On May 6, 2026, Anthropic announced two major changes for every paid Claude Code subscriber: five-hour rate limits doubled and peak-hour throttling removed on Pro and Max. Behind it sits a compute deal with SpaceX adding more than 220,000 NVIDIA GPUs and 300 MW from the Colossus 1 data center. Here's what actually changes for your daily work.
Anthropic published the change on its official blog under the title «Higher usage limits for Claude and a compute deal with SpaceX». The three key points, with no marketing gloss:
Everything applies immediately. Nothing to update, no reinstall of Claude Code, no config change. If you open your terminal today, the new limits are already active.
Anthropic has never published exact numbers for Claude Code limits. The figures floating around come from empirical measurements by the developer community. We use them here with that caveat: useful approximations, not contracts.
| Plan | Before (5 h, Sonnet) | After | Peak hours |
|---|---|---|---|
| Pro · $20/month | ~10-45 messages | ~20-90 messages | No throttling |
| Max 5× · $100/month | 5× Pro | ~10× Pro effective | No throttling |
| Max 20× · $200/month | 200-800 prompts/session | ~400-1,600 prompts/session | No throttling |
| Team / Enterprise | Variable per seat | Doubled per seat | Unchanged |
Numbers vary with message size, model choice, and context length. A conversation with many heavy files burns quota faster than a string of short questions.
If you pay $20/month for Claude Pro, before this change you could expect 10-45 messages per 5-hour rolling window using Sonnet. After the change that range moves up to roughly 20-90 messages in the same window.
In practice, this turns Pro into a viable option for half-day work sessions on a single mid-sized project. Before, many developers hit the limit by mid-afternoon and had to wait for reset; now most can finish the workday with margin to spare, except on very intensive projects.
Two things that did not change:
The Max 5× plan ($100/month) gave roughly 5× the Pro quota. After the doubling, the net effect for Max 5× users is equivalent to having ~10× the original Pro quota in the 5-hour window.
The Max 20× plan ($200/month) is where the change is most noticeable: sessions that previously delivered 200-800 prompts now hover around 400-1,600. Combined with the removal of peak-hour throttling, this plan becomes the «don't worry about quota» option for full-time developers.
Practical recommendation: if you were hitting Pro's limit several times a week, the upgrade to Max 5× already made sense. After this change the ROI improves further, because Max 5× now performs like Max 20× did a week ago for many use cases.
The headline talks about «doubling» because it sounds clean, but the change that hits daily work hardest is the removal of peak-hour throttling on Pro and Max.
Before, Anthropic silently dropped effective limits during periods of highest demand in the United States. For users in Europe, LATAM and Asia, that often coincided with their late-afternoon work window. The result: you ran out of Claude Code at the worst possible time.
With the change, those caps disappear. Capacity is still finite, but allocation becomes more uniform across the day. Developers in European, LATAM and Asian time zones are the biggest winners: they were working during what was peak for Anthropic, and now it isn't anymore.
The third change affects whoever calls Anthropic's API directly, not Claude Code. Anthropic says it has raised Claude Opus rate limits «considerably». No specific figures published.
This matters for three groups:
If your application uses Opus through the API and you implemented retry logic with exponential backoff, it's worth reviewing the new response headers and simplifying code if the new limits are sufficient.
The limit change didn't come out of nowhere. It's enabled by a compute deal between Anthropic and SpaceX announced the same day. The data:
For context: 300 MW is comparable to the continuous electrical consumption of about 250,000 average homes. That capacity is dedicated entirely to training and serving Claude models.
In his statement, Elon Musk said he decided to lease capacity to Anthropic after spending time with its leaders and being impressed by their work to make Claude AI «good for humanity». The line is notable because Musk simultaneously runs a lawsuit against OpenAI and develops his own AI at xAI.
Anthropic has been the AI top-three company with the worst capacity-to-demand ratio throughout 2025 and early 2026. While OpenAI and Google operate own-built data centers at scale, Anthropic has been signing deal after deal with AWS, Google Cloud, and now SpaceX.
For end users this translated into two visible problems:
With 300 MW of additional capacity coming online within the first month, both should ease. The doubling of limits is the visible consequence, but the subtext is «we're exiting the compute scarcity regime». For anyone building products on Claude, that's the real story.
Knowing how much you have left avoids the frustration of hitting the limit mid-task. Three paths:
/usage command inside Claude Code: shows the approximate percentage of quota consumed in the current 5-hour window.anthropic-ratelimit-* headers tell you remaining requests, input tokens and output tokens.
Good practice: when starting a large task, run /usage first. If you've already burned 60% of your window, consider splitting the task into two sessions or saving it for after the window resets.
Anthropic introduced the weekly cap in August 2025 to throttle extremely intensive users who were eating capacity disproportionately. That cap remains active and was not doubled by this announcement.
The weekly cap works as a parallel ceiling: even if you have margin within each 5-hour window, if your total usage over the past seven days exceeds a threshold, effective limits drop.
For most users the weekly cap is invisible. It only affects a minority who use Claude Code heavily for many hours a day across consecutive days. If that's you, the May 6 change helps within each 5-hour window but doesn't lift the weekly ceiling. For those cases, the option remains Max 20× or enterprise API usage.
The SpaceX deal isn't an isolated event. Anthropic has signed similar deals with Amazon (Trainium / AWS) and Google (TPU) over the past two years. The trend line is clear: aggressively add compute capacity to avoid falling behind OpenAI.
Three things worth watching over the next six months:
If you've read this far, you have the full map of what changed, why, and how it affects your subscription. What comes next depends on your situation:
And if Pro is still tight even after the doubling, the practical recommendation is to try Max 5× for a month before jumping to Max 20×. Most full-time developers find their sweet spot there.