Anthropic is flipping a pretty important switch for power users of Claude today: fast mode for Claude Opus 4.7 is now live in research preview on the API and in Claude Code, giving developers a way to keep Opus-level intelligence while significantly cutting response times.

If you’ve been following the Claude ecosystem over the last couple of months, this move won’t feel random. Opus 4.7 is Anthropic’s current flagship model, positioned as their most capable option for complex reasoning, long-running agents, and serious coding work. It already supports a 1 million token context window, up to 128k output tokens, adaptive thinking levels, and the same toolbox of features you get across the Claude 4.x line, like tools, file handling, and strong multilingual support. Benchmarks released around launch showed a clear jump over Opus 4.6, including about a 13% lift on a 93-task coding benchmark and better performance on tough software engineering tasks that earlier models struggled with. In other words, Opus 4.7 is not just a minor patch; it is the model Anthropic expects you to reach for when you actually care about correctness and robustness in production.

Fast mode doesn’t introduce a new, smaller model. Instead, Anthropic runs the same Opus 4.7 weights under a different inference configuration designed to push more tokens per second. According to their docs, setting the speed parameter to “fast” on Opus 4.7 can deliver up to roughly 2.5x higher output tokens per second compared to standard speed, with the core capabilities and behavior unchanged. The gains are mostly about throughput once the model starts talking, not the time to first token, so you should expect long answers and agentic loops to complete materially faster, even if the initial pause before the response appears doesn’t shrink quite as dramatically. That distinction matters when you’re running workflows that generate thousands of tokens in one go, like code refactors, long-form analyses, or multi-step plans.

The tradeoff is price. Fast mode for Opus 4.7 sits firmly in premium territory at six times the standard Opus rates. For fast mode, Anthropic’s current published pricing for both Opus 4.6 and 4.7 is around $30 per million input tokens and $150 per million output tokens, compared with roughly $5 and $25 per million for regular Opus 4.7 usage. That multiplier applies across the full context window, even if you are pushing toward the high end of the 1M token limit, and it stacks with other pricing modifiers like prompt caching and data residency add-ons. In plain terms, if you flip fast mode on for a heavy agent pipeline without thinking about volume, the bill can surprise you much faster than a single call’s price might suggest.

Anthropic is treating fast mode as a beta “research preview” feature for now, and access is still gated. API customers who want to experiment with fast Opus 4.7 need to join a dedicated waitlist, which is separate from general Claude API access, so Anthropic can throttle demand and collect feedback before they roll it out more broadly. Once you’re approved, you add a speed flag to your API calls and operate under a separate set of rate limits from the standard Opus pool. If you push beyond those limits, the API answers with a 429 and a retry-after header, making it clear that fast capacity is being managed as its own lane.

For developers living inside Claude Code, the story is a bit more immediate. Anthropic’s dev account has already said that fast mode for Opus 4.7 is live there as an opt-in option and is scheduled to become the default model for fast mode on Claude Code very soon. The official docs walk you through flipping it on, and from that point, you’re effectively getting Opus 4.7’s reasoning at a noticeably higher speed for coding sessions, refactors, or code review flows. This comes after a period of experimentation where Anthropic tweaked default reasoning effort levels in Claude Code and listened to pushback from users who felt quality suffered when they tried to shave latency too aggressively. Now, fast mode is a more explicit, paid lever: if you care enough about speed to pay six times the normal rate, Anthropic gives you that path without quietly downgrading the core model.

Beyond Claude’s own interface, fast Opus 4.7 is also starting to appear in the broader ecosystem of AI tools used by developers. Anthropic has confirmed that fast mode is available in research preview on partner products like Cursor, Emergent Labs, Factory, v0, Warp, and Windsurf, all of which are leaning hard into AI-assisted coding and agent workflows. That means you may not need to touch the raw Anthropic API at all to benefit; you could be writing code in a specialized editor or running agents in a hosted platform that simply flips fast mode on under the hood for certain tasks. Given that Opus 4.7 is already exposed through multiple infrastructure providers like AWS, Google, Azure, and Anthropic’s own platform, this fast-mode option is likely to show up behind the scenes in more dev tools over the coming weeks.

So who actually needs fast Opus 4.7, especially at that premium price? The use cases Anthropic seems to be prioritizing are agentic and asynchronous workflows that are bottlenecked by response time rather than single-call cost. Think large codebase analysis, multi-stage debugging, or long-running planning tasks where each step kicks off additional work. In those scenarios, standard Opus 4.7 already does the job, but a 2x or more speed-up per output can shrink end-to-end loop times in a way that changes what’s feasible. Developers experimenting publicly have pointed out that shorter waits per loop make it more realistic to keep Opus in the loop for iterative agents instead of dropping down to a cheaper, weaker model just to reduce latency.

At the same time, there’s a real debate about whether “fast” is always better when you’re talking about premium reasoning models. Some practitioners note that getting faster responses is the easy axis; the harder part is making sure that fast mode preserves the careful calibration and reliability that makes Opus useful for agents in the first place. Other high-end models have gotten fast before, but the failure mode has often been confidently wrong answers at high throughput, which is arguably worse than a slower, more cautious model. Anthropic insists that fast mode runs the same Opus 4.7 model weights rather than a separate lightweight variant, but the community will be watching closely to see whether behavior stays consistent in real workloads under the new configuration.

There’s also a cost nuance under the surface. On paper, Anthropic has kept base Opus 4.7 pricing unchanged at $5 per million input tokens and $25 per million output tokens, which looks competitive given its benchmark performance. But Opus 4.7 has a new tokenizer and can use substantially more tokens on certain tasks, especially at higher reasoning efforts, which effectively raises the total dollars you pay per task even before you touch fast mode. Add a 6x multiplier for fast mode on top of that, and it becomes even more important to profile how many tokens your agents actually consume over a full run instead of relying on simple “per million” price comparisons.

On the infrastructure side, the broader provider landscape around Opus 4.7 helps frame what fast mode is competing with. Independent benchmarking of Opus 4.7 across providers like Amazon, Azure, Google, and Anthropic shows output speeds in the tens of tokens per second, with Amazon and Azure often coming out slightly ahead on raw throughput and Google leading on time to first token. Those numbers are for existing “max effort” adaptive reasoning runs, so fast mode on Anthropic’s own platform is an attempt to push into that same territory while keeping the full Opus capability profile. If you’re already on a cloud provider’s managed offering, you may end up comparing fast-mode Anthropic directly with their hosted Opus or competing frontier models when optimizing for both latency and cost.

Taken together, fast mode for Claude Opus 4.7 is less about a flashy new model and more about saying: if you’re willing to pay, you no longer have to choose between “smart” and “fast” within the Claude ecosystem. You get the same flagship model that lifted benchmarks and improved agentic coding, now with a dial that trades dollars for throughput in a controlled, opt-in way. For teams building serious agent workflows or large-scale coding tools, that’s a powerful new knob to have, but it’s one that will demand careful cost monitoring and real-world testing to see whether the speed gains justify the premium.