Claude Opus 5.5 Ships Cheaper, With Four Breaking Changes

What Anthropic Actually Shipped

On September 22, Anthropic released Claude Opus 5.5, and the framing this time is different from the last Opus release. Where Opus 5 was pitched as a step-change in capability, Opus 5.5 is pitched mostly as an efficiency release. The model ID is `claude-opus-5-5`, it is live now on the Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and Anthropic says it matches the performance of Fable 5.1, its most capable public model, on most tasks while running meaningfully cheaper in practice, since it tends to finish the same task using fewer tokens.

That framing lines up with how I read the actual spec sheet. The context window stays at 1M tokens, the max output stays at 128k tokens, both identical to Opus 5. Nothing changed there. What did change is the default effort level, which drops from high on Opus 5 down to medium on Opus 5.5. Effort is the lever that controls how much the model thinks before it answers, and a lower default means a request that used to run at high effort automatically now runs lighter unless you explicitly ask for more. That is not a capability cut so much as a bet that the model needs less thinking by default to hit the same bar, and it is consistent with Anthropic's own note that real-world workloads see the biggest savings because the model completes tasks in fewer tokens overall.

Latency lands in the middle of the current lineup too. Anthropic's own comparison table puts Opus 5.5 at "moderate" latency, slower than Sonnet 5 and Haiku 4.5, faster than Fable 5.1. That is the same tier Opus 5 sat in, so nothing surprising there. What is worth noting is retirement: Anthropic committed to keeping Opus 5.5 active for at least a year, not sooner than September 22, 2027, which is the standard runway the company gives a model at launch and a reasonable signal that this is meant to be a daily driver, not a stopgap before the next Opus.

The Four Breaking Changes Worth Checking Before You Upgrade

This is the part of the release that matters most if anything in a RAXXO tool, or any project, already calls the Claude API against an Opus model. Anthropic lists four breaking changes on Opus 5.5, and unlike a routine model swap, these can fail a request outright rather than just changing the output quality.

The first is that thinking can no longer be disabled at all. On Opus 5, you could turn thinking off as long as your effort level was high or below, xhigh and max were the only levels that forced thinking on. On Opus 5.5, adaptive thinking is always on, full stop, at every effort level. If a setup explicitly disables thinking anywhere in its request, that call needs to drop the flag entirely rather than just adjust the effort level.

The second is forced tool use. Setting `tool_choice` to `any` or to a specific `tool` used to force the model to call a tool on that turn. On Opus 5.5, both of those choices return an error. Only `auto` and `none` are supported now. Anything that depended on forcing a specific tool call, a common pattern for structured output or a guaranteed function call, needs a different approach on this model.

The third is subtler: thinking blocks are now tied to both the specific model and the specific conversation that produced them. Carrying a thinking block over into a different model or a resumed conversation elsewhere is no longer something to rely on. It is the kind of change that will not show up in testing, only in production, if a system is passing thinking blocks between sessions or models.

The fourth is narrower but still worth flagging for anyone using computer use: the earlier `computer_20251124` tool version is no longer accepted on the Claude API or Google Cloud with this model. Anthropic replaced it with a newer toolset earlier in the year, and Opus 5.5 is the point where the old version actually stops working rather than just being discouraged.

There is a fifth change that will not break a request but can go quiet in production: text that used to stream between tool calls now comes back inside `thinking` blocks, and that text is empty at the default display setting. Anything that streamed that text to a user as a live progress indicator will go silent between tool calls unless the display setting is changed to return it. The first three of these four breaking changes also apply to Fable 5.1, so this is not an Opus-only migration if a project already moved to Fable 5.1 back on September 1.

Where the Pricing Actually Lands

Anthropic cut the price of Opus 5.5 by exactly 20 percent on both input and output tokens compared to Opus 5, and that cut carries through consistently across every related rate: the 5 minute cache write, the 1 hour cache write, and Fast mode pricing all dropped by the same 20 percent. That is a straightforward, uniform cut rather than a headline number that only applies to one usage pattern.

The one rate that moved by more than 20 percent is the cache read price. A cache hit on Opus 5.5 now costs half of what the standard cache multiplier would charge, a better rate than Opus 5 offered on cached reads relative to its own input price. For anything that leans on prompt caching, and most agent loops with a long system prompt or repeated context do, that compounds with the base price cut rather than sitting on top of it.

None of this changes what Opus 5 already offered: it was priced identically to Opus 4.8 with no increase at all. Opus 5.5 is the tier below Fable 5.1 getting meaningfully cheaper again, on top of a generation that had already held the line on price while adding capability. The pattern across the last two Opus releases has been the same: hold or cut the price, push the improvements into capability and efficiency instead of a bigger bill.

Fast mode also carries through to Opus 5.5, still in research preview, still available on the Claude API only, still running roughly two and a half times faster than the default at a premium rate. That premium also dropped 20 percent alongside everything else, so the trade for lower latency did not get more expensive even as the baseline got cheaper.

Why This Is Different From a Routine Point Release

A 20 percent price cut alone would be a nice-to-know, not something I'd write about. What makes Opus 5.5 worth a full piece is that it pairs a real price cut with breaking changes that can silently fail existing code, and that combination is exactly where I pay closest attention as someone who runs Claude Code daily inside agentic workflows.

The thinking and tool_choice changes both point at the same shift Opus 5 already started: Anthropic is optimizing this model line for long, autonomous agent loops rather than single-turn requests where a developer manually steers behavior with flags. Always-on adaptive thinking and the removal of forced tool choice both push toward "let the model decide," which tracks with how these models get used in practice now, running inside Claude Code, resolving multi-step tasks, calling their own tools without a human confirming each step.

The five-platform simultaneous launch is also worth noting on its own. Opus 5 launched across four platforms back in July. Opus 5.5 adds Claude Platform on AWS to that list, Anthropic's own AWS Marketplace offering, on top of the Claude API, Bedrock, Google Cloud, and Microsoft Foundry it already covered. That is the widest same-day availability an Opus release has had yet, and it means nobody building on a specific platform has to wait for parity this time either.

I wrote about the last Opus release, including the context window and effort-level changes it introduced, in Claude Opus 5 Is Here: Fable 5 Intelligence at Half the Price, and about the Fable and Mythos 5.1 release earlier this month in Claude Fable 5.1 and Mythos 5.1: What Changed on September 1. Reading the three releases together, the direction is consistent: fewer manual levers, more model judgment, and a price curve that keeps bending down rather than up.

Bottom Line

Claude Opus 5.5 is a real release, not a routine version bump, but not for the reason most headlines are leading with. The price cut is real and uniform across the board, 20 percent on nearly every rate tied to this model. The part that actually needs attention this week is the four breaking changes bundled into the same release: thinking that can no longer be turned off, forced tool use that now errors out, thinking blocks that no longer travel between models or conversations, and an older computer use tool version that stops working outright.

For anyone running Opus 5 in production today, the upgrade is not a drop-in swap. It is worth reading Anthropic's own migration notes before flipping the model ID, especially if a workflow leans on forced tool calls or explicitly disabled thinking anywhere. For everyone else just watching where the field is heading, the signal is the same one Opus 5 already sent in July: Anthropic keeps making its mid-tier model cheaper and more autonomous at the same time, and the gap to the flagship keeps closing from below rather than the flagship pulling further ahead. I will keep watching how Opus 5.5 holds up once more real workloads move onto it, and write again if the picture changes.