Claude Rate Limits Explained: API, Chat, and Cowork Usage Caps

Anthropic doubled Claude's five-hour rate limits on May 6, 2026, powered by a compute deal with SpaceX that added over 300 megawatts of new GPU capacity and more than 220,000 NVIDIA GPUs. The capacity boost was necessary because Claude enforces a single usage pool across three surfaces that consume tokens at very different rates, and most users discover this only after getting throttled mid-task.

The three surfaces are claude.ai (web and desktop chat), Claude Code (the terminal-based coding agent), and Claude Cowork (the desktop agent that handles multi-step tasks autonomously). Each surface sends different volumes of tokens per interaction, but they all draw from the same account-level budget. A morning of heavy Claude Code usage leaves fewer tokens for an afternoon Cowork session. A long chat conversation with file attachments eats into your Claude Code quota for the evening.

Anthropic measures your budget in tokens, not messages. A short chat question might use 200 input tokens and produce 500 output tokens. A Claude Code session that reads five source files, generates edits, and verifies the results can consume 50,000 tokens in a single loop. A Cowork task that researches a topic, drafts a document, and iterates on its own feedback can burn through 200,000 tokens before you see a final result. The same "one interaction" costs very different amounts depending on which surface you use.

Your budget resets on two overlapping timers. The five-hour rolling window starts from your first interaction and replenishes five hours later. If you send your first message at 9 AM and hit the limit at 10 AM, the window resets at 2 PM (five hours from when it started, not when you stopped). A separate weekly cap limits total usage across an entire week, resetting at a fixed time assigned to your account. Hitting either limit pauses your access until the relevant timer expires.

When you hit a limit, Claude shows a message with the reset time. Your session data stays intact. No work is deleted and no extra charges are added. You wait, or you switch strategies.

Max plans carry two weekly sub-limits: one across all models and a separate, lower cap for Sonnet usage specifically. Running out of your Sonnet allocation mid-week still leaves your Opus budget available, but not the reverse. This catches users who default to Sonnet for everything and never touch Opus.

For teams running agentic workflows that might get interrupted by rate limits, persisting output to a shared workspace prevents lost progress. A platform like Fast.io keeps files accessible through its MCP server regardless of whether your Claude session is active.

Anthropic does not publish exact token quotas for consumer plans. The "5x" and "20x" in Max plan names refer to usage multipliers relative to Pro, not to specific token counts. Community testing and independent analysis fill in approximate numbers, but these shift with server-side changes and vary based on model selection, context size, and effort level.

Here is what each tier provides in practice, based on community-reported estimates after the May 2026 doubling.

Free ($0/month)

Limited access to Sonnet with no Opus. Claude Cowork is not available. Enough to evaluate Claude on small tasks, but not for sustained work of any kind.

Pro ($20/month, $17/month annual)

The baseline paid tier. Community estimates put Pro at roughly 90 chat messages per five-hour window after the May 2026 doubling. Claude Code prompts cost more per interaction because agentic loops read files and execute commands, so expect fewer coding prompts per window. Cowork is included, but a single complex Cowork task can consume what would otherwise be 5 to 20 chat messages worth of tokens. At those rates, daily Cowork users regularly exhaust Pro within an hour of focused agentic work.

Pro works well for chat-primary users and occasional Claude Code sessions. Once Cowork becomes part of your daily routine, the math breaks down quickly.

Max 5x ($100/month)

Five times Pro's quota. Independent testing suggests roughly 450 chat-equivalent messages per five-hour window. The larger budget makes Cowork practical for regular use, because the 5 to 20x consumption multiplier cuts into a much bigger pool. This is the realistic starting tier for anyone who uses Cowork more than occasionally.

Max 20x ($200/month)

Twenty times Pro. At this level, you are unlikely to hit the five-hour cap during normal work, even with heavy Cowork usage. The weekly cap becomes the binding constraint instead of the rolling window. This tier is built for developers who run Claude as their primary working environment throughout the day.

