AI coding assistant pricing compared in 2026

The sticker prices, side by side

Here is what the four main AI coding assistants charge individual developers per month, as listed on their own pricing pages in September 2026. All figures are in US dollars and exclude tax. This post is about cost alone, so if you are still deciding which tool suits your work, start with our comparison of the best AI coding assistants and what each one is good at and come back here to price it.

GitHub Copilot: Free, Pro $10, Pro+ $39, Max $100.
Cursor: Hobby free, Pro $20, Pro+ $60, Ultra $200.
Claude Code: included with Claude Pro at $20 a month, or $17 a month on annual billing. Claude Max starts at $100 and offers a choice of 5x or 20x Pro usage. The free Claude plan does not include Claude Code.
OpenAI Codex: included with ChatGPT. Free, Go $8, Plus $20, Pro from $100 with 5x or 20x limits, with a $200 tier at the top.

That list is where most comparisons stop, and it is close to useless on its own. Copilot Pro at $10 and Cursor Pro at $20 are not the same product at half the price. What separates them is the included usage behind the fee, and each vendor measures that in a different unit.

Four vendors, four incompatible units

This is the part that makes a like-for-like comparison genuinely hard, and it is worth understanding before you pick a tier.

GitHub prices its allowance in dollars

Copilot moved to usage-based billing in June 2026. Usage is now metered in GitHub AI Credits, and GitHub's billing documentation on doc.github.com fixes one credit at $0.01. Each paid plan gets base credits matching the subscription price plus a variable "flex allotment" on top:

• Pro at $10: 1,000 base plus 500 flex, so 1,500 credits, worth $15.

• Pro+ at $39: 3,900 plus 3,100, so 7,000 credits, worth $70.

• Max at $100: 10,000 plus 10,000, so 20,000 credits, worth $200.

Credits do not roll over. The allowance resets at 00:00 UTC on the first of each calendar month, regardless of when your billing date falls. Code completions and next edit suggestions are not billed in credits at all and stay unlimited on every paid plan.

Cursor splits usage into two pools and describes them as multiples

Cursor's pricing page no longer publishes a dollar figure for included usage. Pro+ is described as 3x Pro limits on Agent, Ultra as 20x, and Teams Premium as 5x Standard. Its models and pricing docs explain the structure: there are two pools, one for Cursor's own models (Grok 4.6, Grok 4.5, Composer 2.5) and one for third-party models charged at each model's API rate. The plan table marks both pools as simply "Included."

Cursor does publish the most useful budgeting guidance of any vendor, though. Its docs estimate that daily agent users typically run $60 to $100 of total usage a month, and power users running multiple agents or automations often exceed $200.

Claude Code sells multiples of Pro, measured per session

Claude Code has no separate price. It is bundled into every paid Claude plan and draws from the same usage pool as chat, so a heavy day in the web app eats into your terminal budget. Anthropic's pricing page describes limits as multipliers on a rolling five-hour session window, with weekly limits stacked on top: Pro gives at least 5x Free per session, Max gives 5x or 20x Pro. When you hit a limit you can wait for the reset, upgrade, or switch on usage credits and keep working at standard API rates.

Codex prices in credits, but not in dollars

OpenAI meters Codex in credits per million tokens. The Codex pricing documentation publishes a full rate card: GPT-5.6 Sol costs 100 credits per million input tokens and 500 per million output, Terra 50 and 300, Luna 5 and 30. It also gives per-model message estimates for each five-hour window, such as 10 to 100 Sol messages on Plus and 200 to 2,000 on Pro 20x.

What it does not publish on that page is what a credit costs. The docs say credit purchase prices and discounts depend on your plan or agreement. So unlike GitHub's credits, you cannot convert a Codex allowance into dollars from the pricing page alone.

Doing the arithmetic anyway

Two of these units can be turned into money with published numbers, which is as close to a fair comparison as the market currently allows.

For Copilot, the conversion is direct. Pro+ gives you $70 of credits for a $39 fee. As a scale check, Anthropic lists Claude Opus 5 at $5 per million input tokens and $25 per million output. Spent entirely one way, $70 is roughly 14 million input tokens or 2.8 million output tokens. Real agent sessions land somewhere between those two figures, and GitHub prices each model against its own published rate card, so treat this as an order of magnitude rather than a quote.

For Codex, you can price a task in credits even without the dollar conversion. A session that pulls in 200,000 tokens of context on GPT-5.6 Sol and writes 30,000 tokens of output costs 20 credits for input and 15 for output, so 35 credits total. OpenAI's own note is that GPT-5.6 usage averages 5 to 30 credits per message, which makes a long agent run visibly more expensive than a chat exchange.

