Plan-Based vs Token-Based Billing: How Claude Code, Codex, and Cursor Charge You (2026)

Two Ways AI Coding Tools Bill You

Every AI coding tool in 2026 charges you in one of two ways, and understanding which one you are on is the single biggest lever for controlling your bill. The two models are plan-based billing and token-based billing.

Plan-based billing is a fixed monthly subscription. You pay a flat fee — $20, $100, $200 — and in return you get a bundle of usage governed by rate limits or message quotas. Your bill is completely predictable: it does not matter whether you send one prompt or one thousand, the number at the bottom of the invoice is the same. The tradeoff is that when you hit the plan's caps, you either slow down, wait for a reset, or get pushed onto a higher tier.

Token-based billing is pay-as-you-go. You are charged per token — the small chunks of text a model reads and writes — with separate rates for input (what you send) and output (what the model generates). There is no monthly floor and no cap: light months cost almost nothing, heavy months cost a lot, and the total scales linearly with exactly how much work you do.

The cleanest way to understand the difference is to look at three popular tools that each started life leaning toward one model: Claude Code, OpenAI Codex, and Cursor.

Claude Code — The Token-Billing Archetype

Claude Code is the clearest example of pure token-based billing. Run it against the Anthropic API and you pay only for the tokens each session consumes, priced per million tokens by model:

• Claude Opus 4.8: $5 input / $25 output per million tokens — the flagship for complex reasoning and architecture.
• Claude Sonnet 4.6: $3 input / $15 output — the everyday workhorse for most coding.
• Claude Haiku 4.5: $1 input / $5 output — fast and cheap for simple edits and completions.

There is no monthly floor with this setup: a quiet week costs a few dollars, while a heavy agentic sprint with large context windows and multi-file operations can run into the hundreds. But Claude Code is not only token-billed. As of mid-2026 you can also drive it from a Claude Pro ($20/month) or Claude Max ($100 or $200/month) subscription, where the same tool runs against a plan quota instead of a metered API key. Same product, two different meters.

Codex — The Plan-Billing Archetype

OpenAI Codex approaches it from the opposite direction. For most users, Codex is something you get bundled inside a ChatGPT plan — ChatGPT Plus at $20/month or ChatGPT Pro at $200/month. You do not see a per-token line item; instead your usage is governed by rate limits that reset on a schedule. Once you hit the cap, you wait for the reset window rather than paying an overage. That is plan-based billing in its purest form: fixed cost, quota-limited throughput.

But Codex is not locked to the subscription either. The underlying coding model, GPT-5.3 Codex, is available through the API at $1.75 input / $14 output per million tokens. Teams that need automation, CI integration, or usage that outgrows a plan's rate limits can drop down to metered token billing and pay for exactly what they consume — the same shift Claude Code users make in reverse.

Cursor — The Explicit Hybrid

Cursor bakes both models into a single product. You start with a subscription — Cursor Pro at $20/month or Business at $40/month — which includes a bundle of usage. That is the plan-based layer, and for many developers it is all they ever pay. When you exceed the included allowance, Cursor switches you to token-priced overages on its own Composer models:

• Composer 2.5 Fast: $3 input / $15 output per million tokens.
• Composer 2.5 Standard: $0.50 input / $2.50 output per million tokens.

So a light Cursor user experiences pure plan-based billing — a flat $20 every month. A heavy user experiences a hybrid: the $20 base plus token-metered overages once the included usage runs out. Cursor makes explicit what Claude Code and Codex offer as separate paths: the two billing models are not rival products, they are two ends of the same spectrum.

Side-by-Side: Plan vs Token for Each Tool

Plan prices and quota terms shift frequently, so treat the subscription figures above as accurate for mid-2026 and check each provider before committing. The per-token API rates are the standard published rates for each model.

Which Billing Model Should You Pick?

Plan-based billing wins when you code interactively most days, want a predictable invoice your finance team can approve, and your usage sits comfortably inside the plan's caps. A developer in the editor six to eight hours a day almost always gets more value from a flat $20–$100 plan than from paying per token.

Token-based billing wins when your usage is bursty (heavy sprints followed by quiet weeks), you need backend automation or CI pipelines that a human-oriented subscription cannot serve, you want to route different tasks to different models, or you need per-project cost attribution that a flat plan obscures.

The most important takeaway for 2026: Claude Code, Codex, and Cursor all support both models now, so the real skill is knowing which meter you are currently on and whether it matches how you actually work. Many teams run both — subscriptions for daily interactive coders and token billing for automation and occasional contributors. To find your own break-even point, plug your usage into the AI Cost Estimator and compare the monthly token cost against the subscription options.