What AI Coding Agents Actually Cost in 2026
Claude Code, Copilot, Cursor, and Codex all sell you a monthly number. None of them sell you the number that matters — what happens the day you actually use the thing.
- GitHub Copilot killed flat pricing June 1, 2026 — Pro ($10/mo) and Pro+ ($39/mo) now burn AI Credits at published per-token API rates once you're past the included amount.
- Cursor's tiers are Pro $20 (a $20 credit pool), Pro+ $60 ($70 in credits), Ultra $200 ($400 in credits) — the plan price and the usage pool are two different numbers.
- Claude Max is $100/mo for 5x Pro usage or $200/mo for 20x — Anthropic doesn't publish the raw prompt count, just the multiplier off Pro.
- OpenAI folded Codex into ChatGPT Plus ($20/mo) and Pro ($200/mo) with a rolling 5-hour usage window, then added pay-as-you-go credits for when you blow through it.
As of August 2026, AI coding agents bill on two layers: a flat subscription ($10-200/mo) plus a usage pool that's now metered in tokens against published API rates, not flat 'requests.' GitHub Copilot switched to this model June 1, 2026 (AI Credits); Cursor has run credit pools since 2025; Claude Code and OpenAI Codex use rolling usage windows with paid overage credits. Read the credit-to-token math before picking a tier, not the sticker price.
I priced out four coding agent stacks this week for a client migration, and every single vendor’s pricing page told me a different kind of half-truth. Not lies — just the number that flatters them most, front and center, with the number that actually determines your bill buried two clicks down or not published at all.
Here’s what’s real, sourced off each company’s own pricing docs, as of this week.
The subscription number isn’t the usage number
The single biggest gotcha across all four tools: what you pay monthly and what you get to spend are increasingly different figures.
Cursor is the clearest about it. Pro is $20/month and buys a $20 credit pool — 1:1. Pro+ is $60/month but buys $70 in credits. Ultra is $200/month for $400 in credits. That’s a real discount at the top of the ladder, but it means comparing “$60 vs $200” tells you nothing until you convert to the actual usage number.
GitHub Copilot just joined this pattern. On June 1, 2026, GitHub retired flat premium-request pricing and moved Copilot to AI Credits — consumed by actual input/output/cached token counts against each model’s published API rate, not a flat “1 request = 1 unit” count. Copilot Pro ($10/mo) gets $10 in monthly credits. Pro+ ($39/mo) gets $39. Code completions and Next Edit suggestions stay free and don’t touch the pool — only agent and chat work does.
The receipts
| Tool | Tier | Price | What you actually get |
|---|---|---|---|
| Claude Code | Pro | $17-20/mo | Claude Code included, “more usage” than Free, no published prompt count |
| Claude Code | Max 5x / Max 20x | $100 / $200/mo | 5x or 20x Pro’s usage multiplier, still no raw number published |
| Claude Code | Team | $20-25/seat/mo | Claude Code + Cowork, seat-based |
| GitHub Copilot | Pro | $10/mo | Unlimited completions, $10/mo in AI Credits for agent/chat |
| GitHub Copilot | Pro+ | $39/mo | Premium models incl. Opus, $39/mo in AI Credits |
| GitHub Copilot | Business / Enterprise | $19 / $39 per seat/mo | Matching AI Credits per seat, admin spend controls |
| Cursor | Pro | $20/mo | $20 credit pool, frontier model access, MCPs/hooks |
| Cursor | Pro+ | $60/mo | $70 credit pool (3x Pro’s agent limits) |
| Cursor | Ultra | $200/mo | $400 credit pool (20x Pro’s agent limits) |
| OpenAI Codex (ChatGPT) | Plus | $20/mo | Codex in a rolling 5-hour usage window, pay-as-you-go credits past that |
| OpenAI Codex (ChatGPT) | Pro | $200/mo | Highest included Codex window, same credit top-up option |
None of these vendors publish an exact “N tokens per dollar” table for the subscription tiers — that math only shows up once you’re buying overage credits at raw API rates. Anthropic and Cursor both describe Pro/Max tiers purely in multiplier language (“5x,” “20x,” “3x”) without an absolute baseline, which is a deliberate hedge: it lets them absorb model-cost changes without reprinting the pricing page.
Where the trap actually is
The trap isn’t the cheap tier — it’s the middle tier bought out of anxiety. Copilot Pro+ at $39/mo and Cursor Pro+ at $60/mo both exist because people hit a wall on the entry tier and jump two rungs “to be safe.” But agentic coding burns unevenly: a routine day costs nothing, and one long-running background agent can eat a week’s credit pool in an afternoon — you’re billed on tokens generated, not flat requests.
The value play is staying on the entry tier — Copilot Pro or Claude Pro — until you can point to a specific week you actually ran out, then upgrading on real usage data instead of a guess. If you’re already committed to Copilot’s Agent HQ workflow or weighing BYOK against Copilot’s managed models, the token-metered world makes that call higher-stakes — a bad model pick now shows up directly in the credit burn, not just output quality.
Claude Code and OpenAI Codex both describe their tiers in the same vague multiplier language — read it straight off the source before committing a team.
Business and Enterprise tiers on both Copilot and OpenAI’s side skip the guessing and go to admin-controlled credit budgets — worth it once more than two or three people draw from the same pool, exactly the direction the production AI stack conversation has been heading all year.
#TheAIMogul
Bottom lineThe value play is Claude Pro or Copilot Pro at the entry tier if you're a single builder doing normal-volume work — both hand you real agentic coding for $10-20/mo. The trap is the $39-60 middle tier bought 'to be safe': that's exactly where people discover their agent burns a full day's credit pool on one gnarly refactor, because every tier now settles in tokens, not vibes.