For Cursor and Claude Code, no equivalent sum is possible from the pricing pages. You buy a multiple of a tier you have not measured yet, which means the only reliable method is to run a tier for two full billing cycles and read the usage dashboard.

The upgrade nobody mentions is free

There is one number that barely moves across all four tools: the cost of ordinary autocomplete. Copilot's completions and next edit suggestions are unlimited and unmetered on every paid plan. Cursor includes unlimited tab completions from Pro upward. If you use these tools mainly as smart autocomplete, the cheapest paid tier will last you indefinitely and the entire pricing debate is irrelevant to you.

Everything expensive is agentic. Chat, agent mode, code review, CLI sessions and cloud agents are what consume credits, and a single agent run across a large codebase can involve dozens of model calls. Model choice compounds it: Cursor's rate card puts Claude Fable 5.1 at $10 per million input tokens against $0.75 for Gemini 3.8 Flash, more than a tenfold spread for the same request. Routing routine edits to a cheaper model is a bigger lever than any plan change, and our breakdown of how ChatGPT, Claude and Gemini differ in practice covers which jobs are safe to hand to the cheap tier.

Team seats: what you are actually buying

At team scale the seat price stops being the interesting number.

• Copilot Business: $19 per granted seat, contributing 1,900 credits per user to a shared enterprise pool. Copilot Enterprise: $39 per seat, 3,900 credits.

• Cursor Teams: Standard $40 per user, Premium $120 per user with 5x Standard agent limits.

• Claude Team: Standard seat $20 per seat on annual billing ($25 monthly), Premium seat $100 annual ($125 monthly) with 5x Standard usage. Seat types can be mixed on one account.

• ChatGPT Business: $20 per user on annual billing ($25 monthly), minimum two users.

Now compare the ratios. A Copilot Pro subscriber pays $10 and receives $15 of credits, a 1.5x return, and Max returns 2x. A Copilot Business seat costs $19 and receives exactly $19 of credits, a flat 1.0x. The business seat is not buying you more usage per dollar. It is buying policy control, pooled credits and central administration, and the individual plans are quietly the better usage deal.

Cursor adds a charge that only appears at team scale: a Cursor Token Rate of $0.25 per million tokens on third-party model requests, applied on top of the model's API price and charged even on bring-your-own-key usage. Opting into regional data residency adds a further 10% on eligible models.

Which tier fits which developer

Mostly autocomplete, occasional chat. Copilot Pro at $10 or Cursor Pro at $20. You will rarely touch the ceiling, and the free tiers are enough to judge the editor before you pay.

Daily agent use on one project. Cursor's own estimate of $60 to $100 of monthly usage points at Pro+ at $60. The equivalents are Copilot Pro+ at $39 with $70 of credits, or Claude Pro at $20 if your sessions are bursty rather than continuous.

All day, multiple agents. The $100 to $200 tiers exist for this and nothing below them will hold. Cursor Ultra, Copilot Max, Claude Max and Codex Pro all sit in this band.

Learning, not shipping yet. Stay on the free tiers. None of these subscriptions will teach you anything a free plan will not, and the money is better spent elsewhere while you are training, as our look at what a coding bootcamp costs against its job outcomes sets out.

A team. Start people on individual plans, watch the usage dashboards for two billing cycles, then buy seats for the people who need governance rather than for everyone at once. Every one of these vendors lets you mix seat types.

FAQ

Is Claude Code sold separately from Claude?
No. It is included in all paid Claude plans and shares one usage pool with chat on web, desktop and mobile. There is also a pay-as-you-go route through the API if you would rather not subscribe.

What happens when I run out of included usage?
It depends on the meter. Copilot's credits reset on the first of the calendar month, and you can set a dollar budget to keep working at $0.01 per credit. Cursor bills on-demand usage in arrears at the same rates. Claude and Codex reset on rolling five-hour windows, so you may only need to wait, and both let paid users buy additional usage.

Which is cheapest overall?
For autocomplete-only work, Copilot Pro at $10. For agent-heavy work, the honest answer is that nobody can tell you without your usage data, because three of the four vendors do not publish included usage in a unit you can convert to money.

Do the free tiers do anything useful?
They are enough to evaluate. Copilot Free includes 2,000 completions a month with limited chat and agent usage, Cursor Hobby gives limited agent requests and Composer access with no card required, and Codex has limited free access. Claude's free plan is the exception, as it does not include Claude Code.

Before you subscribe

Pricing in this category changed several times in 2026, including Copilot's move to token-based credits in June. Every figure above was checked against the vendors' own pages in September 2026, and every one of those pages is worth opening before you pay, because plan names, credit allowances and model rate cards all move faster than the comparison articles that cover them.

If you are choosing right now, the practical sequence is unchanged: run the cheapest paid tier for two full billing cycles, read your usage dashboard, and let the measured number pick the plan.