Claude Code Limits Doubled | How SpaceX Colossus 1 Changes the Game
For paid plan users who've been frustrated by Claude Code's 5-hour window stopping work mid-task, here's a breakdown of the change that took effect on May 6, 2026: limits were instantly doubled across all plans. We also cover the scale of the SpaceX Colossus 1 contract behind this, and what the Opus API bandwidth expansion means in practice for engineers who use the API heavily — from large-scale code analysis to continuous agent execution.
Starting May 6, 2026, Claude Code's 5-hour window limits were instantly doubled across all paid plans, and the Max plan had its peak-hours throttling removed entirely. No settings changes or plan upgrades are required — existing users can simply keep working under the new, higher limits during daytime hours.
The bigger practical story is the +1500% input and +900% output bandwidth expansion for the Opus API. This expands the room to use Opus without downgrading to Sonnet for large-scale code analysis, long-form document generation, and continuous agent execution — directly boosting throughput for engineers who use the API at scale.
This expansion is backed by the SpaceX Colossus 1 contract: over 220,000 GPUs and 300+ MW of compute capacity, which Anthropic plans to integrate into its own infrastructure during May 2026. Note that orbital AI compute capacity remains at the concept/expression-of-interest stage and is not a confirmed plan.
目次 (12)
- The Day the "5-Hour Wall" Disappeared — Full Details on the Rate Limit Doubling and Same-Day Rollout
- What Is Colossus 1? — The Scale and Significance of Anthropic's Infrastructure Deal with SpaceX
- What +1500% Opus API Bandwidth Means in Practice — What Changes for Heavy API Users
- Three New Managed Agents Features (Dreaming / Outcomes / Multi-Agent) — Another Major Announcement on the Same Day
- Dreaming — How Agents Self-Improve Through "Reflection" (Research Preview)
- Outcomes — Define "Success" to Improve Task Completion Rates (Public Beta)
- Multi-Agent Orchestration — Lead and Specialist Role Division (Public Beta)
- What to Check Right Away — How to Verify Your New Limits and Rethink Long-Running Task Design
- Check Your Current Rate Status with the /usage Command
- Free Your Long-Running Task Design from the "Must Finish in 5 Hours" Constraint
- Use the Max Plan's Removed Peak Throttling to Your Advantage in Scheduling
- Sources
The Day the "5-Hour Wall" Disappeared — Full Details on the Rate Limit Doubling and Same-Day Rollout
Claude Code had previously enforced a token consumption cap within a fixed time window (commonly called the "5-hour window"). Many engineers have experienced the model stopping mid-task during long code analysis sessions or large-scale refactoring jobs.
According to Anthropic's official announcement, this change doubles the rate limits within the 5-hour window across all paid plans. Here's the breakdown by plan:
| Plan | Before | After |
|---|---|---|
| Pro | Limit within 5-hour window | 2x within same window |
| Max | Throttling during peak hours | Peak throttling removed entirely |
| Team | Limit within 5-hour window | 2x within same window |
| Enterprise (seat-based) | Limit within 5-hour window | 2x within same window |
For Max plan users in particular, peak-hour throttling has been eliminated entirely. You can now work at the same throughput during Japanese daytime hours as you could at night. For long-running agent tasks, this significantly reduces the risk of hitting a limit and having execution stop partway through.
No settings changes or plan changes are needed. Users already on a paid plan will automatically benefit from the doubled limits as of May 6, 2026. Anthropic expanded backend capacity to make this instant rollout possible — and the SpaceX compute deal is what made it happen.
What Is Colossus 1? — The Scale and Significance of Anthropic's Infrastructure Deal with SpaceX
The contract that enabled this limit increase is an agreement between Anthropic and SpaceX. SpaceX's Colossus 1 is a large-scale GPU cluster built in Texas, and Anthropic has contracted for its entire compute capacity (Anthropic official announcement).
The scale is stated in concrete figures. The compute capacity secured exceeds 300 megawatts, with over 220,000 GPUs including NVIDIA H100, H200, and GB200 models. This is orders of magnitude beyond a typical data center — securing this much capacity at a single facility is an unprecedented level of scale.
As the name Colossus 1 implies, the SpaceX-developed facility was designed from the ground up for large-scale AI computation. Anthropic plans to integrate this capacity into its own infrastructure and bring it online during May 2026 (official announcement).
Anthropic's infrastructure strategy extends well beyond Colossus 1. Multiple large-scale compute agreements are running in parallel: a deal with Amazon for up to 5 gigawatts (with approximately 1 GW expected online by end of 2026), a 5-gigawatt deal with Google and Broadcom (targeting a 2027 launch), and a strategic partnership with Microsoft and NVIDIA (Azure capacity at roughly $30 billion scale). Colossus 1 is one piece of that larger puzzle, but it's the specific compute foundation directly responsible for the rate limit increase announced today.
Anthropic also notes that it has "expressed interest" in developing future orbital AI compute capacity (multi-gigawatt scale) in connection with SpaceX (official announcement). However, this is currently at the concept/expression-of-interest stage and should not be taken as a confirmed plan.
What +1500% Opus API Bandwidth Means in Practice — What Changes for Heavy API Users
Alongside the end-user rate limit doubling in Claude Code, an equally significant change has been made for API users. The API rate limits for the Claude Opus model have been dramatically expanded.
