Published · 5h agoProduct2 min read
Opus 5.5 returns the text between tool calls as empty thinking blocks
Claude's platform docs list four Opus 5.5 breaking changes that return errors and a fifth that changes the response shape without failing any request. A pipeline gated on whether the call succeeded ships the quiet one.
Not a builder's beat, but builders have a standing stake in it.See today for builders
What happened
- An application that streamed that text to users as progress updates goes quiet between tool calls until it sets a display value that returns the text.
- Early testers ran long coding tasks for hours with little oversight, according to claude.dev, which says the model's biggest gains over prior Opus versions are on multi-step work.
- The same guide says a long run sometimes stops to report instead of going on. The report is a summary that names the next step and stops there, or a list of choices that do not block the work.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint The SDK's limits are both quantitative. max_turns counts tool-use turns only; max_budget_usd caps spend against a threshold. Neither one can stop a run that is producing well formed output in a shape you did not expect.
- exposure A run split across parallel subagents puts many more changed files behind one final message, and claude.dev's own example prompt puts the burden back on the operator to check each subagent's evidence.
- decision Whoever owns design review has to write down the specific patterns to exclude, because claude.dev says a general instruction like "avoid a generic look" mostly swaps one default style for another.
A review step sits at the end of the loop, and the loop is long. The agent SDK repeats turns until Claude produces output with no tool calls, then closes the session with a ResultMessage carrying the final text, token usage, cost and session ID [9][11]. Inside a turn it runs the requested tools and feeds each result back to the model, and control never passes to your code [8]. A reviewer reads the last message.
Four of the five documented changes to code already running on Opus 5 are listed as breaking. Thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and on the Claude API and Google Cloud the earlier computer use tool is not accepted [2]. The fifth returns a valid response. Text between tool calls comes back in thinking blocks whose text is empty at the default display setting [3], so one change in five arrives as a success [7].
Throwing the work away is cheap in tokens. Output runs $20 per million and one response caps at 128K tokens [1][5]. A discarded maxed-out reply costs $2.56 [20]. One edit needs no watching at all. claude.dev says to remove "think carefully" and similar lines from prompts and saved instructions, and reports that in its testing in a chat product, removing the line made replies start sooner with no clear drop in quality [17].
Maggie Appleton of GitHub Next [19] quotes the HCI researcher Lucy Suchman: "Plans are best viewed as a weak resource for ad hoc activity" [18]. claude.dev's suggestion is a plan of that kind, a short rule in CLAUDE.md about when to stop and ask and when to keep going [21].
The two things that matter are whether a change fails a request and whether a user sees the difference. The four breaking ones fail at the call site, where your error handling already is. The display change succeeds and reaches everyone who was reading a status line. The only check for it is a person opening the app and watching one run from the first tool call to the last.
What to watch
- A default display value that returns the text between tool calls would close the progress-update gap, and no application would have to change.
- Whether the output-300k-2026-03-24 beta header on the Message Batches API becomes generally available. GA would raise what a discarded run costs.
- Whether stop-to-report behaviour on long runs becomes a documented setting rather than a rule operators write into CLAUDE.md.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Claude Opus 5.5 is priced at $4 / $20 USD per million input / output tokens.
ReportedView cited source - [2]
Four breaking changes affect code already running on Claude Opus 5: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and, on the Claude API and Google Cloud, the earlier computer_20251124 computer use tool is not accepted.
ReportedView cited source - [3]
A further change alters the response shape without failing any request: text between tool calls comes back in thinking blocks whose text is empty at the default display setting.
ReportedView cited source - [4]
An application that streams that text to its users as progress updates goes quiet between tool calls until it sets a display value that returns the text.
ReportedView cited source - [5]
Claude Opus 5.5 has a max output of 128K tokens, and adaptive thinking is always on and cannot be turned off; depth is controlled with the effort parameter.
ReportedView cited source - [6]
On the Message Batches API, Claude Opus 5.5 supports up to 300k output tokens with the output-300k-2026-03-24 beta header.
ReportedView cited source - [8]
A turn is one round trip inside the loop: Claude produces output that includes tool calls, the SDK executes those tools, and the results feed back to Claude automatically, without yielding control back to your code.
ReportedView cited source - [9]
Turns continue until Claude produces output with no tool calls, at which point the loop ends and the final result is delivered.
ReportedView cited source - [11]
A ResultMessage marks the end of the agent loop and contains the final text result, token usage, cost, and session ID.
ReportedView cited source - [10]
max_turns / maxTurns counts tool-use turns only, and max_budget_usd / maxBudgetUsd caps turns based on a spend threshold.
ReportedView cited source - [13]
Early testers had Opus 5.5 run long coding tasks for hours with little oversight, and compared with prior Opus models its biggest gains are on multi-step work, such as carrying a change through a large repository until the tests pass.
ReportedView cited source - [22]
Early testers had Opus 5.5 coordinate parallel subagents on long audits and migrations with little oversight, and claude.dev's example prompt instructs: give each service to its own subagent, and when a subagent reports back, check its evidence.
ReportedView cited source - [15]
With no design direction, Opus 5.5 falls back on a few default styles; a general instruction like "avoid a generic look" mostly swaps one default for another, while a list of specific patterns to leave out works much better.
ReportedView cited source - [16]
On a long task Opus 5.5 sometimes stops to report instead of going on: a summary that names the next step without taking it, an offer to continue, or a list of choices that don't block the work.
ReportedView cited source - [17]
claude.dev advises removing "think carefully", "think step by step" and similar lines from prompts and saved instructions, and says that in its testing in a chat product, removing a "think carefully" line made replies start sooner, with no clear drop in quality.
ReportedView cited source - [21]
claude.dev advises putting a short rule in the CLAUDE.md file about when to stop and ask and when to keep going, and to keep permission prompts on for destructive commands.
ReportedView cited source - [18]
Lucy Suchman, quoted by Maggie Appleton: "Plans are best viewed as a weak resource for ad hoc activity."
- [19]
Maggie Appleton is a designer, engineer and researcher at GitHub Next, the R&D lab arm of GitHub.
ReportedView cited source - [7]
Of the five documented changes to code already running on Claude Opus 5, four are listed as breaking and one returns a valid response, so one change in five does not surface as a failed request.
Derived - [20]
A single maximum-length Opus 5.5 response of 128K output tokens costs $2.56 in output tokens at list price, so discarding and rerunning it costs $2.56.
Derived
Sources & coverage · 3 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- maggieappleton.comSep 13Planning with Agents: Divided Worlds, Boundary Objects, and Thicker Interfaces
- platform.claude.comHow the agent loop works
- platform.claude.comClaude Opus 5.5 - Claude Platform Docs
Additional citations
- Lucy Suchman, quoted in Maggie Appleton's talk write-up