Anthropic Launches Claude Opus 5.5: 40% Cheaper, 30% Faster Output
Anthropic says that on typical tasks with default settings, the model costs about 40% less to run than Opus 5 and writes out answers more than 30% faster. Existing prompts should mostly carry over, but API settings need a second look.
Anthropic released Claude Opus 5.5 on September 22, along with prompting guidance for existing Opus 5 users. A prompt is the instruction you give an AI. The new guidance covers how much effort the model puts into a task, progress updates on long-running jobs, handling complex visual material, and refusals triggered by safety guardrails.
Will your existing prompts still work?
For the most part, yes, according to Anthropic. What has changed is how the model reasons through a problem before producing an answer. Opus 5.5 keeps its ability to scale thinking up or down to match the task permanently switched on. In Opus 5 you could turn that feature off; in the new model you cannot.
The default effort level has also dropped from high to medium. Effort level is the dial users adjust to balance how much work the model does against speed and cost. Developers who connect Claude to their own products should check whether their existing settings still behave the same way on the new model. Separate from rewriting prompts, it is worth auditing the parameters you send along with each model call.
How much have cost and speed changed?
Anthropic says that in tests on typical tasks using default settings, Opus 5.5 cost about 40% less to run than Opus 5. Output speed, meaning how quickly the model writes out its answer, improved by more than 30%.
For developers building on the model, base pricing is $4 per million input tokens and $20 per million output tokens, each 20% below Opus 5. A token is a unit for counting fragments of text. Input, which is what you send, and output, which is what the model writes back, are billed at different rates. That means actual savings will vary by product, depending on how much text flows in each direction.
On performance, Anthropic says Opus 5.5 matches its own Claude Fable 5.1 on most tasks. For coding work, the company also published a cost and performance comparison against GPT-6 Astra. If your use case is not coding, it is worth benchmarking the model on the specific jobs you plan to hand it.
Who can use it in Korea?
Claude Pro, Max, Team, and Enterprise subscribers, along with developer customers, have had access to Opus 5.5 since launch. It is also available through Amazon Web Services (AWS), Google Cloud, and Microsoft Foundry. Korea is among Anthropic’s officially supported regions, and both Claude on the web and the desktop app support Korean.
Before putting the model to work, it is also worth considering what kind of requests you will be making. Some requests in biology and cybersecurity may be refused by safety guardrails. For biology research, vetted institutions have been able to apply to the Life Sciences Verification Program since launch. For cybersecurity, Anthropic says it will extend coverage of its Cyber Verification Program to Opus 5.5 within the next few weeks. Whether the cost and speed gains the company describes hold up on real work done in Korean is something users will have to judge for themselves in practice.