AI Coding Agent Pricing: Cursor, Claude Code, Codex, GitHub Copilot, Replit, and Grok
- 23 hours ago
- 8 min read
Every major AI coding tool now bills by consumption rather than by a flat monthly fee, and all six platforms compared here made that shift within the same twelve months — so the number on the pricing page tells you almost nothing about what you'll actually pay.
........
Cursor runs six tiers from free to $200/month, spending a credit pool that depends on which model you pick — Auto mode is unlimited, manual frontier-model selection is not.
Claude Code has no separate subscription: it's bundled into Claude Pro ($20), Max ($100 or $200), or billed pay-as-you-go through the API.
OpenAI Codex moved from message-based estimates to token-metered credits in April 2026, riding on whatever ChatGPT plan you already pay for — Free through $200 Pro.
GitHub Copilot replaced Premium Request Units with token-based AI Credits on June 1, 2026, at unchanged list prices from $0 to $100.
Replit bills every agent action as a credit-consuming checkpoint — including failed ones — which is why real-world Replit bills commonly run 3-4x the subscription sticker price.
Grok Build 0.1 is the cheapest of the six on a pure per-token basis: $1 input / $2 output per million tokens on the API.
··········
WHY COMPARING THESE PRICES IS HARDER THAN IT LOOKS.
None of these six tools sell a fixed amount of AI work for a fixed price anymore, and they each broke from flat pricing in a different way.
Cursor moved from a simple 500-fast-requests-per-month Pro plan to a credit pool where different models cost different multiples of a credit.
Codex moved from an averaged per-message cost to a direct token-to-credit mapping in April 2026.
Copilot moved from Premium Request Units — a flat count regardless of how much a model actually had to think — to per-token AI Credits on June 1, 2026.
Replit and Grok Build never had flat pricing to begin with: Replit bills by the agent action, and Grok Build is a raw per-token API model.
The result is that "how much does X cost" no longer has a single honest answer for any of these tools — it depends on which model you route to, how large your context is, and how often the agent fails and retries.
··········
CURSOR: SIX TIERS, AND A CREDIT POOL THAT MOVES WITH YOUR MODEL CHOICE.
Cursor's pricing spans free to $200 a month, and the subscription fee only buys a credit pool — it doesn't cap what you can spend.
Hobby is free, with limited Agent requests and Tab completions — enough to evaluate the tool, not to code on daily.
Pro, at $20 per month, is the standard individual plan and the one most reviewers point new users toward.
Pro+ at $60 gives roughly three times Pro's usage credits, and Ultra at $200 gives roughly twenty times Pro's usage plus priority access to new features — a tier explicitly positioned as comparable to Claude Code Max at the same $200 price point, the difference being that Cursor is a full visual IDE while Claude Code is a terminal.
Teams runs $40 per seat per month, layering centralized billing, usage analytics, and SSO on top of Pro-equivalent access; Enterprise is custom-priced with pooled organization-wide usage.
··········
Cursor plans at a glance
Plan | Price | What it's for |
|---|---|---|
Hobby | Free | Evaluation only |
Pro | $20/month | Standard individual daily use |
Pro+ | $60/month | ~3x Pro's credits, power users |
Ultra | $200/month | ~20x Pro's credits, full-time agentic work |
Teams | $40/user/month | Pro-equivalent access + admin controls |
Enterprise | Custom | Pooled org usage, SCIM, dedicated support |
··········
The real cost driver is model selection: Auto mode is unlimited on every paid plan, and switching to it instead of manually picking a frontier model on every request is the single biggest lever on an actual monthly bill.
··········
CLAUDE CODE: NO STANDALONE PRICE, JUST YOUR CLAUDE PLAN.
Claude Code isn't sold separately — it rides on whichever Claude subscription you already have, or on raw API billing.
Claude Pro at $20 a month covers short coding sessions in smaller codebases and includes both Sonnet and Opus model access.
Claude Max, at $100 or $200 a month depending on the tier, targets everyday use in larger codebases with more generous usage limits — the $200 tier is the one Cursor's own pricing explicitly benchmarks itself against.
