One unit: the API-$ equivalent
API-$ equivalent: what the task would cost at the vendor's own API list price. One unit lets us compare a week of work against very different plans. It works because:
- Claude's meter is priced at API rates. The internal credit table mirrors API prices per token, and cache reads are free (she-llac, alldonesites).
- GitHub Copilot's AI Credits are dollars (1 credit = $0.01).
- Cursor's pools are dollars at API price.
ChatGPT's Codex limits are published as messages per 5 hours, so the “Codex local message” row is counted in messages and only against ChatGPT plans.
Capacity figures, and where each comes from
Weekly capacity is taken as stated when a weekly figure exists. Monthly allowances become weekly by dividing by 4.33 weeks per month (so Copilot weekly = monthly ÷ 4.33). Claude's 5-hour window figures for Pro and Max 5x are derived: the measured Max 20x window ($262) times she-llac's credit ratio between the plans.
| Plan | Figure | Value | Kind | Basis |
|---|---|---|---|---|
| Claude Pro | Weekly capacity (API-$) | $66 | inferred | alldonesites inferred figure: Pro about $66 of API value per week |
| Claude Pro | Per 5-hour window (API-$) | $13.10 | derived | Max 20x measured window $262 x (Pro 550k / Max 20x 11M credits per 5h, she-llac ratio) |
| Claude Max 5x | Weekly capacity (API-$) | $550 | inferred | alldonesites inferred figure: Max 5x about $550 of API value per week |
| Claude Max 5x | Per 5-hour window (API-$) | $78.60 | derived | Max 20x measured window $262 x (Max 5x 3.3M / Max 20x 11M credits per 5h, she-llac ratio) |
| Claude Max 20x | Weekly capacity (API-$) | $1,099 | measured | alldonesites: about $1,099 of API value per week (about 4.2 full windows), measured on 4 accounts, 21 Sep to 13 Oct 2026 |
| Claude Max 20x | Per 5-hour window (API-$) | $262 | measured | alldonesites: about $262 of API value per 5-hour window (token range 676M to 1,917M per window) |
| GitHub Copilot Pro | Monthly credits (API-$) | $15 | official | GitHub Copilot plans page |
| GitHub Copilot Pro | Weekly equivalent (API-$) | $3.46 | derived | $15 a month ÷ 4.33 weeks per month |
| GitHub Copilot Pro | Task cost factor (default model) | 0.8x | derived | Claude Opus 5.5 GitHub rate $4.00 per M input ÷ $5.00 reference; other models on the methodology page |
| GitHub Copilot Pro+ | Monthly credits (API-$) | $70 | official | GitHub Copilot plans page |
| GitHub Copilot Pro+ | Weekly equivalent (API-$) | $16.17 | derived | $70 a month ÷ 4.33 weeks per month |
| GitHub Copilot Pro+ | Task cost factor (default model) | 0.8x | derived | Claude Opus 5.5 GitHub rate $4.00 per M input ÷ $5.00 reference; other models on the methodology page |
| GitHub Copilot Max | Monthly credits (API-$) | $200 | official | GitHub Copilot plans page |
| GitHub Copilot Max | Weekly equivalent (API-$) | $46.19 | derived | $200 a month ÷ 4.33 weeks per month |
| GitHub Copilot Max | Task cost factor (default model) | 0.8x | derived | Claude Opus 5.5 GitHub rate $4.00 per M input ÷ $5.00 reference; other models on the methodology page |
Independent sources disagree in places: Classmethod measured much larger Max weekly caps during an August promo and found the 20x weekly cap about 3.5x the 5x cap. We use the newer alldonesites figures and say which are inferred.
GitHub Copilot: per-model rates
Copilot bills each model's per-token rate in AI Credits (1 credit = $0.01). We scale every task's API-$ by the chosen model's GitHub input rate / $5. Only models whose GitHub output rate is exactly 5x the input rate (same shape as the reference) are offered, so one factor is exact for input and output tokens. The default is Claude Opus 5.5; pick another model in the calculator.
GitHub also bills cached input (e.g. Claude Opus 5.5 $0.20 per M) and cache writes. Our task ranges come largely from Claude-meter measurements, where cache reads are free, so agentic-coding rows can cost more on Copilot than shown.
| Model | Input / M | Cached input / M | Output / M | Factor vs $5 reference |
|---|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | 0.8x |
| Claude Opus 5 | $5.00 | $0.50 | $25.00 | 1x |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | 0.4x |
| Claude Haiku 5.5 (prompts up to 100K) | $0.10 | $0.01 | $0.50 | 0.02x |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | 2x |
| GPT-6.1 Sol (up to 272K) | $2.00 | $0.10 | $10.00 | 0.4x |
| GPT-6 Astra (up to 272K) | $10.00 | $1.00 | $50.00 | 2x |
| GPT-6 Luna (up to 272K) | $0.10 | $0.01 | $0.50 | 0.02x |
| Gemini 4 Argon (promo price until 1 Dec 2026) | $2.00 | $0.10 | $10.00 | 0.4x |
| Kimi K3 | $3.00 | $0.30 | $15.00 | 0.6x |
Source: GitHub Docs: Models and pricing for GitHub Copilot (per-token rates in AI Credits), checked 2026-10-20.
