Claude Code cost per month, per engineer, and how to cap the bill

As of 5 October 2026, Anthropic puts the average Claude Code cost per month at $150 to $250 per developer, measured across enterprise deployments. That works out at about $13 per active day, and 90% of users stay under $30 a day.¹ Spend per engineer has a long tail, though. In Gartner Peer Insights research reported by Computer Weekly, 23% of tech leaders spend $200 to $500 per developer per month on coding-agent tokens, and 6% of organisations pay more than $2,000.²

For an engineer who uses an agent on most working days, a seat plan costs less than API billing at Anthropic’s averages, provided the work fits inside the plan’s limits. The caps that protect a budget are hard limits per person and per workload, plus a separate budget for the most expensive models.

Below are the published spend figures, plan prices and limits for Claude Code, Codex and Cursor, the cost of a single task, and worked budgets for teams of 1, 5 and 25 engineers. After those come the caps that stop spend, the settings that cut token use and a policy template. Prices and limits are as of 5 October 2026.

What teams spend per engineer on coding agents

On 24 June, Gartner predicted that AI coding costs will overtake the average developer’s salary by 2028, as token consumption rises and vendors move from seat licences to consumption-based pricing.³ The same day, Gartner analyst Nitish Tyagi told The Register that bills were rising from $20 or $100 to $2,000 to $5,000 per developer per month, and could reach $20,000 in extreme cases.⁴

The published figures measure different things, so read the third column before comparing them.

The Databricks figure shows how fast spend can move. When the company gave GPT Astra to a control group with no cost mitigations, the average developer spent 60% more than before. Databricks says an overnight rise of that size is very hard to plan for across more than 10,000 users.⁶ It now gives every user four budgets:

• A monthly maximum across all models.
• A daily runaway limit, which the user can raise in Slack.
• A quality frontier budget, which reserves a fixed share of the monthly budget for GPT Astra and Claude Fable. Databricks says these cost 2 to 3 times the next tier down.
• An experimental budget for new, untested models.

Databricks doesn’t publish the amounts, and the policy template at the end of this post borrows its structure.

Codex pricing vs Claude Code and Cursor plans

All three vendors sell a $20 entry plan and a $200 plan for heavy individual use. OpenAI added a $500 tier, Pro 500, at its DevDay event on 29 September.⁷ The table lists prices per person per month in US dollars, before tax, as of 5 October 2026.

Each vendor meters included usage differently. OpenAI estimates that a five-hour window on Plus or Standard Business holds 15 to 160 local Codex messages on GPT-6.1 Sol, or 5 to 45 on GPT-6 Astra, depending on the task. Pro plans have no five-hour limit, though weekly limits may apply. Fast mode uses the allowance at 2.5 times the standard rate, and GPT-6 Astra Ultrafast at 8 times.¹⁰

At DevDay, OpenAI said Pro 500 includes 25 times the usage of Plus.⁷ New Pro 200 subscribers get a lower allowance, and earlier subscribers keep the old one until 29 October.¹¹ The Next Web reported that the cut halves Pro 200’s Work and Codex usage.¹³

Cursor gives each plan a monthly allowance in two pools, one for its own models and one for third-party models billed at API prices, then switches to on-demand usage. On Teams and Enterprise, third-party models also carry a Cursor Token Rate of $0.25 per million tokens.⁵

Claude Code usage limits since 14 September

Claude subscriptions have a session limit that resets every five hours and a weekly limit on top. Max gives 5 or 20 times Pro’s usage per five-hour session, and a Team Premium seat gives 5 times a Standard seat.⁸ Anthropic’s pricing page doesn’t state these allowances in tokens or dollars, so a seat can’t be converted into an API budget exactly.

From 13 May to 13 September, Anthropic raised weekly Claude Code limits by 50% on Pro, Max, Team and seat-based Enterprise plans. Since 14 September they have been 25% above their level before the promotion.¹⁴ That is about a sixth less than during the promotion (1.25 divided by 1.5 is 0.83), so a team that sized its seats over the summer can now hit weekly limits it didn’t hit in August. The five-hour limits didn’t change.