The alternative is pay-as-you-go through the Claude API at standard per-token rates, with no per-seat fee at all — the natural fit for teams that would rather pay for what they consume than for a seat that might sit idle.
Anthropic's own guidance for newcomers is to start on Pro and upgrade only once real usage patterns are clear, rather than guessing upfront.
··········
OPENAI CODEX: BUNDLED INTO CHATGPT, NOW METERED BY THE TOKEN.
Like Claude Code, there's no such product as a standalone "Codex subscription" — it's a feature that rides on your ChatGPT plan, and since April 2026 it's billed by token consumption rather than by message.
The lineup runs Free (limited trial), Go at $8/month, Plus at $20/month, and two separate Pro tiers: $100/month for 5x Plus's usage, and $200/month for 20x — split apart in April 2026 specifically so heavy users stop subsidizing light ones and vice versa.
Business runs $20-25 per user per month depending on billing term, and Enterprise and Edu are quote-only.
One credit costs four cents. A multi-file refactor priced out at $2.40 on the flagship Sol model against roughly ten cents on the cheap Luna tier — same task, 24 times the cost, and the only variable was which model handled it.
Plus and every tier above it cover Codex on the web app, CLI, IDE extension, and iOS; an API key skips the ChatGPT plan structure entirely and bills strictly per token with no ceiling.
··········
Codex tiers and what they multiply
Tier | Price | Usage multiplier vs. Plus |
|---|---|---|
Free | $0 | Limited trial |
Go | $8/month | Baseline, lighter use |
Plus | $20/month | 1x (the reference point) |
Pro (5x) | $100/month | 5x Plus |
Pro (20x) | $200/month | 20x Plus |
Business | $20-25/user/month | Shared workspace pool |
··········
GITHUB COPILOT: THE JUNE 2026 SWITCH FROM REQUESTS TO TOKENS.
Copilot's list prices didn't move in 2026, but what sits behind each price changed completely on June 1, when Premium Request Units were replaced by token-metered AI Credits.
The lineup is Free, Pro at $10/month, Pro+ at $39/month, Max at $100/month, Business at $19/seat/month, and Enterprise at $39/seat/month — prices identical to the old Premium Request era.
Each paid tier now bundles a monthly credit allowance instead of a request count: Pro includes roughly $15 in credits, Pro+ roughly $70, and Max roughly $200, with one credit equal to one cent.
Inline code completions and next-edit suggestions stay free and unmetered on every paid plan — only chat and agent-mode usage draws from the credit pool.
Pro and Pro+ don't just differ on credits: Opus-class models are no longer available on the base Pro tier as of the 2026 changes, while Pro+ keeps them. A price table alone won't show you that.
··········
Copilot plans and their credit allowance
Plan | Price | Included AI Credits |
|---|---|---|
Free | $0 | Limited, no card required |
Pro | $10/month | ~$15 |
Pro+ | $39/month | ~$70 |
Max | $100/month | ~$200 |
Business | $19/seat/month | Org-pooled |
Enterprise | $39/seat/month | Org-pooled |
··········
REPLIT: EVERY AGENT ACTION IS A BILLABLE CHECKPOINT — EVEN WHEN IT FAILS.
Replit's meter runs whether or not the agent actually succeeds at the task, which makes it the hardest of the six to budget for in advance.
Starter is free with daily-refreshing agent credits and one publishable app — a trial tier rather than a working setup.
Core runs $20-25 per month (depending on billing term) and includes a matching amount in monthly usage credits, covering AI, compute, and deployments together rather than as separate line items.
Pro, which replaced the older Teams plan in February 2026, runs $100 a month flat for up to 15 builders — not per seat — with pooled, tiered-discount credits and one month of rollover.
A single tracked billing cycle put numbers on that: 632 Agent checkpoints at $0.25 each plus 965 Assistant checkpoints at $0.05 each, for $206.25 in checkpoint charges alone — with debugging iterations, failed fix attempts, and operations that hang mid-execution billed exactly the same as the ones that worked.
That's the mechanism behind a pattern trackers have flagged repeatedly: active teams commonly end up paying three to four times Replit's headline subscription price, and hosting an always-on app keeps drawing from the same credit pool long after the build itself is finished.
