into a redesign. Acceptance criteria should include what the agent is meant to leave alone, so extra changes are observable departures from the task rather than automatically counted as useful initiative.
Choose the work envelope before choosing the budget
The implication for everyday deployment is to define what completion means before increasing effort. A narrow bug fix, an exploratory audit and an architecture redesign justify different stopping rules. The model should receive a budget appropriate to that envelope, with escalation when evidence warrants a broader investigation.
Sonnet 5.5’s release also needs current pricing context: Anthropic subsequently cut its cache-read price with the October 7 Haiku launch. Historical launch comparisons should not be mistaken for today’s complete bill. The more durable lesson is operational. A stronger agent is valuable when it knows which work to finish, not merely when it can find more work to do. That is a measurable quality of the model-and-workflow combination, and an increasingly important one as additional effort becomes easier to purchase.
Anthropic official release and methodology
Questions, news tips or corrections? Contact [email protected].
