Claude Code vs Codex vs Cursor: Pricing in 2026
The Short Answer
Same headline price, three completely different billing models. That is the whole comparison.
| Claude Code | Codex | Cursor | |
|---|---|---|---|
| Entry | Claude Pro $20/mo | Bundled with ChatGPT (free tier up) | Hobby free |
| Mid | — | ChatGPT Plus $20/mo | Pro $20/mo, Pro+ $60/mo |
| Top individual | Max from $100/mo | ChatGPT Pro $200/mo | Ultra $200/mo |
| Billing shape | Weekly usage allowance | Token-based credits | Plan allowance + overage |
| Budget tier | — | ChatGPT Go $8/mo | Hobby free |
| Overflow | API key billing | Add credits | Usage-based pricing |
Verified August 20, 2026.
The Structural Difference
Claude Code meters a weekly allowance. You buy a plan, you get a weekly budget of agent work, and when it runs out you wait or switch to API billing. Max plans offer 5x or 20x Pro’s usage from $100/month. For heavy users this is the clearest upper bound in the category — you know your worst case is the subscription price.
Codex meters token credits inside ChatGPT. OpenAI moved Codex to token-based credit pricing during 2026, so real consumption depends on input, cached input, output, model and task complexity. Access now extends down to the free tier, with ChatGPT Go at $8/month (raised from $6 in July 2026) as the budget entry point. Practical daily use starts around Plus at $20/month. Codex’s shape is built for parallel, asynchronous execution — batches of background tasks rather than one terminal conversation.
Cursor mixes a plan allowance with usage-based overage. Hobby is free, Pro $20, Pro+ $60, Ultra $200. Cursor’s own documentation notes daily Agent users typically land above the Pro allowance. This is the most uneven cost profile of the three: cheap when you are mostly using tab completion, expensive when the agent is running all day.
What This Means Practically
If your usage is steady and heavy: Claude Code Max. The allowance model prices predictably, and Max tiers have been estimated at roughly 18x cheaper than equivalent API usage for high-volume users. You are effectively buying wholesale inference.
If your usage is spiky: Codex or Cursor. Paying $200/month for a tool you use hard three days a month is worse than credit metering, even at a higher per-task rate.
If you are evaluating: Codex, because it reaches down to the ChatGPT free tier and Go at $8/month, making it the cheapest real trial. Cursor’s Hobby tier is also free but thinner on agent capability.
If you want tab completion plus autonomy: the common pattern is both — Cursor for daily inline coding and completions, Claude Code for complex autonomous tasks, multi-agent orchestration and long sessions where large context matters. Two $20 subscriptions beat one $200 subscription for most individual developers.
The August 2026 Wrinkle
Anthropic’s +50% Claude Code weekly limit boost now runs through August 31, 2026, announced August 18-19 and covering Pro, Max, Team and seat-based Enterprise. Anthropic said it hopes to make it permanent. It has already been extended four times since roughly May 2026 — July 13, July 19, August 19, August 31.
Budget as though it ends on September 1. If your team’s throughput depends on the extra headroom, that is a dependency on a promotion, not a plan.
Separately and more durably: Claude Sonnet 5 API pricing at $2/$10 per MTok is permanent — the increase to $3/$15 that had been scheduled for September 1, 2026 was cancelled per Anthropic’s documentation around August 10, 2026. Two different September 1 stories were circulating in August; one is settled and one is not.
The Lever Subscriptions Hide
All three subscription models share a flaw: they discourage model routing.
When you have paid $200 for a plan, using the frontier model for everything feels free. It is not — it consumes your allowance, and when the allowance binds, that is the bill arriving late. A trivial task run on Opus 5 costs roughly $0.275 at API rates against $0.041 on Gemini 3.7 Flash or $0.0099 on DeepSeek V4 Flash off-peak for the same 30K-in/5K-out shape.
If you routinely hit weekly limits, the fix is usually not a higher tier. It is moving test generation, dependency bumps, docs and mechanical refactors onto a cheaper model, and reserving frontier capability for work that needs it. On Claude Code specifically, habitual Fable 5 use is the most common cause of limit pressure — it is capped at 50% of the weekly allowance and weighs roughly double an Opus session against it.
Bottom Line
Copilot is cheapest to start. Cursor is the most uneven on cost. Claude Code gives the clearest upper range for heavy users. Codex is the cheapest to trial and the best fit for parallel background batches. None of the three is meaningfully cheaper than the others across all usage patterns, which is exactly why they all landed on $200.
Last verified: August 20, 2026.