Claude Opus 5.5 Arrives: Longer Tasks Need Visible Progress
TL;DR
Anthropic releases Opus 5.5 with lower token prices and claims of stronger long-running agents. Its default lack of text between tool calls changes how products should communicate progress and cancellation.
With Claude Opus 5.5, an agent using an older streaming interface may stop displaying text between tool calls. Anthropic documents this default behavior: the model can continue working while users lose one indication of progress. If a product relies on intermediate narration to make waiting understandable, its ability to distinguish “still running” from “stalled” needs to be tested again after the upgrade.Behavior
The official event date is 2026-09-22; the announcement supplies neither a publication time nor a timezone. MacRumors published its report at 12:23 PDT that day, equivalent to September 23 at 03:23 in Taipei. This article carries the Taipei date of September 23, with a reporting cutoff of 21:00. It covers an overnight release without treating the report’s timestamp as the model’s activation time.Announcement Independent report
Opus 5.5 targets software development, computer use and long-running agent work. Anthropic describes a tester completing a code migration involving 680,000 lines in less than a day. That supplies a concrete task scale, but it remains a vendor-presented example. It cannot establish typical delivery speed or demonstrate that arbitrary projects can proceed unattended.Case
Standard API input and output cost $4 and $20 per million tokens respectively, each 20% below Opus 5. Lower rates reduce spending at unchanged usage; different reasoning lengths or tool-call counts can still change the final bill. AWS also announced availability through Amazon Bedrock. This is an accessible service, although a model listing should not be interpreted as identical capacity in every region.Pricing and migration Cloud availability
Tools can keep working while the screen waits for text
Opus 5.5 uses adaptive reasoning, allowing the model to determine its reasoning effort. Updates between tool calls now arrive in thinking blocks whose text is empty by default. The documented display: "updates" beta option or "summarized" setting returns content, which the interface must also read from those blocks. A product that previously treated ordinary streamed text as a progress indicator may continue executing after a model-name change without preserving its previous waiting experience.Model behavior Migration guidance
This interface change matters to me because longer tasks ask users to accept longer periods without a final result. Suppose an agent is modifying several modules while the screen stays blank. A user might submit the instruction again or restart the task. That is a hypothetical interaction risk; the sources provide no observed incidence rate. But if the product cannot distinguish active execution from a failed connection, it leaves that judgment to someone who cannot see the tools’ state.
I would show the current tool action and the last check confirmed as complete, alongside an understandable way to cancel. Those signals should come from execution records, rather than model-generated reassurance masquerading as progress. A tool that has not replied should be shown as awaiting a response; a disconnected session should have an explicitly unknown status. These are product-design choices derived from the documented behavior, not features Anthropic has already supplied for every application.
Nor should the product notify users about every tool action. Short tasks can simply return their results. Work that requires waiting needs enough status information to decide whether to continue, cancel or return later. Exposing extensive technical logs adds reading burden. Helping people understand their available actions is more useful than reproducing every step of the model’s narration.
The release gives developers lower token prices and new claims about long-task capability, but no user research on redesigned waiting interfaces. Fewer restarts caused by unclear progress, without more failures being mislabeled as active work, would support a claim that the interface improves waiting. How long the model can keep working and whether people can confidently let it continue require different evidence.
The cover reuses this site’s Claude Opus 5 branding image, not an Opus 5.5 performance chart. External image downloads failed because hostnames could not be resolved.
Sources:
Related Articles
Claude Unifies Cowork and Launches Docs: Shared Drafting Still Needs an Approved Record
Anthropic combines chat and Cowork and introduces collaborative Docs. Fewer tool switches help, but missing version history, organizational sharing limits and phased access still shape its use.
Claude Finds a New Enzyme System: One Discovery Cannot Schedule the Next
Anthropic reports that Claude identified ART, with experiments detecting associated short RNAs but not establishing enzyme activity. Around 950 agents produced the lead; unsuccessful reruns limit claims about predictable research output.