Claude Code Pricing in 2026: Plans, Usage Limits and the Real Cost of an Always-On Agent

No. On the plan comparison at claude.com/pricing, the Claude Code row reads No for Free and Yes for Pro, Max and Ultra. The Free plan gives you chat on web, desktop and mobile with limited usage every 5 hours and every week, but not the coding agent.

There are two ways to get Claude Code running. A paid Claude plan: you log in with your claude.ai account (/login), and usage comes out of the plan's limits. A Claude Console account: you create an API key, set ANTHROPIC_API_KEY, and every token is billed at API rates. The authentication docs at code.claude.com/docs/en/authentication list both, plus cloud providers such as Amazon Bedrock, Google Cloud and Microsoft Foundry for organizations that already buy Claude there.

Either way you are paying for model usage, and that matters later, because the same claude binary can be billed in completely different ways depending on which credential it finds first.

These are the individual plans as listed on claude.com/pricing. Prices exclude tax; Anthropic notes that prices and plans can change at its discretion.

Pro is the entry point and the only individual plan with an annual option ($200 billed up front). Max and Ultra are billed monthly only. The new detail on both higher tiers is a monthly allowance of API credits for building on the Claude Platform, which means a Max or Ultra subscriber gets some pay-as-you-go budget for their own scripts on top of the interactive allowance.

The plan names changed recently. The @claudeai post announcing the rename (Max 5x to Max, Max 20x to Ultra, and the removal of the 5-hour limit for verified Ultra accounts) drew 6,219,596 views in our sample, 2.9x the account's median. Newer Ultra accounts keep a temporary 5-hour limit until they are verified, according to the same post.

Which plan for Claude Code? Anthropic does not publish a message or token count per plan; the pricing FAQ says how fast you reach a limit depends on the model, the length of your conversations and the size of your files. A practical rule from the plan structure: Pro suits a few focused sessions a day, Max suits all-day use, and Ultra is the plan that removes the 5-hour ceiling for long unattended runs.

Teams of 2 to 150 people buy the Team plan, which mixes two seat types. Both include Claude Code and Claude Cowork, central billing, single sign-on and no model training on your content by default.

The Team plan also lists up to $500 a month in pooled API credits. Enterprise works differently from every other plan: the seat fee buys access and admin features (SCIM, audit logs, role-based access, custom data retention, spend limits), and usage is billed at API rates on top, with a 20-seat minimum. So for an Enterprise rollout the useful planning number is token spend, not the seat price.

Anthropic gives that planning number in the Claude Code cost docs: across enterprise deployments, the average is about $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying under $30 per active day. The docs recommend starting with a small pilot group to set a baseline.

Subscriptions do not meter dollars; they meter usage against two windows. The pricing FAQ describes a limit that resets on a rolling five-hour session window, with weekly limits on top for paid plans. Pro gives at least 5x the usage per 5-hour session of Free, Max gives 5x Pro, and Ultra gives 10x Pro per week with no 5-hour window for verified accounts.

One pool for everything. The FAQ is explicit that activity across Claude on web, desktop, mobile and Claude Code all draws from the same pool. A long chat in the morning leaves less room for a coding session in the afternoon.

What happens at the limit. You can wait for the reset, move to a higher plan, or, on paid plans, turn on usage credits, which keep you working at standard API rates. Anthropic also reserves the right to apply other caps (weekly, monthly, per model or per feature) to keep access fair.

Temporary changes happen. The official accounts announce limit changes on X: in our @claudeai sample, Claude Code weekly limits were raised 50% for a period in spring 2026, extended once in July, and a May 2026 compute partnership was tied to higher usage limits for Claude Code and the API. Treat any fixed number you read in a guide, including this one, as something to re-check in Settings > Usage.

Why usage climbs in long sessions. The cost docs explain that Claude Code sends the full conversation with every request; a one-line question in a session open all day still draws usage for the whole history (at the cached rate if the cache is warm). A first message after a break longer than the cache lifetime reprocesses everything. And a session-scoped scheduled task fires on its interval even while the session is idle, sending the full context each time. For an always-on agent, that last point is the one that bites.

With an API key, Claude Code charges per token at Anthropic's list prices. These are the current models on the API tab of claude.com/pricing (per million tokens, standard tier):

Batch processing saves 50%, fast mode for Opus 5.5 costs 2x standard, and US-only inference costs 1.1x. Haiku 5.5 costs more for prompts over 100K tokens ($0.50 input, $2.50 output). Our Claude API pricing page goes deeper on caching and batch math.

The break-even is about usage shape, not price. If Anthropic's enterprise average ($150 to $250 per developer per month at API rates) is roughly your pattern, Pro at $20 is far cheaper as long as you fit inside its limits, and Max or Ultra exist for people who do not. An API key wins when usage is spiky or tiny (a few dollars of tokens in a month), when you need a hard budget per run (--max-budget-usd in print mode), or when the work runs unattended in CI where a personal subscription login does not belong.

The credential trap. The authentication docs list a precedence order, and ANTHROPIC_API_KEY ranks above your subscription login from /login. In interactive mode Claude Code asks once whether to use the key; in non-interactive -p mode a present key is always used. So a stray ANTHROPIC_API_KEY in your shell profile silently moves scripted runs onto per-token billing even though you pay for Max. For scripts that should use a subscription, the docs describe claude setup-token, which creates a long-lived CLAUDE_CODE_OAUTH_TOKEN for CI pipelines and scripts.

Checking spend. /usage shows plan usage bars for subscribers and a session cost estimate for API users; the docs stress that the dollar figure is computed locally at list price and that the Usage page in the Claude Console is the authoritative bill.

Plan cards price interactive coding. Many people buying Claude Code want something else: an agent that runs all day, watches X for a topic, and reports only what matters. That workload has two meters: the X data it reads and the model tokens it spends judging it. We measured the first and estimated the second.

