Claude Opus 5.5 release cover with a terracotta gateway and flowing lines

Release Watch

Claude Opus 5.5 Release Date: What Changed on September 22

Q: What is the Claude Opus 5.5 model ID?

claude-opus-5-5, a fixed ID with no date suffix. EvoLink uses the same string on both the Chat Completions and Messages endpoints. On Bedrock the ID is anthropic.claude-opus-5-5.

Q: Can I turn thinking off on Claude Opus 5.5?

Not on either path. On Anthropic's native API, thinking: disabled and budget_tokens both return a 400. On EvoLink, the gateway converts those legacy settings to adaptive thinking instead of rejecting them, but they still cannot turn thinking off. Omit the field or send adaptive, and control depth with effort (low to max, default medium).

EvoLink Team

Product Team

September 22, 2026

14 min read

Claude Opus 5.5 was released on September 22, 2026. Anthropic lists it as the newest Opus model, built for long-running agentic coding and knowledge work, at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5 and $25. The model ID is claude-opus-5-5, the 1M-token context window and 128K max output carry over from Opus 5, and thinking is now always on. Four documented breaking changes affect code already running on Opus 5.

For EvoLink users, the model is already callable: the Claude Opus 5.5 model page is the source of truth for the live route, current EvoLink rates, and integration settings. This article covers what Anthropic released, what it changes for developers, and how the week of "Opus 5.2" rumors resolved. The upgrade decision itself, with a cost model and a canary plan, is in Claude Opus 5.5 vs Claude Opus 5.

Claude Opus 5.5 Release Status

Item	Confirmed status
Release date	September 22, 2026
Anthropic lifecycle	Active; recommended starting point for most workloads
Official model ID	`claude-opus-5-5` (no date suffix, same scheme as `claude-opus-5`)
Standard price	$4 / MTok input, $20 / MTok output
Prompt caching	5-minute write $5, 1-hour write $8, cache read $0.20 per MTok (5% of input)
Batch API	$2 / MTok input, $10 / MTok output
Context window	1 million tokens
Maximum output	128,000 tokens (300K on the Batch API with a beta header)
Input and output	Text and images to text
Thinking	Adaptive, always on; default effort `medium`
Knowledge cutoff	June 2026
Published retirement floor	Not before September 22, 2027 on Anthropic-operated platforms
Official availability	Claude API, Amazon Bedrock (`anthropic.claude-opus-5-5`), Claude Platform on AWS, Google Cloud, Microsoft Foundry

These are provider-published facts from Anthropic's release notes, models overview, and pricing page, checked on September 22, 2026. Gateway availability, regional terms, throughput, and billing still vary by channel and account, so verify the contract you actually use.

What Changed in the September 22 Release?

The headline is a price cut on every token dimension with the same context and output limits. Standard input and output fall 20% against Opus 5. Cache reads fall from $0.50 to $0.20 per million tokens, a 60% cut, because Opus 5.5 prices cache hits at 5% of the input rate, where other Claude models use 10% (Fable 5.1 and Mythos 5.1 use 2.5%). Cache writes fall in step with input: $5 and $8 versus $6.25 and $10.

Anthropic's announcement goes further and says Opus 5.5 "costs 40% less to run than Opus 5" and generates output "more than 30% faster." Read those two numbers carefully. The 20% and 60% figures are list-price arithmetic you can verify on the pricing page. The 40% figure is Anthropic's measurement of total cost on its own task mix, which combines the lower unit rates with fewer tokens spent per completed task. It is a vendor estimate for a workload you do not run. Anthropic's own documentation also notes that at a given effort level Opus 5.5 tends to think more per turn than Opus 5, so a request that carries its old effort setting over unchanged may not see the saving the announcement describes.

On capability, Anthropic positions Opus 5.5 for long-running agentic coding, analyst-grade knowledge work, and reading dense charts, diagrams, and screenshots. The announcement reports a FrontierCode score of 54.6%, describes it as beating GPT-6 Astra at roughly 20% of the cost at default effort, and says it "performs at the level of Claude Fable 5.1 on most work." We have not reviewed the evaluation configurations behind those statements, so this article treats them as claims to reproduce on your own traces, not as results.

Four Breaking Changes Existing Integrations Must Test

Code running on Opus 5 does not move to Opus 5.5 by changing the model string alone. Anthropic documents four changes that return a 400 error or silently alter behavior.

Scope note: the 400 errors below are Anthropic API behavior. On EvoLink, the gateway converts legacy thinking: disabled and budget_tokens settings to adaptive thinking instead of rejecting them, while forced tool choice is rejected on both the Chat Completions and Messages endpoints; the Claude Opus 5.5 model page documents the gateway behavior.

