Mastering AI agents
Claude Code Pricing 2026: Subscription vs API Cost, and What I Actually Spend
Claude Code needs a paid Claude plan (Pro at US$20, Max at US$100 or Ultra at US$200 a month) or a Console account billed per token; the free plan doesn't include it. Official November 2026 prices, plus API rates, what to do when you hit a limit, and how to check your usage.
Kelvin Wang · · 8 min read
On a subscription, Claude Code draws from the same allowance as Claude on the web and in the apps, with a cap every 5 hours and another every week. Once it runs out, you wait for the reset or pay for more. Signing in with a Console (API) account works the other way: you pay for what you use, with no allowance to run out of. Every price here comes from Anthropic’s official pages as checked on November 17, 2026, excludes tax, and can change at any time 1.
- Cheapest individual plan that includes Claude Code (Pro, billed monthly)
- US$20
- Claude pricing page
- Average cost per developer per active day across enterprise deployments (at API rates)
- ~US$13
- Claude Code docs
- Average cost per developer per month, same enterprise basis
- US$150–250
- Claude Code docs
How do a subscription and the API differ?
| Item | Subscription (Pro, Max, Ultra) | API (Claude Console) |
|---|---|---|
| How you pay | A fixed monthly fee | Per token actually used |
| Limits | Yes. An allowance every 5 hours and every week (verified Ultra accounts have no 5-hour limit) | No usage allowance; you can set your own spend limit |
| Shared with Claude on the web? | Shares the same allowance | Billed separately; doesn't touch your subscription |
| How you sign in | Run claude and log in to your Claude account in the browser | Log in with a Console account, or set ANTHROPIC_API_KEY |
| When it runs out | Wait for the reset, turn on usage credits, or upgrade | Keep going and keep paying |
Anthropic’s Claude Code cost docs give one set of enterprise figures: at API prices, the average is around US$13 per developer per active day and US$150 to US$250 per developer per month, and 90% of users stay under US$30 per active day 3. Individual usage swings widely with the model, the size of the project and how you work, so treat these enterprise averages as a rough order of magnitude only.
Claude Code subscription plans in 2026
| Plan | Price (US$, before tax) | Claude Code | Usage |
|---|---|---|---|
| Free | 0 | Not included | Limited, every 5 hours and every week |
| Pro | 20 monthly, or 200 a year (about 17 a month) | Included | The baseline, every 5 hours and every week |
| Max | 100 a month | Included | 5x Pro, every 5 hours and every week |
| Ultra | 200 a month | Included | 10x Pro every week; verified accounts have no 5-hour limit |
| Team, standard seat | 25 per seat monthly, or 20 a month billed annually | Included | More than Pro |
| Team, premium seat | 125 per seat monthly, or 100 a month billed annually | Included | 5x a standard seat |
| Enterprise | 20 per seat a month (billed annually, 20-seat minimum), plus usage at API rates | Included | Billed by usage |
claude.com/pricing, checked 2026-11-17 [1]; the free plan doesn't include Claude Code [4]. Team is for 2 to 150 seats. The page lists no annual price for Max or Ultra.
A few things that are easy to misread:
- Ultra is the old Max 20x. Anthropic’s Help Center says Max 20x has been renamed Ultra, existing subscribers were moved over automatically, and the price stayed the same 7. Anthropic also changed how usage is described: Ultra is listed as 10x Pro per week, with no 5-hour limit for verified accounts. Older articles that mention “Max 5x” and “Max 20x” mean today’s Max and Ultra.
- The API credits in Max and Ultra don’t work in Claude Code. The pricing page says Max comes with US$100 a month in API credits and Ultra with US$200 1, but the Help Center states these credits can’t be used in Claude Code or the Claude apps 6. They’re meant for apps you build on the Claude API.
- The allowance is shared with the web app. A long chat with Claude in the browser during the day leaves less for Claude Code that evening.
- Fable has its own rules. On Pro, you pay for Fable with usage credits; Max and Ultra include it for up to half of your weekly limit 1.
- Prices exclude tax. If your card is issued outside the US, your bank may charge a foreign transaction fee on a US-dollar payment; the amount depends on the bank.
How is API usage billed?
When you sign in with a Console account, every request is billed by the token. A token is the unit a model reads and writes text in, and how many tokens a passage takes depends on its language. Claude Code mostly uses three models (Opus, Sonnet and Haiku), priced per million tokens:
| Model | Input | Output | Cache write (5 min) | Cache read |
|---|---|---|---|---|
| Opus 5.5 | US$4 | US$20 | US$5 | US$0.20 |
| Sonnet 5.5 | US$2 | US$10 | US$2.50 | US$0.10 |
| Haiku 5.5 (prompts up to 100K tokens) | US$0.10 | US$0.50 | US$0.125 | US$0.01 |
| Haiku 5.5 (prompts over 100K tokens) | US$0.50 | US$2.50 | US$0.625 | US$0.05 |
Per million tokens. Claude API pricing docs, checked 2026-11-17 [2]. Other models, such as Fable, are listed on the same page.