Fable models have their own rule. On Max and Team Premium seats, Fable 5 and Fable 5.1 can use up to half of the weekly limit, and they use it faster than other models. On Pro and Team Standard seats, Fable runs only on usage credits.¹⁵

What one coding task costs

Specific Labs’ Real-SWE benchmark prices attempts at real work. Its published results cover 10 tasks from private production codebases, run eight times per model in each model’s own harness at high reasoning effort. Its estimated cost per attempt:¹⁶

The most expensive model costs 2.8 times as much per attempt as the cheapest. No model passed half of its attempts, so the cost of a finished task is more than twice these figures. Ten tasks is a small sample, so we use the costs and leave the model ranking aside.

Databricks measured cost per session from its gateway traces, comparing early adopters with their own usage a week earlier and reweighting sessions by type. Opus 5.5 averaged $4.23 a session against $5.94 on Opus 4.8, 29% less. GPT-6 Sol averaged $2.34 against $4.52 on GPT-5.6 Sol, 48% less, after a 50% price cut.⁶ At $4.23 a session, Anthropic’s $13 average day is about three Opus 5.5 sessions, though the two figures come from different groups of users.

Most tokens in an agent session are context that is sent again on every turn and served from the prompt cache. Anthropic’s cost docs show a sample session with 1,200 fresh input tokens, 5,300 output tokens, 940,000 cache reads and 50,000 cache writes.¹ We priced those tokens on four Claude models at API rates as of 5 October 2026, with five-minute cache writes:¹⁷

Cache reads and writes are 80% of the Opus 5.5 total, so anything that breaks the cache costs more than a longer answer does. Opus 5.5 and Sonnet 5.5 both charge $0.20 per million cached tokens read, which is why moving this session to Sonnet saves about a third of the cost. The per-token prices suggest a half.

Claude Code cost per month for teams of 1, 5 and 25

These budgets compare Claude seat plans with API billing, because Anthropic is the only vendor here that publishes an average daily cost. They are illustrations built from the cited figures. We have not measured our own spend for this post. The assumptions:

• A1. A typical engineer costs $200 a month on API billing, the midpoint of Anthropic’s $150 to $250 average. Anthropic’s averages are token costs, which is what API billing charges.¹
• A2. A heavy engineer costs $600 a month, which is $30 per active day over 20 working days. Anthropic says 90% of users stay under $30 a day, so this understates the top tenth.
• A3. One engineer in ten is heavy, rounded up. That is one in a team of 5 and three in a team of 25.
• A4. Seats are billed monthly. One person uses Max 5x at $100 or Max 20x at $200, and teams use Team Premium at $125 per seat.⁸⁹ Team plans need at least two people.
• A5. Seat allowances aren’t published in dollars, so the seat cost is a floor. Engineers who pass their allowance buy usage credits at API rates. The last column shows how much the team can spend on credits before seats cost more than the API.

For 25 engineers, the API figure is 22 × $200 + 3 × $600 = $6,200, and the seats are 25 × $125 = $3,125. Annual billing lowers the seats to $2,500.

Per engineer, a seat pays for itself once the engineer would have spent its price on the API. At Anthropic’s average of $13 per active day:

Anthropic’s $150 to $250 monthly average implies about 12 to 19 active days at $13 a day. That clears the break-even point for every plan in the table, except Max 20x for engineers at the low end of the range.

API billing has a ceiling of its own. Anthropic’s usage tiers cap monthly API spend at $500 on Start, $1,000 on Build and $200,000 on Scale. An organisation that reaches its tier’s cap is paused until the first day of the next month, unless it gets a higher limit sooner.¹⁸ The five-person team above would hit the Build cap before the month ended, and a single heavy engineer would pass the Start cap.

