Anthropic is reportedly close to releasing its next major AI model, and the most important detail is not only that it may be more capable. It is that the company appears to be rethinking how a model should balance speed, reasoning depth, and cost for developers.
A model built around two modes
According to a report from The Information, Anthropic’s upcoming model is described as a hybrid system. The reported design would allow it to switch between “deep reasoning” and fast responses, rather than forcing users to treat those behaviors as separate categories.
That matters because the source describes deep reasoning as more demanding: it consumes more computing. A model that can answer quickly when a task is simple, but spend more effort when a task requires deeper analysis, would give developers a more flexible way to match the model’s behavior to the job.
The report says the model could arrive within weeks. Anthropic has not, in the source article, provided a full technical breakdown of how the switching works, so the important point is the direction: a single major AI model that can reportedly adapt its response style.
The sliding scale is the developer story
The model will reportedly be introduced with a “sliding scale” that lets developers control costs. The source links that directly to the fact that deeper reasoning consumes more computing.
For developers, that framing is practical. Not every request needs the same amount of work from an AI model. Some prompts may need a fast answer, while others may need more careful reasoning over code, documents, or business material.
The reported scale suggests Anthropic wants to make that tradeoff visible. Instead of presenting reasoning as a fixed feature, the company may let developers choose how much model effort is appropriate for a task.
That could be especially relevant in settings where many requests are handled repeatedly. A developer using the model would need to think not only about accuracy or speed, but also about when the extra computing behind deep reasoning is worth the cost.
Coding appears to be a major benchmark
The Information report says Anthropic’s model outperforms OpenAI’s o3-mini-high “reasoning” model on some programming tasks. The source does not list those tasks, so the claim should be read narrowly: some programming tasks, not every coding scenario.
Still, programming is clearly part of the model’s reported pitch. The model is also said to excel at analyzing large codebases, which is a different challenge from answering an isolated coding question.
Large codebases require a model to keep track of structure, dependencies, intent, and context. The source does not explain the exact benchmarks behind that claim, but it does show where Anthropic appears to be aiming: work that involves more than short prompts and simple completions.
The report also says the model performs well on other business-related benchmarks. Again, the source does not identify those benchmarks, so the safest conclusion is that Anthropic’s next model is being positioned for practical developer and business workloads, not only general chat.
Anthropic’s view of reasoning models
Anthropic CEO Dario Amodei hinted at new models in an interview on Monday with TechCrunch’s Romain Dillet. In that interview, Amodei said the company is focused on building its own take on reasoning models that are better differentiated.
He also questioned the idea that normal models and reasoning models should be treated as separate things. That comment lines up with the reported hybrid approach: rather than separating fast models from reasoning models, Anthropic may be moving toward a model that can cover both behaviors.
This is the central idea behind the reported release. Anthropic is not only trying to ship a stronger model. It is reportedly trying to make the model’s reasoning depth more adjustable and more closely tied to developer cost choices.
What to watch next
The source article leaves several details open. It does not provide the model’s name, a firm release date, pricing, or a full benchmark list. It also does not say exactly how developers will use the “sliding scale” in practice.
Those missing details matter because the usefulness of the reported approach will depend on implementation. A cost-control scale has to be understandable, predictable, and tied clearly to model behavior. A hybrid model also has to make good decisions about when fast responses are enough and when deeper reasoning is needed.
For now, the clearest takeaway is that Anthropic’s next major AI model may focus on a more flexible relationship between speed, reasoning, and computing cost. If the report is accurate, the company’s next release could make the reasoning model debate less about choosing one kind of model and more about controlling how much reasoning a model applies.