Thinking cannot be disabled. thinking: {"type": "disabled"} and the manual budget_tokens form both return 400 invalid_request_error. Omit the thinking field or send adaptive, and use effort to control depth. If you disabled thinking to save tokens, use a lower effort level instead.
Forced tool use is rejected. tool_choice types any and tool return a 400, on the token-counting endpoint too. Keep auto, set strict: true on the tool for schema-valid arguments or move the schema to structured outputs, and say in the prompt when the tool applies.
Thinking blocks are bound to the model and the conversation. Opus 5.5 reads thinking blocks from Opus 5 and earlier Opus, Sonnet, and Haiku models, but not from Fable or Mythos. Going the other way, only Claude Fable 5.1 and Mythos 5.1 read Opus 5.5 blocks; a fallback to any other model runs without them. Editing the system prompt, tools, or earlier turns mid-conversation invalidates the blocks, and accounts created on or after August 31, 2026 get a 400 by default when they replay one after such an edit.
The computer_20251124 tool is not accepted on the Claude API or Google Cloud. Declare computer_toolset_20260801 instead and update the agent loop for member tool-use blocks. On Amazon Bedrock the older tool keeps working.

A fifth change breaks nothing but changes what users see: the short notes the model writes between tool calls now come back as progress-update thinking blocks rather than text blocks. At the default display: "omitted" their text is empty, so an application that streamed those notes as progress goes quiet between tool calls. Set display to "updates" (beta, header thinking-display-updates-2026-08-18) or "summarized" and render the non-empty thinking blocks to restore them.

Two defaults also shift. The default effort is medium where Opus 5 ran at high, so a request that omits effort now does less work. And the safety classifiers cover more categories: expect stop_details.category values such as bio and reasoning_extraction alongside cyber, and note that server-side fallback does not retry reasoning_extraction refusals.

How the Opus 5.2 Rumor Resolved

Between September 14 and 21, community posts described Claude Code quietly routing some Opus 5 requests to a newer build they called "Opus 5.2", and third-party pages circulated three candidate names and an internal codename. Our Opus 5.2 release tracker recorded on September 18 that Anthropic documented no such model, ID, or price.

The release settles the identity question the way our tracker said it would: with a dated release note, a model page, a callable ID, documented prices, and a migration guide, all under the name Claude Opus 5.5. What it does not settle is whether the September routing reports were early traffic on this model. Anthropic has not said so, no "Opus 5.2" product was ever announced, and the two names remain separate entities on this site. Treat "Opus 5.2" as a community label whose story ended, not as an alias for claude-opus-5-5.

The same discipline applies to the next name in circulation. Anthropic has not announced Claude Opus 6; our Opus 6 release watch continues to track it as an unconfirmed future generation, and Opus 5.5 is a point release inside the current one.

Opus 5.5, Opus 5, Sonnet 5, and Fable 5.1 Are Different Routing Decisions

Model	Current role	Routing implication
Claude Opus 5	Active predecessor, $5 / $25, retirement not before July 24, 2027	Keep as the measured rollback while 5.5 is canaried; re-test before assuming it is redundant
Claude Opus 5.5	Newest Opus, $4 / $20, Anthropic's recommended default	Evaluate first for agentic coding, long documents, and visual analysis
Claude Sonnet 5	Balanced tier, $2 / $10	Still the default for everyday coding, high-volume agents, and latency-sensitive chat
Claude Fable 5.1	Top tier, $10 / $50	Escalate when Opus 5.5 at high effort still misses the bar

Anthropic's own guidance is to start with Opus 5.5 for most workloads and move to Fable 5.1 only when higher-effort Opus evaluations still fall short. That does not make Opus 5 redundant on day one: a workload that is already validated on it and has not been re-tested has no evidence yet that the newer route is better for that task.

Release Does Not Guarantee Identical Availability on Every Channel

Anthropic lists Opus 5.5 on the Claude API and on AWS, Google Cloud, and Microsoft Foundry from launch day. Three details still differ by channel. Fast mode, priced at $8 / $40 per million tokens, is a research preview on the Claude API only. The older computer-use tool keeps working on Bedrock but not on the Claude API or Google Cloud. And consumer-plan quotas in the Claude apps are a separate matter from API token billing, so a model appearing in a chat picker says nothing about an API key's rate limit.

On EvoLink, the Claude Opus 5.5 route accepts the same claude-opus-5-5 ID on both the OpenAI-compatible Chat Completions endpoint and the Anthropic Messages endpoint, with the same API key you already use. The model page documents which settings the gateway converts for compatibility and which it rejects. Confirm the returned model, usage, and billed dimensions on your own account before production traffic; a released model is not a passed production gate.

What Developers Wanted, and What Still Requires Measurement

The launch answers the questions that were open a week ago: the public name, the API ID, the context and output limits, the token prices, the cache economics, the platform list, and the migration risks are all documented. It does not answer whether Opus 5.5 is the right route for your product.

The 40% cost figure, the 30% speed figure, the FrontierCode result, and "Fable 5.1 level on most work" are hypotheses until they hold on your traces. Because the default effort dropped from high to medium and the model thinks more per turn at a given level, you need a fresh effort sweep before any cost comparison means anything. Anthropic's migration guide says the same: re-run the sweep, re-evaluate model-specific prompt instructions, and test in development before switching production traffic.

