A model upgrade can make an agent interface appear frozen even though the request is succeeding. Claude Sonnet 5.5 returns some text between tool calls inside thinking blocks rather than ordinary display text. The default can omit that content from the user interface, so an application that previously streamed progress may go quiet until the next tool call or final answer.[2][4]
This is a response-shape migration, not a model failure. It illustrates why model upgrades should be tested as protocol changes: the request can return HTTP 200, the tool can run and the final artifact can be correct while the product experience, logs or approval flow silently break.
Sonnet 5.5 specifications and migration changes
| Measure | Value | Context |
|---|---|---|
| Context window | 1 million tokens | Published model limit.Source 2 |
| Maximum output | 128,000 tokens | Published model limit.Source 2 |
| Standard price | $2 input / $10 output per million tokens | Cache writes and reads are priced separately.Source 2Source 5 |
| Provider-reported speed | 30%+ faster than Sonnet 5 | Anthropic also reports up to 30% lower cost per task for most work.Source 1 |
| Default reasoning | Adaptive thinking; high effort on the API | Claude apps and Claude Code use medium effort by default.Source 1Source 2 |
| Silent response-shape change | Text between tool calls may arrive in thinking blocks | An application must deliberately display it or use the documented between-tools configuration.Source 2Source 4 |
| Breaking tool changes | Forced tool choice can return an error | The migration guide also changes computer-use and advisor-tool compatibility.Source 2Source 4 |
The interface can lose progress without losing the result
Tool-using applications often expose short text updates such as “I found the file” or “I am running the test.” Sonnet 5.5 can place text between tools inside a signed thinking block. If the block uses an omitted display mode, the backend receives the reasoning structure while the front end shows nothing. A timeout indicator or an impatient user may then trigger an unnecessary retry.[2][3]
The repair is not to expose every hidden thought. The application should define a stable progress channel: tool started, tool completed, approval requested, artifact saved and task finished. User-facing status can be generated from structured events while private reasoning remains private.
A migration test matrix for a tool-using Sonnet application
| Test | Expected evidence | Possible failure |
|---|---|---|
| Streaming between tools | Visible progress event before and after each tool call | Interface appears stalled while execution continuesSource 2Source 4 |
| Forced tool configuration | Documented supported selection mode and handled 400 errors | A previously valid any-or-specific tool choice now rejects the requestSource 2Source 4 |
| Thinking continuity | Append-only history and supported thinking-block compatibility | Edited history can cause a 400 error; unsupported model blocks can be dropped while the request succeeds.Source 4 |
| Computer use | Current tool identifier for the selected platform | An obsolete computer tool ID is rejected on Claude API or Google CloudSource 2 |
| Long output | Parser, storage and UI tested beyond former limits | The model can return more than the application acceptsSource 2 |
| Cost and effort | Token, cache, latency and acceptance metrics by effort level | A faster model uses an inappropriate high-effort default for routine tasksSource 1Source 5 |
Signed thinking creates a continuity boundary
Anthropic distinguishes model compatibility from history integrity. Sonnet 5.5 can read thinking blocks from some earlier models but drops unsupported blocks while still returning a successful response. On accounts created from 31 August 2026, conversation binding is enforced by default: changing earlier history before replaying a signed block can produce a 400 error. Blocks also remain restricted to the originating account or a linked account.[4]
A durable audit record should preserve the visible status events, tool inputs, tool outputs and final artifact even when private reasoning cannot be carried forward. Operational traceability does not require publishing an internal chain of thought. It requires evidence of what action occurred and what result it produced.
A benchmark gain does not certify an integration
Anthropic reports major gains on terminal coding, knowledge work, visual understanding and computer use, together with lower token consumption on many tasks. Those results justify evaluation. They do not prove that an existing agent harness, parser, approval screen or retry policy survives the model change.[1]
- Pin the old and new model on the same representative tasks and tool schemas.
- Capture the complete response object, not only the final text field.
- Test ordinary success, refusals, tool errors, timeouts, retries and approval pauses.
- Measure accepted-task cost and correction time at low, medium and high effort.
- Confirm that status text, thinking blocks and final output are rendered intentionally.
- Roll back automatically when protocol or acceptance tests fail.
Sonnet 5.5 may be a materially better model while still being a breaking software dependency. The high-standard upgrade treats pricing, capability and response semantics as one release. When the user can see task state, approvals and artifacts independently of private reasoning, a response-shape change becomes a manageable migration rather than a mysterious agent failure.
Sources & further reading
- Introducing Claude Sonnet 5.5 Anthropic · 28 Sep 2026 · Accessed 11 Oct 2026
- Claude Sonnet 5.5 model overview Claude Platform Docs · Accessed 11 Oct 2026
- What’s new in Claude Sonnet 5.5 Claude Platform Docs · Accessed 11 Oct 2026
- Migrating to Claude Sonnet 5.5 Claude Platform Docs · Accessed 11 Oct 2026
- Claude Platform pricing Claude Platform Docs · Accessed 11 Oct 2026