A coding agent can finish a patch quickly and still leave a team with hours of review, repair and reruns. That is the test behind Anthropic's new Claude Opus 5.5: does it deliver more accepted work for the time and money spent?
Anthropic introduced Claude Opus 5.5 on 22 September 2026. The company says the model improves on Opus 5 in demanding coding and professional work while costing about 40% less to run on its typical workloads. That is Anthropic's estimate, not a saving every customer should expect. The more useful question for a team is whether its own tasks finish with fewer corrections and a lower total bill.
What launched with Claude Opus 5.5?
Opus 5.5 is the first model in Anthropic's Claude 5.5 family. In its launch announcement, Anthropic says it is available across its platforms and names Amazon Web Services, Google Cloud and Microsoft Azure. Developers using the Claude Platform can select the model ID claude-opus-5-5. Availability in a particular account, region or cloud setup should still be checked before a migration.
Anthropic positions the model for sprawling jobs: codebase migrations, audits, agent tasks and professional research. It reports stronger results than Opus 5 in its tests and says Opus 5.5 performs at the level of Claude Fable 5.1 on most work. Those are useful claims to investigate, but they do not establish that Opus 5.5 is the best choice for every repository or document workflow.
| At launch | Claude Opus 5.5 | Why a reader should care |
|---|---|---|
| Release | 22 September 2026 | This is a new model decision, separate from the older Opus 4.8 and Opus 5 stories. |
| Model ID | claude-opus-5-5 on the Claude API |
Existing integrations need an explicit model change. |
| Intended work | Long-running coding agents and knowledge work | The strongest case for testing it is a task with several steps and a checkable result. |
| Context and output | 1 million-token context window; up to 128,000 output tokens | These are technical ceilings, not promises that a very long task will succeed. |
| Thinking | Adaptive thinking is always on; default effort is medium |
Developers should retest latency and cost rather than carrying Opus 5 settings across. |
These specifications come from Anthropic's Opus 5.5 model documentation. The context figure is a capacity limit, not a quality score. A team should still test whether the model finds the right evidence and keeps track of the relevant parts of a large project.
There is a less obvious launch detail in Anthropic's safeguards and access notes. Some cybersecurity requests will be routed to Opus 4.8 under Opus 5.5's safeguards, while broader trusted access for verified practitioners is being expanded. If a team's work includes security research, it should test the actual task path and access rules, not assume every request will run on the selected model.
Does Claude Opus 5.5 really cost 40% less?
Anthropic's standard API list prices fell across input, output and cache reads. The comparison below is in US dollars per million tokens, as listed in Anthropic's launch pricing.
| Standard API usage | Opus 5 | Opus 5.5 | Change |
|---|---|---|---|
| Input tokens | US$5.00 | US$4.00 | 20% lower list price |
| Output tokens | US$25.00 | US$20.00 | 20% lower list price |
| Cache reads | US$0.50 | US$0.20 | 60% lower list price |
Anthropic's model documentation also lists five-minute cache writes at US$5 and one-hour cache writes at US$8 per million tokens for Opus 5.5. Batch processing has a separate discount. Fast mode is a research preview on the Claude API with separate access and rates. A workload that uses these features should be priced from its actual usage pattern, not from the three-row table alone.
For a simple illustration, 100,000 input tokens and 25,000 output tokens would cost US$1.125 at Opus 5 list rates or US$0.90 at Opus 5.5 list rates. That is a 20% saving for the same token counts. It excludes caching, tools, retries, taxes and any plan or cloud charges. The arithmetic is deliberately modest: it shows why the headline 40% saving cannot be inferred from list prices alone.
The 40% figure combines lower token prices with Anthropic's claim that Opus 5.5 uses fewer tokens to finish typical jobs. That distinction matters. If your agent makes many tool calls, retries a failed approach or asks a developer to repair its output, the cost of a successful task can move very differently from the price of a single token.
For a business, the useful measure is cost per accepted result. Record the model charges, elapsed time, review time, retries and whether the final output met the same acceptance criteria. A cheaper first attempt that needs substantial repair may cost more overall. A more expensive run that produces a clean, verifiable result may be worth it.
What should readers make of Anthropic's performance claims?
The launch post presents coding, agent and professional-work evaluations, as well as stories from early customers. These provide context about where Anthropic believes the model improved. They are not a controlled trial of your codebase, approval process or data.
