Claude Sonnet 5 Released: 1M Context Window but Token Consumption Up to 1.35x

On July 1, 2026, Anthropic officially launched the Claude Sonnet 5 model. This model is open to free users and Pro users by default, supports a 1M context window, delivers performance close to Opus 4.8, and has a lower listed price.

Core Specifications and Actual Performance

The context length of Claude Sonnet 5 reaches 1 million tokens, a specification that directly expands the usable range of long-document processing and multi-turn conversations. Official tests show that its scores on code generation, long-text summarization, and complex reasoning tasks are close to those of Opus 4.8, while the input price per million tokens is reduced by about 30%.

However, changes brought by the new tokenizer alter actual usage costs. After processing the same Chinese or English text through the new tokenizer, the token count can increase by up to 1.35 times. This means that the actual fees paid by users may exceed the listed price, with the excess reaching up to 35%.

Comparison with Similar Products

Compared with GPT-4o and Gemini 2.5 Pro, Claude Sonnet 5 leads in context length. GPT-4o currently has a context limit of 128k, while Gemini 2.5 Pro has 200k. Sonnet 5's 1M window allows processing an entire novel or complete codebase in one go without segmentation.

In terms of cost structure, Sonnet 5's listed price is lower than Opus 4.8, but due to token inflation, the actual cost per single long-text API call may be on par with or even slightly higher than Gemini 2.5 Pro. In performance, Sonnet 5 scores 8 percentage points higher than GPT-4o on multilingual long-text consistency tests, but its inference speed is about 15% slower than Gemini 2.5 Pro.

Recommendations for Developers and Enterprises

Before calling Claude Sonnet 5, developers should first test the actual token count of their typical texts using the new tokenizer, and then adjust their budget accordingly. For enterprise users dealing with large volumes of Chinese contracts or code, it is advisable to first test short-text scenarios to confirm whether the token inflation rate is within an acceptable range.

For scenarios requiring stable costs, it is recommended to retain API calls to Opus 4.8 as a backup, or add a token count pre-check step to API requests. Free users can directly experience the 1M window, but should note that billing beyond the free quota will be calculated based on actual token count.