What is Claude Opus 5.5, and what changes if you move from Opus 5?

What Opus 5.5 is

Anthropic released Claude Opus 5.5 on September 22, 2026 and calls it the first model in its Claude 5.5 family. The announcement says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, so today Opus 5.5 is the only 5.5 model you can call.

It is a hosted, proprietary model. The API ID is claude-opus-5-5, and it is also listed on Amazon Bedrock (as anthropic.claude-opus-5-5), Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The model page lists a 1M-token context window, 128K max output tokens, a June 2026 knowledge cutoff, and a retirement date no sooner than September 22, 2027.

The price change, stated plainly

List prices per million tokens fall across the board. The figures below come from Anthropic's model page and announcement; check the live pricing page before you budget, because list prices change.

• Input: 4 US dollars on Opus 5.5, against 5 on Opus 5.
• Output: 20 US dollars, against 25.
• Cache reads: 0.20, against 0.50. Anthropic notes cache reads make up most of the cost of agentic and coding work, so this is the line most likely to move a real bill.
• Cache writes: 5 for the 5-minute cache and 8 for the 1-hour cache.
• Batch processing is half price, 2 and 10.
• Fast mode, a research preview on the Claude API only, costs 8 and 40 and Anthropic says it is up to 2.5 times faster.

What Anthropic says improved

Anthropic reports that Opus 5.5 costs about 40 percent less than Opus 5 on typical workloads at default settings, because it is cheaper per token and uses fewer tokens per task, and that it generates output more than 30 percent faster. On its own benchmark table it reports 66.4 percent on Terminal-Bench 4.0 against 52.3 percent for Opus 5, 54.4 percent on FrontierCode against 48.0, and 81.8 percent partial credit on OSWorld 2.0 against 74.0.

Treat those as vendor-reported numbers. Anthropic itself writes that at this capability level benchmark margins have become a less reliable guide to real-world differences. One independent data point: Artificial Analysis listed Opus 5.5 at max effort first of 661 models on its Intelligence Index (score 58) when we checked, and also flagged it as verbose, using 260M output tokens on the index against a median of 91M. Rankings on that site change as models are added.

Four breaking changes to test before you switch

Anthropic's what's-new page lists four changes that can make an Opus 5 request fail on Opus 5.5. Each one returns a 400 invalid_request_error rather than degrading quietly, which at least makes them easy to catch in a test run.

• Thinking cannot be disabled. Sending thinking type disabled, or a manual thinking budget, is rejected. Omit the thinking field or send type adaptive, and use the effort parameter to control depth and cost.
• Forced tool use is not supported. tool_choice set to any or to a named tool is rejected. Use auto, and use strict tool use or structured outputs if you need schema-valid JSON.
• Thinking blocks are tied to the model and conversation. Opus 5.5 can read thinking blocks from Opus 5 and earlier, but not from Fable or Mythos models. For API accounts created on or after August 31, 2026, replaying a thinking block after editing the system prompt, tools, or earlier messages returns a 400 by default. Keep conversations append-only.
• On the Claude API and Google Cloud, the older computer_20251124 computer-use tool is rejected. Move to the computer_toolset_20260801 toolset. Amazon Bedrock still accepts the older tool.

Changes that do not throw an error

Three behavior changes will not fail a request, which makes them easier to miss.

• The default effort is now medium. Opus 5 defaulted to high, so a request that never sets effort will think less than it used to. Set effort explicitly and re-run your quality checks.
• Text the model writes between tool calls now arrives in thinking blocks, which are empty at the default display setting. An app that streams those progress notes to users will go quiet between tool calls until it sets a display value that returns the text.
• There are more safeguard categories. Anthropic says most cybersecurity tasks are re-routed to Opus 4.8, a biology classifier now runs alongside the cyber one, and requests that push the model to reproduce its internal reasoning can be refused. Refusals come back as HTTP 200 with stop_reason refusal, so handle that value and configure a fallback model.

What it means for Claude subscribers

Anthropic says it is raising five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans alongside the launch, and giving subscribers a rate-limit reset they can save and use when they choose. The announcement does not give the new limit figures, so we cannot tell you how much more usage you get.

Who should switch, and who should wait

Switching looks straightforward if you already run Opus 5 with adaptive thinking, auto tool choice, and the newer computer-use toolset, and your bill is dominated by cached input. The per-token cuts alone lower cost, before any token-efficiency gains.

Wait and test first if your integration disables thinking, forces a specific tool, edits earlier turns of a conversation, streams progress text to users, or does security research work that may now be re-routed or refused. Those are the paths that change.

Limits of this answer

RepoRadar has not run Opus 5.5 hands-on. Prices, specifications, and breaking changes come from Anthropic's documentation as checked on September 23, 2026. Benchmark scores are vendor-reported unless marked as Artificial Analysis. Anthropic also notes in its announcement that Opus 5.5 often suspects it is being evaluated, which it says makes pre-release safety testing harder to interpret.

A useful next action

Before changing the model ID in production, search your code for thinking type disabled, budget_tokens, tool_choice any or tool, and computer_20251124. Fix those, set effort explicitly, add handling for stop_reason refusal, then run your existing evaluation set against claude-opus-5-5 and compare cost and quality with Opus 5 on the same tasks.