Claude Opus 5.5 Pricing: 20% Cheaper, 4 Breaking Changes

Anthropic released Claude Opus 5.5 on Tuesday at $4 in / $20 out per million tokens, 20% under Opus 5, with four breaking API changes.

TechCrunch reports the model sets a new state of the art in coding and knowledge work, according to the company. But the number that will move most production bills is not the headline rate. It is the cache read price, which fell three times further.

What it costs

Per million tokens	Opus 5.5	Opus 5	Change
Input	$4	$5	20% cheaper
Output	$20	$25	20% cheaper
Cache reads	$0.20	$0.50	60% cheaper
Cache writes (5m)	$5	$6.25	20% cheaper

Figures from Anthropic's Opus 5.5 announcement.

TechCrunch confirms the output rate independently: $20 per million against $25 on the previous model, with other metrics showing similar price drops.

Cache reads deserve the attention. On agentic and coding work, cache reads are where the majority of spend lands, and they dropped 60% against 20% on the headline rates. If your workload is a long-running agent re-reading the same context on every turn, that row is your invoice.

Two rates sit outside the table. Fast mode is priced separately at $8 / $40 per million for up to 2.5x the speed, available in Claude Code and the Claude Platform. Batch API work remains 50% off input and output.

The 40% claim, and who checked it

Anthropic measures a 40% lower total cost on typical workloads at default settings, plus 30%+ faster output generation. That number is doing two jobs at once: part of it is the price cut, part is the model spending fewer tokens per task. The Verge reports the same 40% figure, framed as cost to run.

Two customers reported the token half from their own side. Box said the model used a third of the tokens Opus 5 did in their evaluations, with answers 40% less verbose and no accuracy loss. Factory called it the first model they would default to at medium effort, matching Opus 5 at high effort with 20 to 25% fewer output tokens.

Anthropic's task-level examples are more aggressive and should be read as vendor-selected best cases. The company claims a 200,000-line codebase audit and fix finished in under three hours where Opus 5 took over 20 hours and burned 2.5x the tokens, and that an early tester completed a 680,000-line migration in less than a day.

Benchmarks, and Anthropic's own hedge

Evaluation	Opus 5.5	Opus 5	Fable 5.1	GPT-6 Astra
Terminal-Bench 4.0 (agentic coding)	66.4%	52.3%	55.8%	57.9%
FrontierCode v1.1	54.4%	48.0%	50.3%	53.3%
CursorBench 4.0	57.8%	46.6%	51.8%	no data
GDPval-AA v2.1 (knowledge work, Elo)	1846	1708	1735	1542
AutomationBench (business workflows)	40.0%	26.9%	31.4%	41.4%
Terminal-Bench-Science 0.1	58.7%	29.0%	52.6%	64.6%

Figures from Anthropic's Opus 5.5 announcement; Opus 5.5 results use adaptive thinking at max effort.

Read the footnotes before you act on that table. Anthropic hedges it directly: at this capability level, benchmark margins are a less reliable guide to real differences, and in the company's own use the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. Terminal-Bench 4.0 carries a standard error of plus or minus 2.6 points for Opus 5.5.

There is a second asterisk. Opus 5.5 was evaluated with production safeguards switched on, so cybersecurity tasks fell back to Opus 4.8 and biology tasks to Opus 5 whenever the safeguards intervened, which likely pulled those scores down. GPT-6 Astra still leads on AutomationBench and Terminal-Bench-Science; everywhere else in the published set, Opus 5.5 is first.

Four breaking changes before you migrate

1. Thinking cannot be disabled. Adaptive thinking is always on, and depth is controlled only through the effort parameter, which defaults to medium.
2. Forced tool use returns an error. Requests that force the model to call a specific tool are rejected outright.
3. Thinking blocks are tied to the model and the conversation. You cannot carry them across.
4. The old computer use tool is rejected. computer_20251124 is not accepted on the Claude API or Google Cloud.

The first three also apply to Fable 5.1, so a team already running Fable has less to fix.

A fifth change alters the response shape without failing anything, and it is the one most likely to reach your users. Text between tool calls now comes back inside thinking blocks, and at the default display setting that text is empty. If your app streams that text as progress updates, it goes quiet between tool calls until you set a display value that returns it. Nothing throws. The UI simply stops talking mid-task.

Safeguards you will feel

This is the first model Anthropic has released since CEO Dario Amodei announced plans to "pace the frontier," and the safety posture shows it. In testing, Opus 5.5 attempted to circumvent boundaries 85 percent less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low severity and self-reported, according to Anthropic. The company says it was tested by outside partners including Frontier Design and METR before release.

The practical cost lands on security teams. Because its biology and cybersecurity capabilities now match Mythos 5.1, Opus 5.5 ships with Fable 5.1 class safeguards: certain cybersecurity-related requests are transparently rerouted to the less powerful Opus 4.8, and flagged biology requests go to Opus 5. Finding and fixing bugs in the normal development lifecycle still works. If your product does offensive security work, some of those requests will be answered by an older model, and the reroute is transparent rather than a visible error.

Availability and spec

The model ID is claude-opus-5-5, and it shipped the same day on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. On Bedrock it is anthropic.claude-opus-5-5. Vercel has it live on AI Gateway, describing it as a step-change improvement over Opus 5 with its biggest gains in agentic coding, long-running agent tasks and knowledge work.

The rest of the spec: 1M token context, 128K max output — up to 300K on the Batch API with the beta header — a June 2026 knowledge cutoff, and retirement no sooner than 22 September 2027. Claude Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.

What to watch

• Your cache read line, not your output line. The 60% cut is where an agentic workload actually gets cheaper. Pull last month's usage by token type before you model the savings.
• Silent progress streams. Ship the display-value fix in the same PR as the model bump, or your users get a frozen UI on long tool chains.
• Sonnet 5.5 and Haiku 5.5 pricing. Due in the coming weeks, and the deciding factor for anyone routing cheap traffic below Opus.
• Reroute rates on security work. Anthropic says a Cyber Verification Program expansion is coming; until then, measure how often your security prompts land on Opus 4.8.