NEW DELHI, September 25, 2026, 5:08 PM IST — Anthropic has released Claude Opus 5.5 with lower token prices, cheaper prompt-cache reads and broader cloud availability, while introducing migration changes that platform teams need to test before switching production agents.
The model became available on September 22 through Anthropic’s API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Anthropic positions it for long-running coding agents and professional knowledge work. The headline economics are notable: list prices fall to $4 per million input tokens and $20 per million output tokens, both 20% below Claude Opus 5, while cache reads drop 60% to $0.20 per million tokens.
For developers and platform engineers, however, this is not a drop-in price update. Anthropic’s migration documentation lists four breaking changes, including always-on adaptive thinking, errors when tool choice is forced, model-specific thinking blocks and the removal of an older computer-use tool version on the Claude API and Google Cloud. Teams running tool-heavy agents should treat the upgrade as a controlled rollout rather than simply changing a model name.
What Anthropic confirmed
Anthropic says Opus 5.5 uses fewer tokens per task and requires less compute to serve than Opus 5. At default settings, the company estimates that typical workloads will cost about 40% less overall. That figure is a workload estimate, not a 40% cut to every rate: standard input and output token prices are each 20% lower, while the larger cache-read discount can reduce costs further for long-running coding sessions that repeatedly reuse a large prompt.
The model retains a one-million-token context window and supports up to 128,000 output tokens. A beta option for the Message Batches API raises maximum output to 300,000 tokens, and batch processing carries a 50% discount on standard input and output pricing. A separate fast mode can run at up to 2.5 times the standard speed, according to Anthropic, but doubles the standard input and output rates to $8 and $40 per million tokens.
Anthropic also reports more than 30% faster output generation than Opus 5. Its published evaluations show gains on agentic coding, terminal use and knowledge-work tests. Those numbers should be read as vendor-reported results: Anthropic itself cautions that narrow benchmark margins are becoming less reliable indicators of real-world differences, and some comparisons use different effort settings or results reported by other model providers.
Cloud availability widens the deployment choices
AWS says Opus 5.5 is available through Amazon Bedrock and Claude Platform on AWS. Bedrock access includes geographic inference profiles for the United States, Europe, Australia, Japan and global routing, subject to regional availability. AWS also says its Bedrock route supports zero data retention by default and works with managed features such as Guardrails and Knowledge Bases.
Google Cloud lists the model as generally available in its Gemini Enterprise Agent Platform, with United States and Europe multi-region endpoints plus a global endpoint. Its documentation confirms support for computer use, web search, batch predictions, prompt caching, function calling and the memory tool. Anthropic’s own model directory also lists Microsoft Foundry, giving enterprises several procurement and residency paths for the same model family.
That flexibility does not mean the platforms are operationally identical. Region support, quotas, data-handling terms, model identifiers and surrounding guardrail services differ. Platform teams should compare those controls with their existing LLMOps requirements instead of assuming portability at the API boundary.
The migration risks are more important than the model name
The most immediate change is adaptive thinking. It is always enabled in Opus 5.5, with an effort parameter controlling how much reasoning the model applies. Systems that previously disabled thinking to hold latency or cost steady will need new budgets and service-level tests.
Tool orchestration also needs regression coverage. Anthropic warns that forcing tool use now returns an error, while thinking blocks are tied to the model and conversation. Mid-session model switching or reconstructing histories can therefore fail in ways that older integrations did not. On the Claude API and Google Cloud, clients using the older computer-use tool version must migrate to a supported version.
A practical rollout should pin the model version where the provider permits it, replay representative agent traces, and compare task completion, cache-hit rates, tool-call failures, latency percentiles and total cost per successful task. The cheapest token is not useful if a workflow needs more retries or human review. Conversely, the lower cache-read rate could materially improve the economics of repository-scale coding agents and retrieval-heavy workflows that repeatedly reuse instructions and context. Teams designing those systems can also revisit how they structure context with retrieval-augmented generation.
What remains uncertain
Anthropic’s 40% typical-workload saving depends on token use and caching patterns, so actual savings will vary. The company’s strongest capability and efficiency claims also await broader independent testing across production codebases, languages and cloud configurations.
The release nonetheless sharpens a clear market trend: frontier-model competition is shifting from benchmark scores alone toward cost per completed task, long-session reliability and operational controls. For DevOps and platform teams, Opus 5.5 is best evaluated as a new runtime dependency with attractive economics, not as a transparent version upgrade.
Sources
Primary details come from Anthropic’s launch announcement and model and migration documentation. Cloud availability was checked against AWS’s launch guidance and Google Cloud’s model documentation. Independent launch context came from TechCrunch.