Seven Production Signals Worth Monitoring

Signal	Evidence to retain	Why it matters on Opus 5.5
Route identity	Requested and returned model IDs	Fallback to a non-Fable model drops thinking blocks silently
Effort	The level actually sent, per request	Omitting it now means `medium`, not `high`
Tool behavior	Tool chosen, arguments, `strict` validation, retries	Forced tool choice is gone; `auto` does not guarantee a call
Conversation integrity	Append-only history, thinking blocks passed back unchanged	Edited history returns `400` on newer accounts
Progress output	Non-empty `thinking` blocks between tool calls	Streaming UIs go quiet at the default display setting
Cost	Input, cache writes, cache reads, output, tool fees per accepted task	The 60% cache-read cut only reaches cache-heavy workloads
Safeguards	`stop_reason`, `stop_details.category`, fallback used	New refusal categories, and one of them is not retried

Store these fields in one evaluation record. A quality win without route identity or billing evidence is incomplete, and a cheaper trace that needs more retries or human repair can still cost more per accepted task.

Who Should Test Opus 5.5 Now?

Prioritize an evaluation when your workload runs long agentic coding sessions in a real repository, produces analyst-grade documents from large evidence sets, or reads dense charts and screenshots where earlier models needed cropping tools. The cache-read cut matters most when a large stable prefix is served from cache on most turns.

Do not move because it is newer. Everyday coding, high-volume agents, and latency-sensitive chat usually belong on Sonnet 5, and a workload already validated on Opus 5 deserves a matched replay before it moves. Teams that disabled thinking on Opus 5 for latency reasons should budget for a prompt review as well as a code change, since the only remaining lever is effort.

Production Rollout Checklist on EvoLink

Gate	What to verify	Rollback condition
Identity	Requested and returned model both read `claude-opus-5-5`	Unexpected model or opaque fallback
Request shape	No `thinking: disabled`, no `budget_tokens`, no forced `tool_choice`, explicit `effort`	Any `400` on a replayed production request
Quality	Accepted-task rate holds or improves at the chosen effort	Regression in any critical task class
History	Thinking blocks survive the supported conversation flow and any fallback path	Invalid history after edits or downgrade
Cost	Cost per accepted task improves after the effort sweep	Token or review cost exceeds the Opus 5 baseline
Reliability	Error rate and p95 latency stay within the service objective	Capacity or latency breaches the gate

Start with replay, then shadow traffic, then a narrow canary. Keep model selection in configuration so a failed promotion is a config change, not a rewrite.

Evaluate Claude Opus 5.5 on EvoLink Read the Opus 5.5 vs Opus 5 upgrade decision

FAQ

When was Claude Opus 5.5 released?

Anthropic released Claude Opus 5.5 on September 22, 2026, and lists it as active with a retirement floor of September 22, 2027 on the platforms it operates.

Is Claude Opus 5.5 available through an API?

Yes. Anthropic documents it on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. On EvoLink it is a live production route; the model page has the current rates and settings.

What is the Claude Opus 5.5 model ID?

claude-opus-5-5, a fixed ID with no date suffix. EvoLink uses the same string on both the Chat Completions and Messages endpoints. On Bedrock the ID is anthropic.claude-opus-5-5.

How much does Claude Opus 5.5 cost compared with Opus 5?

Anthropic's list price is $4 / $20 per million input / output tokens against Opus 5's $5 / $25, a 20% cut. Cache reads fall from $0.50 to $0.20, a 60% cut. Anthropic's "40% cheaper to run" is a total-cost estimate on its own task mix, not a list price.

Can I turn thinking off on Claude Opus 5.5?

Not on either path. On Anthropic's native API, thinking: disabled and budget_tokens both return a 400. On EvoLink, the gateway converts those legacy settings to adaptive thinking instead of rejecting them, but they still cannot turn thinking off. Omit the field or send adaptive, and control depth with effort (low to max, default medium).

Is Opus 5.2 the same thing as Opus 5.5?

Anthropic never announced an "Opus 5.2"; the name came from community routing reports in mid-September. Opus 5.5 is the model Anthropic actually released. Whether the earlier reports described early traffic on this model is unconfirmed, so we keep the two names separate.

Does Opus 5.5 replace Claude Opus 5?

Not automatically. Opus 5 remains active with a retirement floor of July 24, 2027. Anthropic recommends starting new workloads on Opus 5.5, but a workload already validated on Opus 5 should be re-tested before it moves.

Where can I compare Opus 5.5 with Opus 5 before switching?

Use the Claude Opus 5.5 vs Claude Opus 5 upgrade guide for the cost model, compatibility risks, and canary plan.

Sources

Facts last verified September 22, 2026. Performance, speed, and total-cost statements are attributed to Anthropic; EvoLink users should measure their own route and workload.

All Posts

#Claude Opus 5.5 #Claude Opus 5 #Anthropic #model release #migration