According to the official announcement, the expansion is +1500% on the input side and +900% on the output side. This means the highest-performance Opus model can now be deployed reliably for large-scale workloads. The change has the biggest impact in three areas: large-scale code analysis, long-form document generation, and continuous agent execution.
For large-scale code analysis use cases, the feasibility of batch processing that ingests an entire large codebase in one shot increases dramatically. Previously, Opus's input bandwidth was often the bottleneck that forced a downgrade to Sonnet or Haiku — now there's a much better chance you can keep using Opus throughout.
For long-form document generation, the +900% output bandwidth expansion hits directly. If you're running batches of automated document generation via the API, a roughly 10x increase in throughput means you can process significantly more volume at the same effective cost. Teams using Opus to generate large volumes of technical documentation or specifications should see substantially shorter processing times.
For continuous agent task execution, architectures that run multiple subtasks in parallel will see a reduced risk of a single task blocking on bandwidth constraints and halting the whole pipeline. For teams running agents that handle multiple subtasks over extended periods, improved stability translates directly into higher output volume.
For a comparison of Anthropic's models, see the Claude Model Comparison Guide. For cost optimization considerations, Claude Code Cost Optimization offers practical guidance.
Three New Managed Agents Features (Dreaming / Outcomes / Multi-Agent) — Another Major Announcement on the Same Day
On the same day as the rate limit increases, three new features were added to Claude Managed Agents (9to5Mac report). These tend to get overshadowed by the rate limit news, but for anyone doing agent-based development, these changes are just as significant.
Dreaming — How Agents Self-Improve Through "Reflection" (Research Preview)
Dreaming is a scheduled memory processing system. It periodically reviews past agent sessions and memory stores to surface trends — recurring mistakes, converging execution patterns, shared configuration preferences across the team — and updates the memory accordingly. The system automatically discovers long-term patterns that no single session would reveal, enabling agents to improve over time (9to5Mac). You can choose between a mode where memory updates happen automatically or one where you review changes before they're applied.
Outcomes — Define "Success" to Improve Task Completion Rates (Public Beta)
Outcomes lets you describe the success criteria an agent should aim for in rubric form. Evaluation is handled by an independent "grader" in a separate context, which prevents the agent's own reasoning biases from contaminating the assessment. If a task fails the rubric, the grader identifies the problem area and the agent retries. Internal tests reportedly showed task success rates improving by up to 10 percentage points compared to standard prompting (9to5Mac), with larger improvements on more difficult tasks.
Multi-Agent Orchestration — Lead and Specialist Role Division (Public Beta)
In multi-agent orchestration, a lead agent breaks a job into pieces and delegates them to specialist agents. Each specialist has its own model, prompt, and tools, and works in parallel on a shared filesystem while contributing to the lead agent's overall context. Because events are persisted, the lead agent can check in with and adjust other agents mid-execution. The trace functionality in Claude Console gives complete visibility into which agent did what, in what order, and why — a significant operational advantage (9to5Mac).
A deeper dive into all three features is coming in a separate article.
What to Check Right Away — How to Verify Your New Limits and Rethink Long-Running Task Design
Here's a practical rundown of what engineers should do after the May 6, 2026 rollout.
Check Your Current Rate Status with the /usage Command
In the Claude Code CLI, the /usage command shows your current consumption and remaining limit. From v2.1.118 onward (release notes), /cost and /stats have been merged into /usage, so if you're used to the old commands, check your version.
/usage
Running this displays:
- Current usage: Tokens consumed in the current 5-hour window
- Remaining in window: How much capacity is left relative to the limit
- Next reset time: When the 5-hour window resets
- Plan-specific limit: The cap set for your plan
This is the quickest way to see the doubled limit in concrete numbers. Max plan users who previously experienced throttling during peak hours should pay particular attention to how much remains during daytime hours.
Free Your Long-Running Task Design from the "Must Finish in 5 Hours" Constraint
With limits doubled, you can reconsider task breakdowns that were designed around the assumption of needing to complete within 5 hours. For example, a full codebase refactoring plan that takes 10 hours no longer needs to be artificially fragmented — there's now a real chance you can submit larger units of work in a single session.
For team leads, this is also a good time to revisit expected completion times (SLAs) for agent-based tasks using Claude Code. Coarser task granularity is now more viable without hitting limits, which could reduce design overhead while increasing output volume.
Use the Max Plan's Removed Peak Throttling to Your Advantage in Scheduling
Teams on the Max plan can now effectively schedule heavy tasks during business hours that they previously had to push to early morning or late night. Moving large-scale analysis or batch processing into normal work hours makes it easier to align review cycles with the whole team.
For a full overview of Claude Code, see the Claude Code Complete Guide. For cost-conscious usage strategies, Claude Code Cost Optimization covers the details.
Sources
- Anthropic official announcement "Higher limits for Claude and a compute deal with SpaceX" https://www.anthropic.com/news/higher-limits-spacex
- Claude Code release notes (v2.1.118–v2.1.132) https://github.com/anthropics/claude-code/releases
- 9to5Mac "Anthropic updates Claude Managed Agents with three new features" https://9to5mac.com/2026/05/07/anthropic-updates-claude-managed-agents-with-three-new-features/
- Bloomberg "Anthropic Is Making Its Claude Chatbot More Appealing to Consumers" https://www.bloomberg.com/news/articles/2026-05-07/anthropic-is-making-claude-chatbot-more-appealing-to-consumers
