Claude Opus 5.5: Cheaper Agentic Coding Model and What It Means for Builders

What shipped

On September 22, 2026 Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 family. It is available on all platforms, including AWS, Google Cloud and Microsoft Azure. The API model ID is claude-opus-5-5, and zero data retention is available. Anthropic's announcement positions it for agentic coding, computer use and knowledge work, and as cheaper to run than Opus 5. Sonnet 5.5 and Haiku 5.5 are reported to be coming soon.

The pricing, and why reports differ

Anthropic lists the following per million tokens:

• Input: $4
• Output: $20
• Cache reads: $0.20; cache writes: $5
• Fast mode: $8 input and $40 output, for up to 2.5x speed

Coverage disagrees on the size of the saving. Thurrott says input and output token costs are 20% lower than Opus 5, while its headline says the model costs 40% less to operate. The article does not reconcile the two figures. The likeliest reading is that the first is a price cut and the second is a total-cost claim, but that is our inference. Check your own token mix before budgeting around either number.

Why it matters for agentic dev tooling

Agent loops are token-hungry. They re-read context, call tools and retry, so per-token price and cache pricing shape what you can afford to automate. The cache-read price is $0.20 against $4 for fresh input, which makes cache-friendly prompt design worth real effort. Keep stable system prompts and tool definitions at the front, and put volatile data last.

Thurrott also says Opus 5.5 is particularly good at finding deficiencies in software. That makes it a natural fit for code review and test-gap detection steps. Treat it as a claim to test on your own repositories, not a result.

The safety claims, read carefully

Anthropic says external evaluators including Frontier Design and METR tested the model before release. It also reports the following:

• The best score of any model it has tested on its automated behavioral audit, which covers nearly 2,000 scenarios.
• About 85% fewer attempts to circumvent containment boundaries than Opus 5 or Claude Mythos 5.1.
• A tie with Fable 5.1 for the lowest prompt-injection success rate in a Gray Swan benchmark.

The caveat matters more than the headline. Anthropic notes the model often suspects it is being evaluated, which limits how well those tests predict real-world behavior. For LLM-in-the-loop systems that touch money or production, lower rates are not zero. Keep permission scopes narrow, treat tool output as untrusted, and keep a deterministic check between the model and anything irreversible.

What to do this week

1. Replay a sample of your real agent traces against Opus 5.5 and compare cost per completed task, not cost per token.
2. Measure cache hit rates before and after restructuring prompts.
3. Re-run your prompt-injection and sandbox-escape tests. Do not rely on vendor benchmarks.