Claude Code stopped. That doesn't mean the model is broken. In 2026, the usual cause is a Session Limit, a weekly allowance, or a separate Opus cap. All three run at once. The errors look alike. The fixes do not.
Anthropic does not publish exact token numbers. Plans are described as multiples of Pro. This article follows the official rules: weekly allowances on Pro, Max 5x, and Max 20x; the 5-hour rolling window; what actually burns tokens; when limits reset; and what it costs to keep going on API or Usage Credits after you hit the wall.
Know the three limits first
A Claude Code subscription has three gates at once. Session and weekly allowances are shared across models—switching /model will not save you. Only the Opus cap is model-specific.
You've hit your session limit · resets 3:45pm
You've hit your weekly limit · resets Mon 12:00am
You've hit your Opus limit · resets 3:45pm
| Limit | What it does | Shape | Does switching models help? |
|---|---|---|---|
| Session (5 hours) | Stops short bursts | Rolling window; usage expires gradually | No — shared across models |
| Weekly allowance | Caps the week | Resets on your account cycle | No — shared across models |
| Opus cap | Protects the most expensive model | Counted separately | Yes — switch to Sonnet / Haiku and keep going |
The Free plan does not include Claude Code. The web app, Desktop, and Claude Code share the same pool.
How Pro and Max weekly allowances compare
Prices follow the official Claude pricing page (September 2026):
| Plan | Monthly price | vs Pro | Best for |
|---|---|---|---|
| Pro | $20 (~$17 billed yearly) | 1× | Single sessions, intermittent use |
| Max 5x | $100 | 5× Session + weekly allowance | Full-time daily Claude Code |
| Max 20x | $200 | 20× Session + weekly allowance | Multi-session, large repos, long-running agents |
Community estimates (not an official token table—treat them as a feel): Pro is roughly 40–80 hours of effective Sonnet compute per week; Max 5x is about 140–240 hours of Sonnet plus some leftover Opus; Max 20x can reach 240–480 hours of Sonnet. How long you last depends on the model, context length, and tool output—not how many messages you send.
Since May 6, 2026, the 5-hour window on paid plans is about twice as large, and peak-hour cuts were dropped. Weekly allowances have seen +50% promos. Trust what /usage shows on your account—do not treat last year's posts as a contract.
Don't mix Fable 5 with the weekly allowance
Fable 5 is a more expensive high-end model, not the default. Max / premium seats can spend up to 50% of the weekly allowance on Fable 5—that is a slice of the same pool, not an extra 50%. On Pro, Fable 5 usually bills Usage Credits from the first request.
For whether a Cursor subscription is worth it, see Cursor Pro real cost analysis.
Session Limit: a rolling 5 hours, not a clock-hour wipe
Session Limit measures tokens burned in the last 5 hours, not a daily reset at 9 a.m.
- Fill it from 9–11 a.m. and capacity comes back through the afternoon—not in one instant at a round hour.
- The error names a time, such as
resets 3:45pm. That is when the window is empty. Capacity is already recovering while you wait. - Parallel sessions on the same account share this pool. Open 3 windows and you burn about ×3. Agent teams in plan mode are officially about 7× a normal session.
Max does not remove Session Limit. It only makes the bucket 5× or 20× larger. Two Max 5x accounts ($200) cost the same as one Max 20x, but you get two independent windows—often the better deal when you work in parallel.
Token burn: the five things that actually spend the quota
Limits count tokens / compute, not turns. One Opus sweep of a large repo can cost more than fifty short Haiku replies.
- Model choice: Opus / Fable cost several times Sonnet or Haiku. Use
/model sonnetor Haiku for mechanical work; switch back to Opus for hard problems. - Long context: The full conversation is resent on every request. After a day-long session, even "continue" pays for the whole history.
- Cache misses: Under a subscription, Prompt Cache lasts about an hour. On Usage Credits or an API key it often drops to 5 minutes. A long lunch, and the first afternoon request bills as a full input.
- Tool output: A
grepthat dumps 4,000 lines, a full test log, or a 6,000-line source file all enter context and get resent on later turns. - Parallel and scheduled work: Multi-agent and scheduled tasks fire with full context. Idle sessions are nearly free; background summaries are usually cheap; scheduled tasks are not.
To save quota: /clear unrelated work first (free)—do not /compact first (compaction is itself a large request). Keep CLAUDE.md around 200 lines; put details in Skills and load them on demand. For how to write Skills, see Claude Code Skills templates.
