Original URL: https://getagenttools.com/blog/claude-sonnet-5-5/

# Claude Sonnet 5.5: the same token prices, a different migration checklist

28 Sep 2026Marek.

- claude
- agent-ops
- coding-agents

Claude Sonnet 5.5 launched on September 28, 2026, with the same base API token prices as Sonnet 5. Anthropic advertises faster work and lower task costs. For operators, the release also brings an integration question: will your agent still call tools and report progress after the model changes?

 

- **Release:** September 28, confirmed in Anthropic's model documentation.
- **Price:** $2 per million input tokens and $10 per million output tokens on the Claude API.
- **Speed:** Anthropic advertises a 30% improvement over Sonnet 5. This is a vendor claim, not a GAT measurement.
- **Migration:** existing Sonnet 5 API integrations face five documented breaking changes.
- **Visibility:** some progress text moves into thinking blocks, which can leave an interface quiet without an API error.

 

If you use [Claude Code](https://getagenttools.com/tools/claude-code/), keep the model and its surrounding tool runner separate in your evaluation. The API details below concern applications built on the Messages API. This is a launch analysis based on documentation, not a hands-on verdict.

  

## What does Claude Sonnet 5.5 cost?

 

Sonnet 5.5 keeps Sonnet 5's standard Claude API rates. Input costs $2 per million tokens; output costs $10. Cache reads cost $0.20 per million tokens. A five-minute cache write costs $2.50, and a one-hour write costs $4. These are usage prices, separate from a Claude subscription.

 

| Claude API rate, per million tokens | Sonnet 5 | Sonnet 5.5 |
| --- | --- | --- |
| Input | $2 | $2 |
| Output | $10 | $10 |
| Cache read | $0.20 | $0.20 |

 

Anthropic's [pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing) lists both models at those rates. Check your provider's billing terms if you use a cloud partner.

 

The [Anthropic homepage](https://website.anthropic.com/) advertises up to 30% lower costs for most work and 30% faster operation. That does not mean a 30% cut to the token rate. Treat the headline as a claim about completing work, then measure your own workload.

 

My suggested comparison is cost per accepted result. Include retries and human corrections. A cheap first response loses its advantage if somebody has to repair the output before using it.

 

## Where is Sonnet 5.5 available?

 

Anthropic's [Sonnet 5.5 model page](https://platform.claude.com/docs/en/models/sonnet-5-5/overview) lists the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The Claude API identifier is `claude-sonnet-5-5`. Bedrock uses `anthropic.claude-sonnet-5-5`.

 

The model page specifies a one-million-token context window and a standard maximum output of 128,000 tokens. Platform support does not prove that a particular account has access. Check the account and endpoint you will actually use before scheduling a migration.

 

## Which API settings change from Sonnet 5?

 

Anthropic documents five breaking changes in [what's new in Sonnet 5.5](https://platform.claude.com/docs/en/models/sonnet-5-5/whats-new-sonnet-5-5). These deserve a check before changing a production model ID:

 

| Existing behavior | What to check |
| --- | --- |
| Thinking disabled | `thinking.type: disabled` now fails. `between_tools` turns off up-front thinking at `high` effort or below. |
| Forced tool calls | `tool_choice` types `any` and `tool` now fail. Use `auto` and validate the resulting behavior. |
| Edited conversation history | Replayed thinking blocks can fail validation after earlier history changes. Default enforcement depends on account age and platform. |
| Older computer-use tool | Claude API and Google Cloud require `computer_toolset_20260801`; Bedrock still accepts `computer_20251124`. |
| Advisor model pairing | Sonnet 5, Opus 4.7, and Opus 4.8 are rejected as advisors for a Sonnet 5.5 executor. |

 

Those are API integration changes. They do not mean every person selecting Sonnet in a chat interface must edit a request body.

 

For forced tools, the [migration guide](https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide) adds an important qualification: strict tool schemas do not force a call. The model can still answer in text. On Bedrock, strict tool use is unavailable for this model; validate inputs in your application.

 

My first regression case would be a task that must use a tool to finish. Check that the call happened. Valid JSON alone cannot prove that.

 

## Why could an agent stop showing progress?

 

Longer notes between tool calls can arrive as `thinking` blocks instead of ordinary text. With adaptive thinking's default `display: "omitted"`, their text is empty. Brief remarks can remain text. The request can succeed while the interface stops showing the updates its user expects.

 

The migration guide explains how to request readable summaries with `thinking.display` when using adaptive thinking. With `between_tools`, progress summaries return without that setting. Parse response blocks by type and preserve thinking blocks in tool loops, including empty ones.

 

This is the failure I would put beside the usual error-rate check. A screen that only shows final output cannot tell you whether the work was observable along the way. Our [agent monitoring article](https://getagenttools.com/blog/ai-agent-monitoring/) covers that distinction between a successful request and a useful trace.

 

## How would I evaluate the upgrade?

 

I would replay a small set of completed tasks with known acceptance criteria. Keep the inputs, tools, and success checks fixed. Record elapsed time, billed usage, tool calls, corrections, and whether progress remains visible.

 

Test a failed tool call as well as a successful one. Include a resumed conversation if your application supports it. Keep a copy of the old configuration so a failed comparison does not become an emergency migration.

 

That is a proposed evaluation, not a report of tests already run. GAT has not established a Sonnet 5.5 performance verdict. The launch gives a reason to run the comparison; the result should decide the switch.

 

For an existing automation, start with the request and response contract. Then measure whether Sonnet 5.5 finishes your actual work faster and at lower cost.

Put this into practice

[See the tool list](https://getagenttools.com/tools/)[Train your team](https://getagenttools.com/#training)