July 2026. Companion to levelcode-paid-plans-model-economics.md (credit scheme) and levelcode-free-tier-economics.md. Grounded in current vendor docs (sources at end).

cursor.png (free/Hobby account, at the limit) shows two unrelated things:

1. Plan usage exhausted: the "You've hit your usage limit" banner + Upgrade CTA, and the model picker locked (🔒 next to Composer 2.5 Fast). Hobby ≈ 2,000 completions + ~50 premium agent requests/mo (deliberately unpublished / adjustable).
2. Context Usage panel (9% of 200K): a transparency feature showing what fills the context window — system prompt 486, tool definitions 8.8K, rules 3.7K, MCP 1.3K, subagent definitions 850, conversation 3.4K. This is not plan usage — it's context composition.

claude.png (Claude Code, Max plan) shows:

1. Context window 365.4k / 1.0M — again context, not plan.
2. Plan usage: "5-hour limit · Resets 3:30 PM · 59%" and "Weekly · all models · Resets Jul 12 · 35%" — the two rolling budgets described below.

• Dual limit: a 5-hour session budget + a rolling 7-day weekly cap. The session opens at your first prompt and covers everything for the next five hours; the timer follows the account, not the clock. The weekly cap is the absolute ceiling for a rolling week, aimed at "24/7" heavy users, and some plans carry a separate weekly Opus budget.
• Shared pool: Claude Code and Claude chat draw from the same session pool.
• Unit of account is opaque: budgets are compute-based; Anthropic publishes ratios, not tokens (Max 5x / Max 20x = 5× / 20× Pro's session capacity; 5-hour limits were doubled across paid plans on May 6, 2026). Expensive models burn the budget faster.
• At the limit: hard stop with a reset countdown; Max plans can buy extra usage at API rates; model fallback (Opus → Sonnet) stretches budgets.
• Surface: Settings → Usage progress bars; in-product meter like the screenshot.

What this design buys: the 5h window smooths infrastructure load and prevents budget hoarding into single mega-bursts; the weekly cap bounds worst-case cost per subscriber (the "cap is the margin floor" principle). The opacity gives Anthropic tuning freedom but generates constant user confusion/anxiety — the top complaint.

• Unit of account is dollars: each paid plan includes a monthly credit pool ("included usage", ≈ plan price — Pro = ~$20/mo) that depletes by the actual inference cost of each request. No request counts — a $20 pool buys ≈225 Sonnet requests or ≈650 GPT-4.1 requests. Resets on the billing date.
• Free (Hobby): small fixed allowances (≈2k completions, ≈50 premium requests), unpublished exact numbers; at the cap → the banner + locked premium models.
• At the limit (paid): opt-in on-demand usage billed in arrears (pay-as-you-go overage), or keep working via Auto.
• The Auto escape valve: Auto mode (Cursor picks the model — usually their own Composer) does not draw from the included pool; it's their margin lever — cheap self-hosted inference presented as an unlimited-feeling lane.
• Surface: in-chat banner (screenshot), dashboard with per-model spend, usage emails; plus the Context Usage breakdown panel for context transparency.

What this design buys: dollars are honest and models are instantly comparable; overage converts power users into revenue instead of churn. Weakness: no rate governor — a user can burn a month's pool in a weekend, and "dollar anxiety" makes some users under-use the product.

Monthly compute budget B per plan comes from the credit-scheme doc (Pro $14.43 / Pro+ $33.05 / Max $51.67 / Ultra $88.91). On top:

1. Session governor (Claude-style): 5-hour rolling cap = B / 20 (Pro ≈ $0.72 ≈ 19 Kimi turns per 5h at full burn). Allows ~3× the average day in a burst, smooths gateway load, kills runaway-loop damage.
2. Weekly governor: rolling 7d cap = B / 3.2 (Pro ≈ $4.51). Bounds a compromised/abusive account to ~31% of monthly budget per week while letting legitimate crunch weeks through.
3. At-limit ladder (graceful, not a wall):
   • 75% / 90% of any window → quiet meter color change + toast.
   • Session cap hit → offer the free lane: "Session limit reached — continue on gpt-oss-120b (doesn't count) or wait 1h 12m." (0.05× cost ≈ noise; the user never stops working — better than both competitors.)
   • Weekly/monthly cap → Cursor-style banner + single CTA (upgrade tier or opt-in overage at cost+20%, Max/Ultra only), with Claude-style reset time.
4. Auto lane: "Auto" model choice routes gpt-oss/Kimi and is marketed as unlimited on paid plans — honest footnote: subject to the weekly governor.
5. UI: one popover, exactly the claude.png layout (context bar + plan bars with reset times) — we already ship the context donut (PLAN §M5); add the two plan bars from the Redis ledger. Free tier gets the cursor.png treatment: banner + 🔒 on gateway models, BYOK always unlocked (our unique third lane).
6. Ledger truth: meter actual provider cost per request (M10 usage ledger); publish per-model multipliers on the pricing page; recompute monthly.

Why this wins: Cursor's honesty (dollars, published multipliers) + Claude's load-smoothing and abuse bounds (rolling governors) + a degradation lane neither has (gpt-oss continuation instead of a hard stop) + BYOK as the pressure-release valve no competitor offers at all.

• https://support.claude.com/en/articles/9797557-usage-limit-best-practices
• https://www.morphllm.com/claude-code-usage-limits
• https://usagebar.com/blog/claude-code-weekly-limit-vs-5-hour-lockout
• https://cursor.com/help/models-and-usage/usage-limits
• https://cursor.com/docs/models-and-pricing
• https://www.verdent.ai/guides/cursor-usage-limits-explained
• https://www.finout.io/blog/what-happened-to-cursor-pricing-2026-guide-5-cost-cutting-tips