The workload. A live search for "claude code" plus a cost word (limit, pricing, usage, plan, subscription, Max, Pro) returned 599 posts over 13.5 hours: 44.4 posts an hour, about 1,064 a day, with a peak of 58 in one clock hour. 48.7% of 60-second polls would have found nothing; at 5-minute polls only 4.3% would.

The tokens. The posts had a median of 270 characters, which is about 68 tokens; the mean was 211.2 tokens because long posts pull it up (p90 480). Token counts here are an estimate (characters divided by 4). We assumed 400 tokens of instructions per triage batch and 30 output tokens per verdict.

What the numbers say. At API list prices, a Haiku-based monitor costs less than half of a Pro subscription, data included. A Sonnet-based monitor at 60-second polls ($48.75) costs more than Pro and less than Max. The polling interval mainly changes the input-token bill, because every non-empty poll repeats the instruction prompt: 15.61 million input tokens a month at 60 seconds against 10.05 million at 5 minutes, for the same 31,934 posts.

Why not just run it on the subscription? You can (Claude Code in print mode with a CLAUDE_CODE_OAUTH_TOKEN), but every call then draws from the same 5-hour and weekly pool as your interactive work, and the docs warn that scheduled tasks resend the full context on every fire. A stateless print-mode call per poll, with only the new posts as input, is the pattern that keeps either bill small.

The data line in that table uses TwitterAPI.io's published prices: 15 credits per tweet returned, $0.15 per 1,000 tweets, with a 15-credit minimum per request, so a poll that finds nothing costs $0.00015 and a full 20-tweet page costs $0.003. You pay only for the data returned, and no X account is needed for reads.

That minimum sets a polling floor regardless of topic: checking every 60 seconds for a month is 43,200 checks, or $6.48 if every one came back empty; every 5 minutes it is $1.30. Our measured monitor sits above the floor because about half its 60-second polls returned posts.

On X's own pay-per-usage rate card (docs.x.com/x-api/getting-started/pricing) the same 31,934 posts a month would cost about $159.67 in post reads, roughly 33x the $0.15 per 1,000 rate. X also charges $0.010 per user read and offers $20 in credits on a first eligible card purchase plus a match of the first auto-recharge up to $50.

New TwitterAPI.io accounts that pass the automatic bot check get up to $1 of free credits (at most 100,000 credits, no card, one per person or network, expiring after 31 days). Before the first payment the rate limit is 0.2 requests per second, which is plenty for one monitor polling every 60 seconds; paid accounts get 20 requests per second. Purchased credits never expire.

Pricing news reaches users through X first, so we read the latest timeline posts of both official accounts (reposts excluded): 185 posts from @claudeai covering 390 days and 178 from @AnthropicAI covering 332 days.

Product news lives on @claudeai. More than a quarter of its posts mention Claude Code, against 4.5% on @AnthropicAI, and its posts about plans, prices and limits reach about twice the median views of its other posts. The limit-related posts were the outliers: the post extending higher Claude Code weekly limits drew 28,235,025 views, 13.3x the account median.

Practical takeaway. If you budget around a plan's limits, following @claudeai (or monitoring it with the script below) is how you learn about a temporary boost or a rename the day it happens, before the pricing guides catch up.

Pick the model per job. The model aliases in Claude Code (haiku, sonnet, opus, fable) let you route cheap triage to Haiku and keep Opus for hard reasoning; in our monitor the model choice moved the monthly bill from about $6 to about $64 with identical data.

Keep context small. Because every request resends the conversation, /clear between unrelated tasks and a lean CLAUDE.md save more than any plan change. The /usage breakdown on paid plans attributes usage to skills, subagents and MCP servers and suggests fixes, such as trimming a large CLAUDE.md.

Cap unattended runs. In print mode, --max-turns limits agentic turns and --max-budget-usd stops a run once estimated API spend reaches a set amount (the docs note spend can pass the cap, so leave headroom).

Check which credential is active. If scripted runs should use your subscription, make sure ANTHROPIC_API_KEY is not set in that environment; if they should bill to the API, prefer a dedicated Console key with its own spend limit.

Do not poll faster than the topic moves. For the X side, 5-minute polls cut the data bill from $7.95 to $4.85 a month on our measured topic, and on a quiet topic the bill stays close to the polling floor.

Question. What does an always-on Claude Code agent that watches X have to process each month, and what does that cost; and how do the official Anthropic accounts communicate plans and limits? Sources. Part A: a live /twitter/tweet/advanced_search for "claude code" plus a cost word, newest first, 599 posts from 528 distinct authors over 13.5 hours. Part B: /twitter/user/last_tweets for @claudeai and @AnthropicAI, about 200 posts each. Collected in 2026 through our rate-limited read client, 50 requests costing about $0.14 in credits.

Method. Polling windows were simulated over the covered span (60 s and 300 s). Tokens are an estimate (characters / 4), not tokenizer counts; Anthropic's tokenizer may count differently, and the triage prompt size (400 tokens) and output size (30 tokens per post) are our assumptions. Model costs use API list prices from claude.com/pricing, standard tier, no batch discount and no prompt caching. Topic labels on the official timelines are keyword matches on post text.

What the sample cannot show. 13.5 hours of one search is a snapshot: volume swings with news (a plan change can multiply it), weekends differ from weekdays, and a monitor on a different topic will see a different rate. Subscription costs for the same workload cannot be computed, because Anthropic does not publish per-plan token allowances. 74.0% of the sample was English; 29.4% were replies.

Privacy. Aggregates only. Search results come from private individuals, so raw pages stayed off the repository and only de-identified rows (hour, language, length, engagement counts) are kept. Only the two Anthropic organization accounts are named.