··········
GROK BUILD: THE CHEAPEST PER-TOKEN RATE OF THE SIX.
Grok Build sits apart from the other five because it's priced as a raw API model rather than as a seat-based subscription product.
grok-build-0.1 costs $1.00 per million input tokens and $2.00 per million output tokens on the xAI API — a third of what Grok's own general-purpose flagship, Grok 4.6, charges for output.
The Grok Build agent CLI itself is bundled into SuperGrok and X Premium+ subscriptions rather than sold on its own, so a developer who only wants API access never needs to touch the consumer subscription tiers at all.
Because reasoning is always on and not configurable for this model, there's no lower-effort setting to dial down cost the way Codex's tier ladder allows — the token rate itself is the only lever.
··········
SIDE BY SIDE: WHAT $20 A MONTH ACTUALLY BUYS.
$20 is the one price point where four of these six tools directly overlap, and each one hands you a different kind of tool for it.
Cursor Pro at $20 buys a credit pool sized for daily individual coding, spent down faster or slower depending on how often you bypass Auto mode.
Claude Pro at $20 buys Claude Code access with both Sonnet and Opus models, sized for short sessions in smaller codebases rather than continuous all-day agentic work.
ChatGPT Plus at $20 buys Codex on web, CLI, IDE extension, and iOS, running background tasks rather than the real-time interactive style Cursor is built around.
GitHub Copilot Pro is only $10 — half the other three — and buys roughly $15 in credits for chat and agent-mode usage on top of unlimited free inline completions.
··········
What $20/month buys, side by side
Tool | $20 tier | Style of work |
|---|---|---|
Cursor | Pro | Real-time IDE, interactive editing |
Claude Code | Claude Pro | Terminal agent, short sessions |
OpenAI Codex | ChatGPT Plus | Background cloud tasks |
GitHub Copilot | (Pro is $10, not $20) | Inline completions + light chat/agent |
··········
THE HIDDEN-COST PATTERNS THAT REPEAT ACROSS ALL SIX.
Read across these six pricing pages and the same handful of traps show up again and again, no matter which company wrote them.
Model selection is the single biggest lever on every credit-based tool here: Cursor, Codex, and Copilot all charge different multiples of their credit currency depending on which underlying model handles a request, so defaulting to the cheapest capable option is the one habit that pays off on all three.
Replit bills failed and retried work the same as work that succeeds, and the other tools carry a version of that same risk: a coding agent that needs three attempts to get something right has effectively tripled its own token cost, whichever platform it's running on.
Annual billing quietly discounts several of these plans by around 20% (Cursor explicitly, others less transparently), which changes the real comparison between tools if you're willing to commit for a year.
Seat-based team pricing buys administrative control — centralized billing, SSO, usage analytics — rather than more AI capability on Cursor, Copilot, and Claude Code alike; a team without compliance requirements is often better served by individual accounts with expense reimbursement.
··········
HOW TO ACTUALLY CHOOSE.
Price narrows the list; the workflow each tool was built around is what should actually decide it.
For real-time, interactive editing inside a visual IDE, Cursor's $20 Pro tier is the standard entry point, with Auto mode as the main cost control.
For terminal-native, deep-codebase agentic work, Claude Code on Claude Pro or Max is the fit, particularly for teams already invested in Anthropic's models elsewhere.
For background tasks that run while you do something else, Codex on ChatGPT Plus or Pro is built for that specific pattern rather than for live pair-programming.
For teams that want the lowest floor with a genuinely useful free tier, GitHub Copilot's $10 Pro plan and unlimited free inline completions are hard to beat on pure price.
For rapid, full-stack app generation where the platform also hosts the result, Replit is the fit — provided the team budgets for real usage at three to four times the sticker price rather than the sticker price itself.
For anyone building directly against an API rather than through a branded product, Grok Build 0.1's $1/$2 per-million-token rate is worth benchmarking against whatever else is under consideration on token cost alone.
The test that actually settles it: run the same representative task on each candidate, in a real session, and work out cost per accepted result. That number, not the one on the pricing page, is the one worth writing down.
··········
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