Team Standard ($25/seat/month, $20/seat annual)

Per-seat allocations comparable to Pro. Usage pools are managed at the team level with admin controls. Standard seats include Claude Code but not Cowork.

Team Premium ($100/seat/month, 5-seat minimum)

Roughly 5x the standard seat allocation. Includes Claude Cowork. Built for engineering teams where multiple members run agentic sessions concurrently.

Enterprise (custom pricing)

Negotiated quotas, SSO, admin controls, and up to 500K token context windows on some models. Seat-based Enterprise plans received the May 2026 doubling. Consumption-based Enterprise plans are billed on usage rather than caps.

The most common mistake in plan selection is underestimating Cowork's token appetite. Chat users can stay on Pro indefinitely. The moment you find yourself hitting the five-hour cap more than twice a week, Max 5x at $80/month more than Pro almost certainly pays for itself in recovered work time.

Anthropic's own documentation states that "working on tasks with Cowork consumes more of your usage allocation than chatting with Claude" because "complex, multi-step tasks are compute-intensive and require more tokens to execute." In practice, the difference is substantial.

A chat interaction follows a simple pattern: you ask a question, Claude responds. One prompt in, one response out. Even with a long system prompt and conversation history, a typical exchange might total 3,000 to 5,000 tokens.

A Cowork task follows a different pattern entirely. When you ask Claude Cowork to "write a project proposal based on these five documents," the agent reads each document (potentially 50,000+ input tokens), cross-references them, drafts the proposal, reviews its own output for consistency, makes edits, and produces a final version. Each step is a full API call with growing context. A single task can consume 100,000 to 200,000 tokens or more.

The 5 to 20x consumption range depends on task complexity. Simple file operations like renaming and organizing sit at the low end. Research-heavy tasks with multiple document reads and iterative writing push toward the high end. A morning of five Cowork tasks can use what would take two days of pure chat usage.

Several strategies help you stretch your allocation without upgrading plans.

Use the right surface for each task. Save Cowork for tasks that genuinely need autonomous multi-step execution: research across multiple files, document drafting with iteration, complex code refactors. Use standard chat for questions that need a single response. This split alone can triple your effective daily usage.

Start fresh sessions for new topics. A Cowork session that drafted a marketing proposal carries that entire context forward when you switch to analyzing a spreadsheet. The accumulated context inflates every subsequent interaction. Close the task and start a new one.

Choose models deliberately. Sonnet handles the majority of tasks at lower token cost than Opus. Reserve Opus for complex reasoning that Sonnet fails at on the first attempt. You can adjust this per-task in Cowork settings.

Persist your output externally. When a rate limit hits mid-workflow, you need your intermediate results to survive until the next window. Saving Cowork output to a Fast.io workspace means your files stay indexed and searchable through Intelligence Mode while you wait for the reset. Agents can read and write through the Fast.io MCP server, so resuming in a new session picks up where the previous one left off.

Monitor your consumption. Claude tracks usage in Settings > Usage. Check it before starting an intensive Cowork session, especially if you have been chatting or coding earlier in the day.

Keep Agent Output Accessible Between Rate Limit Windows

Fast.io gives your Claude agents persistent storage with MCP access. Files stay indexed and searchable in a shared workspace while you wait for limits to reset. Starts with a 14-day free trial.

Start 14-Day Trial

The API rate limit system operates independently from consumer subscription plans. You do not need a Pro or Max subscription to use the API, and API usage does not count against your chat, Code, or Cowork quotas.

Anthropic organizes API access into three tiers (Start, Build, Scale) plus a Custom tier for enterprise agreements. Each tier sets limits across three dimensions: requests per minute (RPM), input tokens per minute (ITPM), and output tokens per minute (OTPM).

