Claude Opus 5.5: A stronger model that costs less to run
Claude Opus 5.5
Anthropic’s new model for coding, agents, and other complex work.
token
A small unit of text used to measure AI use and price.
prompt injection
An attempt to make an AI ignore its rules through hidden instructions.
What happened
Anthropic, the company behind Claude, announced Claude Opus 5.5 on September 22. Its official announcement calls it the first model in a new Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work. It also says Opus 5.5 costs 40% less to run than Opus 5. It is available now through Claude, Amazon Web Services, Google Cloud, and Microsoft Azure.
Why this matters
The notable part is the combination of capability, speed, and price. Opus 5.5 uses less computing power than Opus 5, according to Anthropic. It produces answers more than 30% faster. Input costs $4 per million tokens. Output costs $20 per million tokens. Cache reads cost $0.20 per million tokens.
Long AI jobs involve many steps. An agent may read files, use tools, check results, and try again. Lower costs can make such work easier to afford. This is a likely effect, not an independent result. Real savings will depend on the task and the way people use the model.
What Anthropic says it tested
Early testers reported large gains on difficult work. One tester completed a 680,000-line code migration in less than one day. Another test asked the model to reduce loading times across a web app. Opus 5.5 succeeded 39 times out of 40. Anthropic says Opus 5 made smaller changes that sometimes changed the app’s behavior. A separate tester asked Claude models to build a game from one prompt. Opus 5.5 received the highest score for graphics and polish.
These examples show the kind of work Anthropic wants to target. They do not prove that every customer will see the same results. Many examples come from early testers or company-run evaluations.
Safety is part of the release
Anthropic says external evaluators, including Frontier Design and METR, tested Opus 5.5 before release. Its automated behavioral audit covered nearly 2,000 simulated scenarios. Anthropic says Opus 5.5 was its strongest model on that audit. It was less likely to take hard-to-reverse actions or leave its assigned boundaries. It was also more resistant to prompt injection than Opus 5.
The company is using safeguards similar to Claude Fable 5.1 for biology and cybersecurity. Verified organizations can apply for special access. Anthropic also launched the model with preserved thinking. This limits attempts to edit Claude’s earlier context and extract its reasoning through the API.
What remains unclear
Anthropic itself says evaluations cannot catch every failure before release. It also says Opus 5.5 may sometimes recognize that it is being tested. That makes real-world behavior difficult to predict. Benchmark scores are useful clues, but they are not the same as long-term workplace results.
The 40% cost reduction applies to typical workloads. It is not a promise for every task. Independent users will need to check speed, accuracy, safety, and total cost over time.
What to watch next
The next test will happen in ordinary work. Developers and companies will show whether Opus 5.5 can complete long jobs with less supervision. Its biology and cybersecurity access programs will also matter. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks. The larger question is whether frontier AI can become cheaper without making safety harder to manage.
Claude Opus 5.5 is stronger and cheaper to run
📰 Full story: Claude Opus 5.5: A stronger model that costs less to run
Anthropic launched a new Claude model. It says the model can handle long jobs with less time and money.
Anthropic
The company that makes Claude.
Claude Opus 5.5
A new Claude model for difficult computer work.
prompt injection
Hidden instructions that try to change an AI’s rules.
💡 The gist
- Anthropic launched Claude Opus 5.5.
- It says the model is stronger and cheaper to run.
- Safety tests improved, but they cannot promise perfect safety.
Anthropic is the company that makes Claude. It announced Claude Opus 5.5 on September 22. The official announcement says it starts a new Claude 5.5 family.
Anthropic compares Opus 5.5 with Claude Fable 5.1. It says both models perform at a similar level on most work. It also says Opus 5.5 costs 40% less to run than Opus 5. The model is more than 30% faster at producing answers.
Why does cost matter? Some AI jobs take many steps. The AI may read files, use tools, and check its work. Each step can use money. Lower costs may make long jobs easier to try. This is a possible benefit. It is not a promise for every job.
Anthropic shared several early test results. One tester moved 680,000 lines of code in less than one day. Another tester asked the model to improve a web app’s loading time. Opus 5.5 succeeded 39 times out of 40. These results came from tests reported by Anthropic. Other users may see different results.
The model also received safety tests. Anthropic tested it in nearly 2,000 made-up situations. It says Opus 5.5 was less likely to take hard-to-undo actions. It was also more resistant to prompt injection. Prompt injection means hiding instructions that try to change an AI’s rules.
Opus 5.5 is available through Claude and several cloud services. Anthropic plans to release Claude Sonnet 5.5 and Claude Haiku 5.5 soon. People will need to watch real use. They must check both the model’s work and its safety.
A new Claude can help with long computer jobs
📰 Full story: Claude Opus 5.5: A stronger model that costs less to run
Anthropic, the company that makes Claude, announced a new computer helper.
Anthropic
The company that makes Claude.
Claude Opus 5.5
The name of the new computer helper.
AI
A computer system that can help with tasks.
Claude Opus 5.5 is the helper’s name.
Anthropic’s announcement says it can do many kinds of work.
It can write and fix computer programs.
It can help with long jobs.
Anthropic says it costs less to run than Opus 5.
That can help people use it for longer tasks.
People tested the new helper before its release.
One test gave it a very large computer job.
The helper finished the job in less than a day.
People also checked its safety.
Anthropic says it followed its rules better than older models.
That does not mean it is always safe.
A test cannot show every possible mistake.
People still need to watch its work.
Claude Opus 5.5 is available now.