Each request Claude Code sends carries the whole conversation, your CLAUDE.md and the tool definitions. Prompt caching stores the repeated part, so the next request pays the cache-read rate for it. On Opus 5.5 and Sonnet 5.5 a cache read costs 0.05x the input price; on most other models it’s 0.1x 2.
The official docs include a sample /usage output that shows where the money goes. That session ran on Sonnet 4.6 and cost US$0.55 in total:
| Item | Tokens | Sonnet 4.6 rate (per million) | Cost |
|---|---|---|---|
| Input | 1,200 | US$3 | ~US$0.004 |
| Output | 5,300 | US$15 | ~US$0.08 |
| Cache read | 940,000 | US$0.30 | ~US$0.28 |
| Cache write | 50,000 | US$3.75 | ~US$0.19 |
| Total | — | — | ~US$0.55 |
Token counts from the example in the Claude Code cost docs [3], Sonnet 4.6 rates from the API pricing docs [2]; costs are my own arithmetic.
What you typed and what the model wrote back come to less than a fifth of the bill. More than 80% is the conversation being read from and written to the cache again and again, so the longer a session runs and the more files it pulls in, the more each message costs. The same token counts at Sonnet 5.5 rates come to about US$0.27. But Claude 4.7 and later models, Sonnet 5.5 included, use a newer tokenizer that turns the same text into roughly 30% more tokens, so the real saving is smaller 2.
What happens when you hit your limit?
When a subscription allowance runs out, Claude Code shows a session limit or weekly limit message with the reset time 3. You have four options 6:
- Wait for the reset. The 5-hour or weekly allowance comes back on schedule.
- Turn on usage credits. Switch them on under Settings → Usage on claude.ai and buy a block of credits; usage beyond your plan is charged at standard API rates 8. You can set a monthly spend limit or turn on auto-reload. If you’re signed in with your Claude subscription, running
/usage-creditsin Claude Code opens that settings page directly 3. - Upgrade your plan. Pro to Max, Max to Ultra.
- Switch to the API. Add credit to a Console account and run heavy work on API billing.
How do you check how much you’ve used?
- Run
/usageinside Claude Code. On a subscription you see your allowance bars and a breakdown of what’s using it: the share taken by skills, subagents and MCP servers, plus suggestions. Pressdorwto switch between the last 24 hours and the last 7 days. The dollar figure under Session is an estimate at API list prices, meant for API users; subscribers aren’t charged that amount 3. - The breakdown only counts this machine. It’s built from local session history, so usage on other computers or on claude.ai isn’t included.
- API users check the Console. Your actual bill is on the Cost page in the Claude Console, and you can set spend limits per workspace. The first time you sign in to Claude Code with a Console account, a workspace called “Claude Code” is created automatically, and all Claude Code usage is tracked there 3.
What makes usage go up?
Anthropic’s cost docs list these causes, along with what to do about each 3:
- Sessions that run too long. Every request sends the whole conversation, so in a session that’s been open all day, even a short question pays to read the full history. Run
/clearwhen you switch tasks. - Coming back after a long break. Once the cache has expired, your first message reprocesses the entire context.
- Staying on Opus. The docs recommend Sonnet for most coding work and Opus for complex architectural decisions; switch with
/model. - Thinking costs money too. Extended thinking tokens are billed as output. For simple tasks, turn it down with
/effort. - Subagents and agent teams. Each subagent makes its own requests, and the docs say agent teams use about 7x the tokens of a standard session when teammates run in plan mode.
- A long CLAUDE.md. It loads into every session. The docs suggest keeping it under 200 lines and moving workflow-specific instructions into skills, which load only when used.
- Scheduled and background tasks. A scheduled task such as
/loopsends your full context on every run, even when you aren’t typing.
How much do I actually spend a month?
As I wrote in the post about this site’s tech stack, I use both Claude Max x20 and GPT Pro. Max x20 is the plan Anthropic now calls Ultra: US$200 a month, US$2,400 a year, and Claude Code runs on that subscription’s allowance. Most people don’t need anything that high. If your usage is light, Pro at US$20 a month is enough to start with Claude Code, and the website cost breakdown is based on a single US$20 plan.
When I first launched this site, I was still on the US$100-a-month Max plan. For a simple website that allowance was fine. But building features like a booking system and finishing the whole site burns through a lot of tokens. My advice is to look on GitHub for an existing open-source project before you start. Starting from one saves far more of your allowance than having AI write everything from scratch.
Alternatives to Claude Code
Claude Code is one of several AI coding tools. OpenAI’s Codex and xAI’s Grok Build do similar work, and they bill in different ways: by subscription, by usage, or both.
Go deeper Claude Code vs Codex vs Grok Build in 2026: Which AI Coding Agent Fits You Who each suits, how you use it, pricing and limits, compared on the same criteria. No winner picked.Sources
Kelvin Wang
Electrical engineer building tools with AI agents, sharing hands-on notes on engineering and investing.
More about me