Anthropic launched Claude Opus 5.5 on September 22, 2026, introducing the first model in its Claude 5.5 family. The company is positioning it as a stronger option for demanding professional work, with lower list pricing than the previous Opus generation and additional safeguards aimed at high-capability use cases.
The launch is real and broadly available, but the most impressive performance and safety figures currently come from Anthropic and the evaluations it commissioned or published. Business buyers should therefore separate three things: confirmed product facts, laboratory results, and results proven inside their own workflows.
Key takeaways
- Claude Opus 5.5 launched on September 22, 2026, through Anthropic and major cloud platforms.
- API list pricing is $4 per million input tokens and $20 per million output tokens.
- Anthropic reports gains in complex reasoning, coding, agentic work and resistance to prompt-injection attacks; those claims still need task-specific validation.
- The sensible next step is a controlled pilot with representative work, fixed pass/fail gates and human review—not an automatic organization-wide switch.
What Anthropic announced
Anthropic describes Claude Opus 5.5 as its most capable model for complex professional tasks. The company highlights coding, research, long-running agentic work and computer-use scenarios. It also says the model is better at recognizing suspicious instructions and maintaining boundaries when untrusted content attempts to redirect it.
Reuters, The Verge and the Financial Times independently reported the launch on September 22. Their reporting confirms the product release, pricing direction and Anthropic’s emphasis on cybersecurity and safeguards. That independent reporting does not turn every benchmark on the company’s launch page into a settled real-world fact; it confirms that Anthropic made and documented those claims.
This distinction matters because a benchmark result answers a narrow question under controlled conditions. A company deploying a model has a wider question: does it complete our work accurately, consistently, securely and economically when it encounters our documents, tools, edge cases and approval process?
Claude Opus 5.5 price and availability
Anthropic lists Claude Opus 5.5 API pricing at:
- $4 per million input tokens
- $20 per million output tokens
- $0.20 per million cache-read tokens
- $5 per million cache-write tokens
Compared with the prior Opus 5 list prices of $5 for input and $25 for output, the headline input and output rates are 20% lower. That simple comparison does not guarantee a 20% lower bill. Total cost also depends on prompt length, output length, caching, retries, tool calls, review time and the share of runs that produce usable work.
A more useful operational measure is cost per accepted result:
Cost per accepted result = total model and tool cost ÷ outputs accepted after review
If one model costs less per token but needs more retries or more human correction, it may be the more expensive option in practice.
Anthropic says the model is available through its own products and API as well as Amazon Bedrock, Google Cloud and Microsoft Azure. AWS separately confirmed Bedrock availability and published regional access information. Availability can still differ by account, region, service tier and organizational policy, so buyers should verify their own console rather than relying on a launch summary.
What improved—and what still needs proof
Anthropic reports better performance across advanced coding, reasoning, agentic tasks and computer use. The company also presents results from outside evaluators, including work intended to test capability and dangerous-capability thresholds.
Those materials are useful evidence, especially when test methods and limitations are visible. They are not a substitute for a buyer’s own evaluation. Four gaps commonly appear between a launch benchmark and production use:
The task gap
A coding benchmark may not resemble a company’s legacy repository, approval rules or deployment tooling. A research benchmark may not measure citation quality against the organization’s private source set.
The context gap
Production prompts contain incomplete instructions, conflicting documents, stale policies and user mistakes. A clean test suite rarely captures all of that noise.
The control gap
A model can perform well in isolation but become risky when it receives broad tool permissions. The design of identity, access, logging and human approval can matter as much as the model.
The economics gap
Token prices are only one line in the cost model. Latency, review minutes, failed runs and incident handling also affect the business case.
The practical conclusion is not that vendor benchmarks are useless. It is that they are evidence for deciding what to test, not evidence for skipping the test.
Why safety is central to the release
Anthropic’s launch materials put unusual emphasis on safeguards. The company says Claude Opus 5.5 uses classifiers and training intended to reduce harmful assistance and resist prompt-injection attempts. Its system card provides more detail about capability evaluations, deployment decisions and known limitations.
Prompt injection is especially relevant for agents and research tools. A model may read a webpage, email, document or issue ticket containing instructions written by an untrusted party. If those instructions can override the operator’s goal, the model may expose data, misuse a tool or take an unauthorized action.
No safety percentage should be read as a universal guarantee. Attack methods change, enterprise integrations create new paths, and a model’s safe behavior can depend on the system around it. Organizations should still apply least-privilege access, allowlisted tools, data classification, action logs and human approval for irreversible steps.
The system card is the right primary source for evaluating Anthropic’s claims. Buyers should read its scope and limitations, then reproduce the most relevant tests in their own environment.
The VALUE pilot framework
The following framework turns launch interest into a decision that can be audited.
V — Verify the use case
Define one bounded job, the users who perform it and what a good result looks like. “Improve productivity” is not a testable use case. “Draft a support response grounded only in approved knowledge-base articles, with citations, in under 45 seconds” is.
A — Assess evidence and risk
List the evidence behind each important claim. Mark it as a confirmed product fact, a vendor-reported test, an independent test or an internal result. At the same time, record the data involved, possible harms and required approvals.
L — Log quality, cost and latency
Use a representative test set and log task success, serious error rate, citation accuracy, refusal correctness, latency, review minutes and cost per accepted result. Keep the same rubric when comparing models.
U — Use a controlled pilot
Start with a small user group and reversible tasks. Restrict data and tools, keep human approval in the workflow, and define a rollback path before the first live test.
E — Escalate or expand
Expand only if the pilot clears every hard safety gate and meets the quality and economics threshold. Escalate failures involving sensitive data, unsafe actions, repeated hallucinations or unexplained tool use to the appropriate security or governance owner.
Decision gate | Example evidence | Stop condition
Task quality | 100 representative cases scored by reviewers | Critical error rate exceeds the agreed maximum
Grounding | Claims trace to approved sources | Unsupported material claim appears in a high-impact output
Security | Prompt-injection and permission tests | Model can access or act beyond the test role
Economics | Cost per accepted result and review time | Total workflow cost exceeds the current process
Operations | Logs, owner, rollback and incident route | No accountable owner or rollback path exists
Who should test now—and who should wait
Teams already evaluating advanced models for coding, research or multi-step agent work have a clear reason to include Claude Opus 5.5 in a controlled comparison. The lower headline API price may also justify retesting an Opus-class model where previous economics were marginal.
Teams should wait before broad deployment if they cannot assemble a representative test set, restrict permissions, monitor outputs or name an accountable owner. A capable model does not repair a weak operating process. It can make that weakness harder to see by producing fluent output more quickly.
Regulated or safety-critical users need domain-specific validation and qualified review. The release materials are not legal, medical, security or compliance approval for a particular deployment.
Frequently asked questions
When did Claude Opus 5.5 launch?
Anthropic announced and released it on September 22, 2026.
How much does the Claude Opus 5.5 API cost?
Anthropic lists $4 per million input tokens and $20 per million output tokens, with separate cache-read and cache-write prices. Contracted and cloud-provider pricing may differ.
Is Claude Opus 5.5 safer than earlier models?
Anthropic reports stronger safeguards and prompt-injection resistance. Those results should be treated as documented launch evidence, not a guarantee for every integration. Test the actual system, permissions and data paths you plan to use.
Should a business replace its current model immediately?
No general answer fits every workflow. Run a controlled, like-for-like evaluation and compare quality, safety, latency, review effort and cost per accepted result.
