Claude Opus 5.5 drops to $4 and $20, and breaks four Opus 5 habits

Swap the model string on an agent that forces a tool call, and you'll get a 400 before the first token. That's the Claude Opus 5.5 launch in one line. The new model, claude-opus-5-5, went live on 22 September at $4 per million input tokens and $20 per million output, down from Opus 5's $5 and $25, and Anthropic's own table puts it ahead of Fable 5.1 on most rows. It also refuses four request shapes that Opus 5 accepted, and the cheapest bill in the world doesn't help a pipeline that won't start.

The price cut is bigger on cache than on output

Twenty percent off the headline. That's the part everyone quoted. The line we'd actually watch is cache reads: $0.20 per million, which is 5% of the base input price. On Opus 5 a cache read was 10% of $5, so fifty cents. A 60% cut. If you run long agent sessions that re-read the same 200K of repository context every turn, that's where the invoice moves, not on output.

The rest of the card is plainer. Five minute cache writes cost $5 per million, one hour writes $8, and batch halves everything to $2 and $10. The context window is still 1M tokens, max output is still 128K on the synchronous API, and batch can go to 300K with the output-300k-2026-03-24 beta header. Knowledge cutoff is June 2026. Retirement is "not sooner than" 22 September 2027, so you've got a year of guaranteed runway if you pin it.

Anthropic also claims the model spends about 40% less than Opus 5 on typical workloads, partly because it writes shorter answers. We can't check that yet. It's the vendor's number on the vendor's mix, and it runs straight into a second fact from the same docs: at a given effort level, Opus 5.5 thinks more per turn than Opus 5, "most of all at xhigh and max". Thinking tokens bill as output. So a cheaper rate and a hungrier model can net out anywhere, and I honestly don't know where they'll land on a real coding loop until someone publishes token counts per task.

One ugly data point already exists. Simon Willison ran his usual SVG test at max effort and the model burned through the whole 128K output budget mid reasoning, twice, at $2.56 a run and close to 20 minutes each. Fable 5.1 finished the same prompt. It's one prompt. It's also the kind of failure you don't want to discover on a nightly batch.

Four request shapes that now return a 400

Here's the list from Anthropic's migration notes, in the order we'd check a codebase. First, thinking can't be disabled. Opus 5 accepted thinking: {"type": "disabled"} at high effort or below. Opus 5.5 answers that, and any manual budget_tokens, with an invalid_request_error. Drop the field or send {"type": "adaptive"}, then use effort as the dial.

Second, forced tool use is gone. tool_choice set to any or to a named tool fails, including on the token counting endpoint. We wrote about the same break on Fable 5.1 three weeks ago, and it's the one that bites structured extraction jobs hardest. The documented fix is auto plus strict tool use, or moving the schema to structured outputs.

Third, the older computer_20251124 tool is refused on the Claude API and Google Cloud. You need the computer_toolset_20260801 toolset there. Bedrock still accepts the old tool, which is a nice trap if you test on one cloud and deploy on another. Fourth, thinking blocks are bound to the model and the conversation. Opus 5.5 reads blocks from Opus 5, but not from Fable or Mythos, and on accounts created from 31 August 2026 a replayed block after an edited system prompt or tool list returns a 400.

Then there's the change that doesn't error at all, which we think is the nastiest. The short notes the model writes between tool calls now arrive as thinking blocks, and at the default display: "omitted" their text is empty. If your UI streams those notes as progress updates, it just goes silent. No exception, no log line. A user thinks the agent froze.

The benchmark table, and where we'd use it

Anthropic's launch page puts Opus 5.5 at 66.4% on Terminal-Bench 4.0, against 55.8% for Fable 5.1 and 52.3% for Opus 5. That's a mid tier model beating the top one by more than ten points on long terminal work, at 40% of Fable's $10 and $50 list price. Other rows are closer: 57.8% on CursorBench 4.0 against Fable's 51.8%, 54.4% on FrontierCode v1.1 against 50.3%. On AutomationBench it scores 40.0%, a hair behind the 41.4% the same page gives GPT-6 Astra.

Two cautions. The page doesn't say which effort level produced each row, and the default is now medium, so your out of the box result may not look like the table. And harnesses disagree: xAI's Grok 4.7 page listed Fable 5.1 at 57.9% on Terminal-Bench 4.0, two points above Anthropic's own figure. No independent run had been published when we wrote this.

Our take: if you're on Opus 5 with thinking already on and no forced tools, move now, pin effort to whatever you ran before, and compare a week of invoices. If you disabled thinking to save money, the math changes, because that option is gone and low effort is the closest substitute. If you're on Fable 5.1 for agent work, we'd run a side by side before paying 2.5x for it again. I'd keep Fable for the hardest one shot tasks for now, mostly because of that max effort result. Sonnet 5.5 and Haiku 5.5 are due "in the coming weeks" per TechCrunch, so the cheap end of the lineup is about to move too.

A quick check that your key can see the model before you touch a config:

Sources

Release date, pricing, cache rates, context, output limits, cutoff, retirement date and availability come from Anthropic's Claude Opus 5.5 model page. The four breaking changes, the silent progress update change, the default effort and the extra thinking per turn are from What's new in Claude Opus 5.5. Benchmark scores and the 40% workload claim are from the Opus 5.5 launch page. The previous $25 output price and the coming Sonnet and Haiku releases are reported by TechCrunch. The max effort test and the 60% cache cut are from Simon Willison. The figures are ours, built from those pages.

Frequently asked questions