Tasks: low, likely and high
Each task has three costs. Ranges were drafted at an Opus-class list price of about $5 per M input and $25 per M output tokens (Sonnet about 0.6x, Haiku about 0.2x). Re-checked 20 Oct 2026: Claude Opus 5.5 lists at $4 / $20 per M (claude.com/pricing), about 20% lower, so the high ends are conservative for Opus 5.5. On GitHub Copilot plans the calculator rescales every row to the chosen model's per-token rate published by GitHub (see methodology).
| Task | Low | Likely | High | Basis |
|---|---|---|---|---|
| Quick chat Q&A | $0.01 | $0.02 | $0.05 | arithmetic |
| Long document summary (about 75 pages) | $0.40 | $0.65 | $1 | she-llac cold-context example ($0.65) |
| Long chat thread (30 turns on a 50k-token context) | $0.50 | $1.50 | $4 | she-llac warm-cache math |
| Agentic coding, 1 hour, one interactive agent | $2 | $8 | $25 | Cursor guidance ($60-100/mo for daily Agent users); alldonesites window ($262 per 5h when saturated, so at most about $52/h) |
| Agentic coding, 1 hour, headless or 3 parallel agents | $10 | $35 | $75 | alldonesites 1.49x headless factor |
| Code review of a pull request | $0.20 | $0.50 | $1.50 | arithmetic; Copilot's legacy pricing charged 13 premium requests per review |
| Deep research report | $1 | $3 | $8 | no vendor publishes it; NotebookLM caps Deep Research at 200/day on Ultra 20x, Perplexity Max at about 500/month (unverified) |
| Codex local message (ChatGPT plans only) | messages | OpenAI publishes Codex limits as messages per 5 hours, not dollars, so this row only counts against ChatGPT plans | ||
The hit day and its range
For each scenario we add up the week day by day and find the first day the running total goes over the weekly cap. The likely scenario gives the headline day. The range runs from the high case (costs most, so it hits earliest) to the low case (hits latest, or never). That's why you'll read things like “Likely hits the weekly cap on Wednesday (range: Monday to never)”.
Off-days
If the likely week doesn't fit, we drop the heaviest days first, one at a time, until the rest fits under the cap. Those are the suggested off-days (or days to move work to another tool or plan).
The 5-hour session check
Claude also limits each 5-hour session. We multiply the plan's per-window figure by the windows you use per working day (1 to 4, default 2). If a day's likely cost is above that, the day is flagged and the plan doesn't count as fitting, even if the weekly total does.
Codex messages
For ChatGPT Plus we multiply OpenAI's per-5-hour message range for your model by your windows per day. Under the low end is clear, inside the range is cloudy (you may hit it), above the high end is a storm. OpenAI says weekly limits may apply but publishes no weekly number.
Cheapest plan that fits
We test every plan with a dollar capacity and pick the cheapest whose likely week fits (ties broken by plan id), plus the cheapest that fits even in the high case. Affiliate status is never read; the calculator has no access to it.
Confidence per vendor
- High confidence Claude (Anthropic): Official multipliers plus two independent measurements of absolute capacity (she-llac credit table, alldonesites live meter). Pro and Max 5x weekly figures are inferred, Max 20x is measured.
- High confidence GitHub Copilot: GitHub publishes hard numbers: monthly AI Credits in dollars (1 credit = $0.01).
- Medium confidence ChatGPT / Codex (OpenAI): OpenAI publishes message ranges per 5 hours for Codex on Plus. Pro tiers only get multipliers and an unpublished weekly allowance.
- Low confidence X Subscription (Grok + Cursor): Multipliers only (4x, 15x, 50x Plus). The size of the base weekly pool is not published.
- Low confidence Cursor (legacy plans): Monthly pools; only rough dollar guidance is published.
- Low confidence Google AI (Gemini): Compute-based multipliers only; hard counts exist only for Flow credits and NotebookLM.
- Low confidence Perplexity: The official limits page was blocked for our fetcher; counts are unverified.
Calibration
For plans that publish only multipliers, you can enter one reading: “Day X of my plan used Y% of my weekly meter”. We divide that day's likely API-$ by Y% to get an implied weekly capacity, then forecast as usual. The reading stays in your URL; nothing is sent anywhere.
Known gaps
- No absolute caps for Gemini, X Subscription, ChatGPT Pro or Perplexity.
- Claude Pro and Max 5x figures are inferred.
- Vendors change limits often (5+ changes in the last 6 weeks), so this needs a weekly refresh job.
- Google search volume was not measured.
Limits change often
Several limits moved in the six weeks before our check. We log them in the changelog. All figures were checked on 20 Oct 2026.