Current limits for Claude Opus 4.x models, which apply as a combined cap across Opus 4.8, 4.7, 4.6, and 4.5:

Start tier

1,000 RPM, 2,000,000 ITPM, 400,000 OTPM. Monthly spend cap: $500.

Build tier

5,000 RPM, 5,000,000 ITPM, 1,000,000 OTPM. Monthly spend cap: $1,000.

Scale tier

10,000 RPM, 10,000,000 ITPM, 2,000,000 OTPM. Monthly spend cap: $200,000.

Claude Sonnet 4.x and Haiku 4.5 share the same RPM, ITPM, and OTPM limits as Opus at each tier. Fable 5 has lower limits: 1,000 RPM, 500,000 ITPM, and 100,000 OTPM at the Start tier.

The API uses the token bucket algorithm, which means capacity replenishes continuously rather than resetting at fixed intervals. Short bursts of requests can still trigger 429 errors even if your average usage is within limits.

One significant advantage for API users: cached input tokens do not count toward ITPM limits for most models. With an 80% cache hit rate and a 2,000,000 ITPM limit, your effective throughput is 10,000,000 total input tokens per minute because the 8,000,000 cached tokens are free. Prompt caching for system instructions, tool definitions, and conversation history dramatically increases real-world throughput without requiring a tier upgrade.

For agent developers building on the API, you can monitor rate limit usage through response headers (anthropic-ratelimit-tokens-remaining, anthropic-ratelimit-tokens-reset) and programmatically through the Rate Limits API. Set workspace-level limits in the Claude Console to prevent one project from consuming your entire organization's quota.

If your agents need persistent file storage between API calls, Fast.io provides MCP access at /mcp so agents can read and write workspace files through a single endpoint. Plans start with a 14-day free trial, with monthly credits and storage scaled to each tier.

Multiple rate limit changes landed in rapid succession during spring 2026. Understanding the timeline helps you separate permanent improvements from temporary promotions that may revert.

May 6, 2026: The SpaceX capacity expansion. Anthropic doubled five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans. The company also removed the peak-hours throttle that had previously reduced Pro and Max quotas by roughly half during weekday mornings (5 to 11 AM Pacific). Both changes were backed by a compute partnership with SpaceX that added over 300 megawatts of capacity and more than 220,000 NVIDIA GPUs through SpaceX's Colossus 1 data center. API rate limits for Claude Opus models also increased. These changes appear to be permanent.

May 13, 2026: Weekly limit boost. Weekly caps increased by 50% across all paid plans. This increase is temporary, running through July 13, 2026, unless Anthropic extends it. On May 15, all usage counters were reset to zero, giving every user a fresh start under the new limits.

June 5 through July 5, 2026: Cowork 2x promotion. The five-hour usage limit for Claude Cowork specifically is doubled during this window. The boost is automatic with no opt-in required. It applies only to Cowork, not to chat, Claude Code, or Claude Desktop. Weekly limits are unchanged. Eligible plans: Pro, Max, Team, and legacy seat-based Enterprise. Free accounts and consumption-based Enterprise plans are excluded.

June 15, 2026: Agent SDK credit separation paused. Anthropic had announced that Agent SDK workloads, the claude -p command, GitHub Actions, and third-party apps authenticating through the Agent SDK would move to a separate monthly credit pool ($20 for Pro, $100 for Max 5x, $200 for Max 20x) instead of drawing from subscription limits. On the planned effective date, Anthropic paused the change, stating it was reworking the approach "to better support how users build with Claude subscriptions." Those surfaces continue drawing from your subscription limits exactly as before.

Current state as of June 27, 2026: Doubled base limits (permanent so far), 50% weekly boost (through July 13), Cowork 2x promotion (through July 5), no Agent SDK credit separation. Both temporary boosts may revert to standard levels after their windows close. If you are evaluating whether Max 5x or Max 20x fits your usage pattern, base your decision on the permanent doubled limits rather than the promotional numbers.