Claude Code usage limits: 5-hour and weekly caps explained

How Claude's usage limits actually work

Claude Code and the Claude apps run on one shared usage system with two layers stacked on top of each other. A five-hour session window tracks short-term activity, and its clock starts on your first prompt rather than a fixed time of day, so a message sent at 10am closes that window at 3pm no matter how much or how little you sent in between. Layered on top of that is a weekly limit that resets on a fixed schedule tied to your account, separate from the rolling session clock.

Both meters draw from the same pool. Claude Code, the web chat, the desktop app, and Cowork all count against the same allowance, so a long agent run in your terminal and a browser conversation eat from the same bucket. What gets consumed isn't just raw token count either: message length, file attachments, tool calls like web search, which model you pick, and the effort level you set all factor in, according to Anthropic's own usage guidance. You can check Settings > Usage in the app for your current session and weekly progress at any time.

Capacity scales with plan. Pro runs $20/month. Max 5x is $100/month with five times Pro's session allowance, and Max 20x is $200/month with twenty times, per Anthropic's Max plan page. Anthropic doesn't publish fixed message counts for any tier, since usage varies too much by task for a flat number to mean much.

Why the numbers keep changing

If you've hit a usage limit mid-refactor and gone looking for the exact rule that stopped you, you've probably noticed the numbers don't hold still. Anthropic adjusts these limits often enough that a specific figure in any blog post, this one included, can go stale within weeks. The most recent example: a temporary 50% weekly-limit boost for Pro, Max, Team, and Enterprise plans ran through September 13, 2026. The next day it was replaced by a permanent 25% increase over the original baseline. Framed as an increase, that actually works out to roughly a 17% cut compared with what people had gotten used to during the boosted period, according to BleepingComputer's reporting.

The pattern tends to repeat: a promotional bump, an extension or two, then a quieter rebaseline once the promotion ends. If you're planning work around a weekly cap, treat any number you read, including the ones on this page, as a snapshot, and check your own dashboard before betting a sprint on it.

Why routine work burns the window just as fast as hard problems

Here's the part that catches people off guard: the meter doesn't know or care whether the work in front of it is hard. A tricky concurrency fix and the tenth re-run of a test suite this afternoon cost roughly the same per token, and an agentic loop that reads a dozen files, runs a linter, and retries a failing test chews through a session window fast, because every tool call, file read, and thinking block draws from the same allowance.

That's exactly the kind of work that fills a Max window before lunch: regenerating boilerplate tests, running a refactor across a dozen files, drafting docs, or letting an agent loop unattended on a long task overnight. None of it needs frontier-level reasoning. All of it counts the same as the genuinely hard problem you were trying to save your session for. Other tools run into the same math with their own shape: Codex CLI has its own usage limits and reset cadence, and Cursor's usage-based pricing works differently but hits a similar wall, since routine work and hard work draw from one pool, and the pool is finite.

Your options when you hit the cap

• Wait it out. The five-hour window clears on its own, and the weekly one resets on your account's fixed schedule.
• Drop to a lighter model or a lower effort level inside Claude, when the task allows it.
• Upgrade to a higher Max tier, which costs more every month whether or not you use the extra headroom.
• Route the routine share of the work somewhere that isn't metered by the same clock.

A flat-rate second lane beside Claude Max

That fourth option is where a flat-rate second lane fits, and it's worth being honest about what that means. NeverMetered isn't trying to replace a frontier plan. For the genuinely hard problem, a Claude Max session with Opus is still probably the right tool. What it's built to take off your plate is the part of the week that's tests, refactors, docs, and long-running agents grinding through a task overnight, the work you'd rather not spend session budget on in the first place.

It runs as a hosted API with OpenAI- and Anthropic-compatible endpoints on one base URL, so pointing an existing setup at it is usually a config change rather than a rewrite. It works with Claude Code and 18 other tools, including Codex CLI, Copilot CLI, Cursor, Aider, Cline, Continue, and OpenCode, through the same set of setup guides. There's no token meter, no five-hour window, and no weekly cap on any paid plan: once you subscribe, your price and how many requests can generate in parallel stay fixed for 12 months, and any change after that comes with 30 days' notice by email.

Instead of a time-based cap, NeverMetered prices by parallel streams, meaning how many requests can generate at once. Send more than your plan's stream count at the same time, and the overflow waits briefly in a fair queue instead of failing outright or running into a reset clock. See current plans for stream counts and a free tier you can try before committing to anything.

Prompts and completions are handled in memory only. They're never written to a database or a log, and no third-party AI provider ever sees them or trains on them. NeverMetered keeps request metadata, things like counts, timings, and status, to run the usage dashboard. The full picture of what runs where is on how it runs.

How Qwen3.8-27B actually performs

The model behind the second lane is Qwen3.8-27B, an open-weight model released under Apache 2.0 in August 2026 with a native 262,144-token (256K) context window. NeverMetered serves it through NInfer, an open-source inference engine, running a 4/5-bit quantized build that can score a little lower than the full-precision numbers below.

*One caveat on that starred row: it isn't a fully apples-to-apples comparison. Opus's 53.4 is Anthropic's own officially reported score, while Qwen re-ran the other three benchmarks on a corrected version of the test set. Treat that pairing with a bit more caution than the rest of the table, which comes from Qwen's own model card, retrieved September 17, 2026.

None of this makes Qwen3.8-27B a frontier model, and that's not how NeverMetered pitches it. It's a solid model for its size class, well suited to tests, refactors, and docs: the kind of work this page has been about the whole time. Live evals comparing the quantized build NeverMetered actually serves against full precision haven't been published yet; that work is in progress. Take the table above as Qwen's own reported numbers, not a promise of what you'll see in production.