Claude Sonnet 5.5 vs Claude Sonnet 5

What each one accepts

Token prices

Charged from a prepaid USD balance, no taxes added. One key runs both, through /v1/chat/completions in the OpenAI format. A failed generation is $0.

New projects should start on Sonnet 5.5: it keeps the window and output limit of Sonnet 5 and costs less per token. The catch is two request patterns that now fail. If your code turns thinking off or forces a specific tool, Sonnet 5.5 answers with a 400, so stay on Sonnet 5 until you have rewritten those parts.

When to pick Claude Sonnet 5.5

• You start a new agent or service and can design it around tool_choice: auto.
• You want a lower per-token rate for your default Claude Code model.
• Your inputs are long documents or whole repositories that fit in 1M tokens.

When to pick Claude Sonnet 5

• Production code sends thinking as disabled and you do not want to touch it this week.
• A pipeline forces one tool on every call, for example to pull data into a fixed format.
• Your effort values were tuned on Sonnet 5 and have not been re-tested on 5.5.

Differences you will notice

• Rates. Claude Sonnet 5.5 is billed at $1.89 / $9.45 per 1M tokens, Claude Sonnet 5 at $3.24 / $16.20 per 1M tokens. The newer one is cheaper on input and on output.
• Turning thinking off. Sonnet 5 accepts thinking: {"type": "disabled"}. Sonnet 5.5 rejects it. Lower effort, or on /v1/messages send thinking: {"type": "between_tools"} at high effort or lower.
• Forced tools. A tool_choice of type any or a named tool fails on Sonnet 5.5. Use auto and say in the prompt which tool you expect.
• Effort scale. Both default to high, but Anthropic recalibrated the levels in 5.5, so one setting can lead to a different amount of reasoning.
• Unchanged. The 1,000,000-token window, 128K output, both endpoints, and the 400 on non-default temperature, top_p and top_k.

How to decide

The token rates favor 5.5, but thinking is billed as output, so effort shapes the real cost more than the rate does. I start chat and simple jobs at low and agents or code at medium, then check usage. If your code has neither disabled thinking nor forced tools, switching is just the model id. Current rates for every Claude model are on the pricing page.

Questions

What is the smallest change to move a request to 5.5?

Delete thinking: {"type": "disabled"} and any forced tool_choice, set a lower effort and tool_choice: auto, and change the model id. Everything else in the request can stay.

Does 5.5 give better answers than Sonnet 5?

It is the next version in the same line, but I have no head-to-head tests of my own yet. Run a sample of your real requests on both before you move production.

Can Claude Code use either model?

Yes. Make one the default and swap with /model mid-session. The setup is in the Claude Code docs.