Anthropic offers more detail than a single headline score, but the measures answer different questions:
| Anthropic-reported result | What it suggests | What it cannot establish |
|---|---|---|
| CursorBench 4.0: 52.5% for Opus 5.5 at medium effort, versus 46.6% for Opus 5 at max effort | Better performance in Anthropic's reported coding-agent setup. | That the difference will hold in your repository; the effort settings also differ. |
| Web-app load-time task: Opus 5.5 succeeded in 39 of 40 runs | It handled Anthropic's specific optimisation test consistently. | That it will improve the performance of your app without regressions. |
| Research report test: 16 of 18 Opus 5.5 reports met Anthropic's source-accuracy bar | A promising result on a constrained research task. | That every figure in an unsupervised business report can be trusted. |
| Early-customer accounts and behavioural audit | Some testers saw gains and Anthropic reports stronger safety results. | A typical customer outcome or the elimination of unsafe agent actions. |
Benchmark results depend on the surrounding agent: tools, instructions, effort setting, budget and permission boundaries. Anthropic's methodology notes describe differing effort settings and safeguards in some comparisons. A score is a lead for investigation, not an upgrade decision by itself.
Opus 5.5 may be especially interesting to teams whose current agent handles large jobs but loses time in repeated edits or review. It is less urgent for a team that mainly asks short questions or drafts routine text and already gets acceptable results at lower cost.
Claude Opus 5.5 migration: what could break?
Changing a model ID is the small part of an API migration. Anthropic's migration guidance identifies changes that can affect existing applications:
| Check before switching | Practical consequence |
|---|---|
| Thinking cannot be turned off | Remove or revise requests that assume an off switch; use effort controls to tune depth. |
| Forced tool choice returns an error | Retest workflows that require a particular tool call. |
| Thinking blocks are tied to the model and conversation | Check code that stores, replays or edits previous responses. |
| An older computer-use tool version is not accepted on the Claude API and Google Cloud | Update tool definitions before switching those integrations. |
The default effort also changes from Opus 5's high to Opus 5.5's medium, according to Anthropic's release notes. That means a comparison at each model's default is a product-experience comparison, while a comparison at the same effort setting asks a different question. Record which one you are testing. Check streaming progress and refusal handling too if your application exposes them to users.
How to test an Opus 5.5 upgrade on your own work
Choose one bounded task that recently caused friction, such as fixing a bug across several files or updating a documented workflow. Start Opus 5 and Opus 5.5 from the same project state, give them the same instructions and tools, set an explicit effort level, and define a clear stopping condition. This is a proposed trial, not a test we have run for this article.
Record four things:
- Quality: Did the result pass the same tests and human review?
- Total effort: How much time did the model and reviewer spend, including corrections?
- Total cost: What was charged for the full task, including retries and tool use?
- Behaviour: Did the agent respect instructions, report uncertainty and stop when it needed approval?
Repeat on a second, different task before deciding. Our guide to AI evals explains why a small set of realistic tasks is more useful than a single impressive demo. For coding agents, the agent harness explainer also helps separate the model's contribution from the tools and controls around it.
For example, a team migrating a small service could define success as passing the existing tests, preserving the documented API and producing a change a reviewer accepts without a second patch. The trial sheet would show the starting commit, model, effort, tool permissions, elapsed time, token and cache charges, number of reruns, test result and review time. This example is a proposed measurement plan, not a reported result.
The older Opus 4.8 launch article remains useful as a dated record of that release. Opus 5.5 is a new decision, with new pricing, safeguards and performance claims. Teams should compare the options they can actually use today.
Claude Opus 5.5 frequently asked questions
Is Claude Opus 5.5 the same model as Claude Fable 5.1?
No. They are separate Claude models. In its announcement, Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work. That is a vendor comparison, not a claim that the models have identical capabilities, safeguards or access rules.
Are Opus 5.5 API tokens 40% cheaper than Opus 5 tokens?
No. Anthropic's pricing explanation shows standard input and output list prices 20% lower than Opus 5. Its roughly 40% cost claim refers to typical complete workloads in its tests, where it says Opus 5.5 also used fewer tokens. Your result will depend on the task, caching, retries and tool use.
Can I replace the Opus 5 model ID and leave the rest of my integration alone?
Do not assume so. Anthropic's migration guidance documents breaking changes involving always-on thinking, forced tool choice, conversation-linked thinking blocks and an older computer-use tool version on the Claude API and Google Cloud. Check those paths, then run your own tests before switching production traffic.
Who should consider Claude Opus 5.5 now?
Technical teams running long coding or agent tasks have the clearest reason to trial it. They can measure whether fewer failed steps and cleaner reviewed output justify a switch. Business teams using Claude for research or analysis can run the same kind of matched test on a report with checkable facts and a known review standard.
Do not change a production workflow on the strength of a launch chart alone. Start with one task, retain human approval for consequential actions and check the live pricing and access terms in your chosen product. Opus 5.5's strongest practical claim is that better work may now cost less to complete. A team can find out by measuring the completed work, not just the model's first answer.
