Anthropic’s new model excels at long-term reasoning for simple to moderately complex problems.
This is Anthropic’s new driving model. Announced this Tuesday, June 30, Claude Sonnet 5 replaces Sonnet 4.6. More autonomous, more agentic, less expensive than its predecessor… Scalded by the catastrophic episode of Fable 5 with the White House, Anthropic has worked on the model and its communication to prove that its latest addition is perfectly harmless. Does this make it a cutting-edge model to favor in everyday life? Answer.
A more autonomous model
Claude Sonnet 5 will completely replace Sonnet 4.6 across the entire Anthropic stack. In Claude Code, Cowork or Claude.ai, Anthropic places it as the default model. And for good reason, Sonnet 5 is a very general model capable of solving any simple to moderately complex task. However, Sonnet 5 does not completely replace Opus 4.8. On SWE-bench Pro (code in agentic mode), it obtains 63.2%, compared to 69.2% for Opus. The gap narrows significantly on Terminal-Bench 2.1 (autonomy in a terminal), with 80.4% against 82.7%, as well as on the autonomous use of a computer, where Sonnet reaches 81.2% on OSWorld-Verified, against 83.4% for its big brother.
Opus therefore maintains the advantage on the most difficult problems, but Sonnet now comes close enough to cover the vast majority of daily tasks. It even stands out on certain reasoning works in long context with a few drawbacks (read below).
It is especially on agentic uses that Sonnet 5 takes on its full meaning. The model is capable of chaining actions with greater autonomy (much less human feedback) than Sonnet 4.6. He thus becomes an excellent model in Claude Code. The model is fast (although limited to 79 tokens per second depending on Artificial Analysis) and reliable enough to fix bugs. The model further reinforces the relevance of the Opus Plan mode. (/opus plan mode): Opus 4.8 can be used to analyze a complex problem, design the architecture and produce a detailed plan, before entrusting its execution to Sonnet 5, in order to have the best of both worlds.
Usual prices for Anthropic
Sonnet 5 is offered at a promotional rate of $2 per million tokens in and $10 out until August 31, 2026. It will then increase to $3 in and $15 out, the usual pricing level for Sonnet models at Anthropic. On paper, Sonnet 5 therefore remains significantly less expensive than Opus 4.8, even though it comes very close on several agentic benchmarks. In practice, it’s more complicated…
A model with a few drawbacks
If Sonnet 5 impresses on paper, it is not without its flaws. The main one concerns its new tokenizer, which allows it to consume up to 30% more tokens compared to Sonnet 4.6. Concretely, its window of one million tokens can accommodate less text, documents or code than that of its predecessor. This inflation also weighs on the bill, particularly during complex tasks carried out on very large contexts with a high level of reasoning. Under these conditions, the total cost of a request can sometimes exceed that of Claude Opus 4.8, despite a lower unit price.
Finally, Anthropic has voluntarily limited certain cybersecurity capabilities of Sonnet 5 in order to reduce risky uses. The model is therefore much less good at cyber tasks and can even perform worse than Sonnet 4.6. A compromise fully accepted by Anthropic.
Should you migrate to Claude Sonnet 5?
Should you migrate to Claude Sonnet 5? In the majority of cases, yes. Sonnet 5 stands out as the best choice for everyday uses, particularly in Claude Code, where its agentic capabilities, speed and autonomy make it an excellent execution model. Opus 4.8, however, remains essential whenever a task requires real depth of reasoning. Outside of code, Sonnet 5 remains excellent for knowledge work tasks, especially in Cowork.
For developers, several points will need to be checked before migrating. First of all, it will be necessary to submit the same requests to the token counting API with Sonnet 4.6 then Sonnet 5, including the entire context in order to measure the real increase in consumption. This not only allows you to forecast your costs but also recalculates the space occupied in the context window. The max_tokens limits also need to be re-evaluated, they cap the entire output, i.e. both the internal reasoning and the visible response. Next, Sonnet 5 enables adaptive reasoning by default, old fixed budgets defined with thinking: {type: “enabled”, budget_tokens: N} are no longer accepted. Finally, it is better to completely remove the temperature, top_p and top_k variables from the queries. Anthropic now recommends controlling tone and variability directly in the prompt system.
Anthropic is moving forward with extreme caution on Sonnet 5, and this is reflected in the model’s design itself. This is not a new frontier model but a more adaptive, more autonomous intelligence that is better sized for daily work. Innovation is above all incremental, all surrounded by numerous safeguards. Sonnet 5 thus becomes probably the most balanced model.
