Anthropic doubled Claude Code's limits by renting 220,000 GPUs from xAI

Anthropic doubled Claude Code’s five-hour rate limits on every paid plan on May 6 and pulled the peak-hours throttle that tightened those limits during US business hours. The lowest API tier on Claude Opus also picked up a 1,500% input-tokens-per-minute and 900% output-tokens-per-minute increase, with the higher tiers moving in step.

The new capacity comes from a SpaceX deal for xAI’s Colossus 1 data center. All of it. That’s 300 megawatts and 220,000+ Nvidia GPUs, coming online inside the month.

Anthropic’s announcement frames the shift as straightforward capacity catch-up: paying customers had been hitting the five-hour limits earlier than the company wanted, and the API tier ceilings on Opus had been the rate-limit-most-complained-about line item from teams running coding agents in production. The fix is more compute, faster than the company could provision through its existing AWS and Google Cloud relationships.

What changed, exactly

Three concrete changes, all live now:

• Claude Code 5-hour limits doubled for Pro, Max, Team, and seat-based Enterprise plans, per Anthropic’s post. The free tier doesn’t move.
• Peak-hours reduction removed for Pro and Max accounts. The previous behavior shaved limits during the US working day so Anthropic could keep capacity available for Enterprise and API; that throttle is gone.
• Opus API rate limits raised across tiers. Tier 1 picks up a 1,500% increase in input tokens per minute and 900% in output tokens per minute, per 9to5Google’s read of the published rates. Higher tiers move proportionally; Anthropic’s tier table on the API docs is the source of truth for the exact new numbers.

If you’ve been running Claude Code at the edge of the five-hour cap, you now have roughly twice the runway before the cooldown kicks in. If you’ve been hitting Opus API throttles during a long agent run, the headroom on input and output throughput is the bigger structural change. Long-context agent loops that previously had to chunk and back off should now stay closer to a continuous flow.

Why now: the SpaceX/Colossus deal

The capacity has to come from somewhere, and Anthropic’s existing cloud capacity hasn’t grown fast enough to absorb the demand. The unlock is a partnership with SpaceX for the full 300 megawatts of xAI’s Colossus 1 supercomputer in Memphis. The exact line in the announcement: “all of the compute capacity from SpaceX’s Colossus 1 data center, 300 megawatts in the form of over 220,000 Nvidia GPUs, will directly improve capacity for Claude Pro and Claude Max subscribers.”

The competitive geometry is the strange part. Colossus is the data center xAI built to train Grok, the model that competes with Claude. xAI sits under SpaceX after the merger that pushed the combined entity past a $1.25 trillion valuation, which we covered alongside the Terafab fab filing. Anthropic is now paying SpaceX, which controls xAI, for the GPUs that were originally provisioned to compete with Anthropic’s flagship model.

Wired’s framing is that the AI race has reached the point where compute scarcity overrides competitive logic: a vendor with idle capacity can monetize it to a competitor faster than its own product roadmap can absorb it. Musk’s posted reaction was that he was “impressed” with the Anthropic team after meetings and “reserves the right to reclaim the compute if their AI engages in actions that harm humanity,” which is the kind of clause that exists nowhere else in the cloud-procurement playbook.

The announcement also flags potential follow-on deals for “multiple gigawatts of orbital AI compute capacity,” tying in SpaceX’s Starlink-derived satellite-data-center concept. Treat that as forward-looking aspiration, not a contract.

The “lazy Claude” reports the limit bump doesn’t fix

A second wave of quality complaints is still active and the rate-limit announcement doesn’t answer any of them. This isn’t the February-to-April set of issues Anthropic postmortemed on April 23, where a length-limit system prompt and a context handling bug were the load-bearing causes. That round was acknowledged and fixed in v2.1.116. The new wave centers on Opus 4.7, which shipped April 16, and is still open.

The recurring patterns developers are reporting on Reddit, Hacker News, and X over the past two weeks:

• Web research regressions versus Opus 4.6. Source-attribution accuracy dropping, weaker contradiction detection, lower citation specificity on the same prompts that worked a month ago.
• Multi-file edit reliability. Increased rates of partial application, files left in inconsistent states after a multi-edit run, and the model bailing into “summarize what I’d do” instead of doing it.
• Tokenizer efficiency loss. The Vibe Coding writeup measured roughly 35% more tokens charged per equivalent prompt under the new tokenizer, which compounds the cost of any agentic loop that wasn’t already retuned. The same writeup flags budget_tokens returning a 400 error on prompts that worked in 4.6, plus hidden thinking tokens by default.
• Refusals and hedging. Higher rate of “I’d recommend you do X yourself” responses on requests Opus 4.6 just executed.

Anthropic hasn’t issued a formal statement on this wave. The framing inside the May 6 capacity announcement is that more compute means more headroom, which is true but orthogonal to the quality complaints. Bigger five-hour buckets and higher API ceilings let an agent run longer; they don’t make the agent’s reasoning step deeper. If the lazy-Claude reports hold up against the next benchmark refresh, the fix is going to have to be a model-side or prompt-side update, not a capacity one.

What this means for you

If you build with Claude Code, the immediate effect is fewer interruptions on long agent runs. The five-hour limit was the most common cap heavy users hit; doubling it pushes the friction point well past most realistic single-session workflows. Pair that with the peak-hours removal and the daytime throttle that frustrated US-based teams disappears entirely. There’s no migration to do; the new ceilings apply automatically.

If you build on the Opus API, audit your retry logic. Code that was tuned to back off aggressively against the old TPM ceiling will now leave a meaningful chunk of the new headroom on the table. Tier 1 specifically goes from a hobbyist-ceiling rate to a small-team-production rate, which changes what’s worth building inside a single API key versus splitting across multiple. Check your provider dashboards for the new published TPM values before changing any concurrency settings.

If you’ve been frustrated with Opus 4.7 specifically, the rate-limit increase doesn’t help. The capacity story is decoupled from the quality story, and Anthropic hasn’t said anything yet about the May wave of regression reports. If your workflow depends on web research, multi-file edits, or token-cost-sensitive agentic loops, run the same task on Opus 4.6 (still available via the API) before scaling up Claude Code usage to fill the new headroom.

If you watch the broader AI infrastructure market, this deal is the proof point that the compute layer has commodified faster than the model layer. A frontier-lab Anthropic paying a frontier-lab competitor SpaceX/xAI for compute capacity is the kind of cross-camp transaction that wasn’t supposed to happen until the next training cycle. It happened this week, and the customer-facing effect is bigger limits today rather than a roadmap promise for next year.