# Switch model and thinking effort in-session
/model sonnet
/effort low
# Unrelated work: rename, then clear — cheaper than compact
/rename api-auth-fix
/clear
How to read reset times
| Limit | How it resets | Where to look |
|---|---|---|
| Session 5 hours | Rolling; the error names a time | Error text, /usage, status line |
| Weekly allowance | On your account cycle (often something like Monday midnight) | /usage, claude.ai → Settings → Usage |
| Opus cap | Separate cycle | Same as above; switch models and continue immediately |
/usage reads local session history. Usage on another machine or on claude.ai is not included. For the full account, trust the web Usage page.
A custom status line can keep both windows visible:
{
"rate_limits": {
"five_hour": { "used_percentage": 23.5, "resets_at": 1738425600 },
"seven_day": { "used_percentage": 41.2, "resets_at": 1738857600 }
}
}
After you hit the wall: wait, step down, or buy overflow
- Session / weekly limits: Switching models does nothing. Wait for the window to refill, or use
/usage-creditsat API prices (you can set a monthly cap; the daily redemption ceiling is commonly $2,000). - Opus cap:
/model sonnetor Haiku immediately and keep working. - HTTP 429: That is API throughput, not a subscription window. Check for a stray
ANTHROPIC_API_KEYin the environment, then lower concurrency.
On Team / Enterprise without billing permission, /usage-credits sends a request to an admin. After you switch to Credits, cache lifetime shortens, so the same work costs more.
For how this relates to AI controlling a computer, see Claude Code and OpenAI Computer Use.
Subscription vs API: which is cheaper
Subscriptions have 5-hour and weekly windows. An API key has neither—you pay per token and per-minute throughput. Usage Credits overflow at standard API prices.
Official list prices (per million tokens, September 2026, Claude API Pricing):
| Model | Input | Output | Cache hit |
|---|---|---|---|
| Haiku 4.5 | $1 | $5 | $0.10 |
| Sonnet 5 | $2 | $10 | $0.20 |
| Opus 5 / 4.8 | $5 | $25 | $0.50 |
| Fable 5 | $10 | $50 | $1 (lower on 5.1) |
Rough math: a mid-size agent job of about 80k input + 8k output is about $0.24 on Sonnet 5; the same scale on Opus 5 is about $0.60. A Prompt Cache hit can drop input to about 10% (about 2.5% on Fable 5.1). Batch API is another 50% off, but it is async—not for interactive coding.
How to choose:
- One or two focused sessions a day → Pro is usually enough.
- Full weekdays, often hitting Session → Max 5x.
- Multiple windows, large repos → Max 20x, or split across two Max 5x accounts.
- Overnight batches you can run async → API / Batch is often cheaper than forcing a higher subscription.
- Occasional overflow → Usage Credits; you do not need to upgrade for the whole month.
FAQ
Does Session Limit count messages?
No. All three limits count tokens / compute. Longer context, more tool output, and a more expensive model burn faster.
Does switching to Sonnet lift Session / weekly limits?
No. Only the Opus cap recovers immediately when you change models. Session and weekly allowances are shared across models.
Does Max still have limits?
Yes. Max multiplies Pro's Session and weekly allowances by 5 or 20. The mechanism does not change.
Does the API have 5-hour and weekly windows?
No. The API bills per token. What you hit is org-level throughput (429), not a subscription window.
Which numbers should I trust?
Trust /usage and the claude.ai Usage page. Hour counts and message counts in third-party posts are estimates. Anthropic can change multipliers and promos at any time.
Spend the quota on real work. Keep the machine up.
Claude Code's limits track cloud tokens, but the agent still has to read the repo, run tests, and open MCP locally. Close a laptop lid and the session dies, the cache misses, and the first afternoon request bills as a full input. A Mac mini idles at about 4W—enough to leave a terminal up 24 hours so /usage and the rolling window follow your pace, not the lid.
Apple Silicon unified memory is steadier for indexing large repos and running tests in parallel. macOS ships with native Unix, Homebrew, Docker, and SSH. Claude Code's computer-use also treats macOS as the main battlefield. Versus a similarly priced Windows box, you get fewer crashes and better unattended runs; Gatekeeper and SIP also lower the risk of leaving an agent up long-term.
If you are already stretching a weekly allowance, a stable, low-power cloud Mac cuts the cost of a session dying and burning the context again—view Kvmzen plans and spend tokens on the code, not on warming up again.