Coding agent economics
What each AI coding subscription is worth if the same usage were billed at public API rates. SemiAnalysis ceilings for Claude & ChatGPT/Codex; directional estimates for SuperGrok / Grok Build.
| Plan | You pay | API-eq (primary) | Range | × multiplier | Confidence | Source |
|---|
Width = API-equivalent dollars at full utilization. Gold tip = sticker price (almost invisible at this scale for top plans).
SemiAnalysis purchased each Anthropic and OpenAI subscription tier, ran long-horizon coding tasks until weekly limits were exhausted, then priced the tokens at public API rates. Widely reported ceilings:
Cursor / Copilot are credit / pass-through products (~1–2× sticker), not lab-style flat-rate subsidies. Gemini rows use community aggregators (medium/low confidence). SuperGrok has no official dollar pool — show the range, not just the midpoint.
Prior market belief was roughly “$200 plan ≈ $2,000 of tokens.” SemiAnalysis found several times that at full utilization — which is why top tiers look like massive compute subsidies.
xAI bills SuperGrok as a shared weekly compute pool (Chat, Imagine, Voice, Build) displayed as percent used. Official docs do not publish a dollar size. Low estimates reverse-engineer token % against Grok 4.5 list rates ($2 / $6 per 1M short context). High figures are informal “value framing,” not drain experiments. Extra Usage Credits after the pool are priced at higher “standard rates” than included usage.