Claude Sonnet 4.5 excels in code, mathematics, aging orchestration. The price, however, is always the same.
This is a new breakthrough in AI for the code. After the success of GPT-5 Codex introduced by Openai in mid-September, many users were starting to abandon Claude Code. Anthropic responds this September 30 by revealing an optimized model for code and agent development: Claude Sonnet 4.5. The American publisher takes the opportunity to present an agency SDK and a new development experience in Claude.
More independent, more precise
Claude Sonnet 4.5 is almost no longer an LLM but a real agent. He is able to code independently for more than 30 hours on complex tasks and requiring several stages. More autonomy but also more details. On the benchmark of reference in agency development SWE BENCH VERIFIE (real code problems), Claude SONNET 4.5 obtains a score of 77.2%, or 2.7 points more than GPT-5 CODEX. He thus obtains the best score, all models combined.
Enough logically, the model also stands out in the Terminal Coding Agentic (50.0% against 43.8% for GPT-5 and only 25.3% for Gemini). More original, it also performed very well in financial analysis, where it largely dominates its competitors (55.3% against 46.9% for GPT-5 and 29.4% for Gemini). Finally, overall, it appears more robust and balanced than Gemini 2.5 PRO and GPT-5 on all STEM benchmarks. Note also that Claude Sonnet 4.5 is much more efficient in benchmarks than Claude Opus 4.1, for a very lower price (we will come back below). Cursor speaks, moreover, in the communication of Anthropic, of a model in the state of the art to solve the problems of complex code in autonomy.
Finally, Anthropic says that its model is excellent in interface use and manipulation. He obtained a Sota score of 61.4% on OSWORLD, benchmark testing the capacity of the model to evolve independently on a computer. Thanks to a Chrome extension, users of the max formulas can also leave the model control their browser. Finally Claude Sonnet 4.5 is the most aligned and secure anthropic model, thanks to a long -standing focus of anthropic on these questions.
A rate in continuity
On the price side, Anthropic remains in its classic pricing policy. Claude Sonnet 4.5 will be offered in the API at the same price as the previous version of Sonnet. Or 3 dollars per million tokens in Input for less than 200,000 context tokens and $ 15 output (the million tokens). For contexts greater than 200,000 tokens, it will take $ 6 inputs and $ 22.50 output. The prompt coaching is always available and reduces the bill.
The difference with Opus 4.1 is notable. As a reminder, Claude Opus 4.1 is billed, any context combined, $ 15 per million tokens at the start and 75 dollars at the exit. The latter is no longer intended to be used and must be replaced, also advises Anthropic.
An SDK to build agents
Finally, in parallel, Anthropic announces the launch of Claude Agent SDK. A platform to develop autonomous agents based on the internal operation of Claude Code. The SDK contains orchestration models, tools, permission management modules and production ready production. It is certainly one of the best SDKs to take advantage of Claude’s strength for use cases around the agentics without a third -party software overlay.
Finally, Anthropic announces a new development experience directly integrated into Claude. Named, “Imagine With Claude”, the latter allows you to design applications from A to Z in a virtualized environment. The agent develops in real time before the eyes of the user, using a GUI and the interface control capacities of Claude Sonnet 4.5.
On paper the demonstration is quite promising, in reality the use cases are still quite vague. The experience is, for the time being, offered to Max subscribers.
