Claude Code pricing: every plan explained

Claude Code doesn't have its own price tag. It's a feature that rides on top of whatever Claude.ai plan you're already paying for, or it runs straight off an API key with no subscription at all. That's usually the first thing that trips people up when they go looking for one clean number: there isn't one, there are eight.

The free tier gets you a small number of chat prompts every five hours and no terminal access. Everything past that is either a seat on a Claude.ai plan or metered API usage, and which one makes sense depends entirely on how you actually work.

What Claude Code costs, plan by plan

A couple of things in that table surprise people. Team Standard seats include Claude Code, not just Premium, according to Anthropic's own seat-management documentation; Premium just buys more usage headroom, not access to the tool itself. And the API route needs no Claude.ai plan whatsoever, you can run Claude Code against a bare API key from day one.

Pro vs Max: the question that actually matters

Most people typing "claude pro vs max" into a search bar aren't comparing feature lists. They want to know one thing: will I get cut off mid-task. Pro promises "at least 5x" the Free tier's per-session allowance, and Max multiplies Pro's allowance by 5 or 20 depending on which tier you pick. All three run on the same rolling 5-hour window, and none of them publish a fixed message count. Anthropic's own help center says the numbers get "tuned over time," which is a polite way of saying don't quote a specific figure back at them, check your account's usage page instead.

One detail matters a lot for Claude Code specifically: usage on Pro and Max is shared between the web chat and the terminal. There's no separate coding allowance sitting off to the side. Burn through your budget on a long agentic session in the morning and you'll have less left for chatting with Claude on the web that afternoon.

In practice, Pro tends to hold up for a handful of sessions a day with modest diffs. Someone running Claude Code for most of a working day, especially on long or agentic tasks, usually finds Max 5x closer to the realistic floor, and people who lean on it constantly end up on Max 20x. For the mechanics of what actually happens when you hit the ceiling, see Claude usage limits; the same rolling-window shape with different multipliers shows up in Codex usage limits and Cursor usage limits too, so it's worth comparing all three if you're switching tools.

How the 5-hour and weekly windows work

The 5-hour window is a rolling clock, not a fixed slot on a calendar. It starts the moment you send your first message and runs five hours from there, then resets. Max plans add a second ceiling on top of that: a weekly cap across every model, resetting at a fixed time assigned to your account. Anthropic's support documentation also reserves the right to apply further weekly or monthly caps "at our discretion" on any plan, so what you see today isn't a permanent guarantee.

If you run out mid-task on a paid plan, Claude Code doesn't just stop working. It offers to keep going on API credits, billed at the standard per-token rate instead of your plan's flat price, so a long session can quietly turn into a metered one without much warning.

According to usage trackers that follow these changes, Anthropic doubled Claude Code's 5-hour limits across Pro, Max, Team, and Enterprise seat plans earlier in 2026 and dropped a peak-hour reduction that used to squeeze Pro and Max accounts during busy periods. Worth knowing, and also a reminder that these numbers move. Check your account's usage page rather than trusting any article's exact figures, including this one.

Team and Enterprise: seats, not a per-developer flat fee

On Team plans, Claude Code ships with every seat, Standard and Premium alike. The difference is usage headroom: Standard seats get a base per-session allowance, Premium seats get noticeably more. A team can mix seat types, put the heaviest Claude Code users on Premium and leave lighter users on Standard, rather than paying the top rate across the board.

Enterprise drops the flat-multiplier model entirely. You pay a per-seat price plus consumption billed at standard API rates, so cost scales with how much the team actually uses instead of sitting on a fixed tier. Team gives you predictable per-seat billing; Enterprise gives you usage-driven billing with more admin and compliance controls layered on top.

Paying per token instead of by seat

You don't need a Claude.ai subscription to run Claude Code at all. Point it at an API key and it bills per token, at the rates Anthropic publishes for each model. As of September 2026, Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, and Anthropic made that price permanent rather than letting a scheduled increase to $3/$15 take effect on September 1. Claude Haiku 4.5 is the cheaper option for routine tasks at $1/$5 per million tokens. On the reasoning end, Anthropic released Claude Opus 5.5 on September 22, 2026 at $4/$20, a 20% cut from Opus 5's $5/$25 and roughly a 40% reduction in typical workload cost, with cache reads at a steep $0.20/M; Opus 5 remains available alongside it. Claude Fable 5.1 tops the range at $10/$50, with a notably cheap $0.25-per-million-token rate on cache reads.

Whether the API route beats a subscription depends entirely on how much you run. Cost-tracking tools that watch real Claude Code API spend put typical usage around $13 on an active day and $150 to $250 a month for a regular developer, with roughly 90% of users staying under $30 on their heaviest days. Below that range, a subscription and pay-as-you-go land close to a wash. Well above it, a Max plan's fixed price starts looking better, right up until you hit its usage ceiling.

When a flat-rate second lane makes sense

None of the plans above change the basic shape of Claude Code pricing: you're capped by a session window that resets every five hours, or you're paying per token with no ceiling at all. Both are reasonable trade-offs for the hardest problems, the ones where you want your best model doing careful, expensive reasoning.

A lot of Claude Code usage isn't that kind of work. It's the stuff that's routine but still eats a session:

• Running the test suite after every change
• Grinding through a large, mechanical refactor
• Writing and updating docs
• Long-running or overnight agent sessions
• Anything repetitive enough that spending your best model's budget on it feels wasteful

That's the kind of work where a flat-rate second lane makes sense alongside your main plan, not instead of it: keep Pro or Max for the tasks that need a frontier model, and send the routine stuff somewhere with no token meter, no 5-hour window, and no weekly cap.

That's the idea behind NeverMetered. It's an OpenAI- and Anthropic-compatible API priced by how many requests run in parallel rather than by the token, so a long agentic session doesn't chip away at a budget the way it would on a metered plan; requests over the limit wait briefly in a queue instead of failing outright. It runs Qwen3.8-27B, an open-weight model that's solid for its size class without pretending to replace a frontier plan on the hardest problems. It plugs into Claude Code, Cursor, Codex CLI, and the rest of the 19 tools NeverMetered supports through the same OpenAI- and Anthropic-compatible APIs your existing setup already speaks. See current plans for exact pricing, how it actually runs for the model, server, and data-handling details, or how it compares to running something locally if that's the alternative on the table. Sign up to try it against your own workload.