If one of the three heavy engineers in the team of 25 spends $2,000 in a month, the API total rises from $6,200 to $7,600. On seats, that engineer draws usage credits instead, up to whatever per-member limit you set.

Enterprise is a third route, at $20 per seat billed annually plus usage at API rates.⁸ On our assumptions, 25 engineers would cost $500 plus $6,200, or $6,700 a month. That is more than Team seats here. Enterprise admins can also set the default model, restrict which models each role can use and cap effort per role.¹⁹

On these numbers, seats cost less for anyone who uses an agent on most working days. API billing suits occasional users and automation. We would buy seats for daily users, run CI and scheduled jobs on API keys in a workspace of their own, and give the heavy tail usage credits with a per-member limit.

Which spending caps are hard, by vendor

By a hard cap we mean a limit that makes requests fail once spend reaches it. Alerts and budgets that only send email don’t count. Simon Willison argued on 3 October that hard caps should be the default for anything billed by usage, because coding agents make it easy to build things that run up bills.²⁰ As of 5 October 2026, every product here has a hard stop somewhere, but several leave it off until you set it, and some apply it to the whole organisation at once.

The Claude Code spending cap on API billing is a workspace limit. The first time someone signs in to Claude Code with a Claude Console account, the Console creates a “Claude Code” workspace for that usage.¹ Give it a spend limit below the organisation’s. A limit you set returns a 400 error, the tier cap returns a 429 with no retry-after header, and Claude Code’s workspace can return a 429 that does carry one.¹⁸ Code that retries on 429 needs to tell them apart.

Organisation-wide caps stop everyone at once. For that reason, Anthropic’s Enterprise guide recommends starting with group and per-user limits.²²

Team defaults differ. Cursor Teams turns on-demand usage on by default and keeps per-member limits for Enterprise. Unless an admin turns on “Only Admins Can Edit Usage Settings”, any team member can change the team-level limit.²⁶²⁷

Enforcement lags. OpenAI says its hard limits aren’t instantaneous, Cursor says usage can briefly pass a limit, and Anthropic says usage can go slightly over before it pauses.²⁴²⁶²² How a backfill bug ran up a $1,100 LLM vision bill covers why provider caps make a late backstop and which guardrails to add in your own code.

Auto-reload removes the stop. Claude’s prepaid usage credits and ChatGPT Business’s pooled credits act as a cap only until someone turns on auto-reload or automatic recharge.²¹²³

Cloud providers need their own controls. For Claude Code on Amazon Bedrock and other clouds, Anthropic’s docs point to the cloud’s budget controls, or to a self-hosted gateway with per-user spend limits.¹ AWS’s new project spend limits pause a project and stop its resources, but AWS is still releasing them to a limited number of customers.²⁸

How to reduce Claude Code token usage

Gartner advises routing simple, high-frequency tasks to smaller models, training developers to send only relevant context, and reviewing the heaviest workflows in sprint retrospectives.³ The vendors’ docs turn that advice into settings.

Default model and effort. As of 5 October 2026, Claude Code’s default model is Opus 5.5 on every Claude plan and on the API, at medium effort.¹⁹ Within one model, effort moves cost by several times. On Artificial Analysis’s test suite, Opus 5.5 cost $1.34 per task at medium, $1.82 at high, $3.46 at xhigh and $5.98 at max.²⁹ Those are general reasoning tasks, so treat the ratio as a guide: max cost 4.5 times medium. Thinking tokens are billed as output, and Anthropic says the default thinking budget can run to tens of thousands of tokens per request.¹

Admins on any plan can cap effort with the maxEffortLevel managed setting and restrict models with availableModels.¹⁹ Codex has the equivalent model_reasoning_effort setting, and OpenAI’s advice is to start at the default effort and raise it only when a task needs deeper planning.³⁰³¹ Anthropic’s docs say Sonnet handles most coding tasks well at a lower price than Opus, though the session table above puts the saving nearer a third than a half.¹

Caching. Anthropic charges 1.25 times the input price to write to its five-minute cache and 2 times for the one-hour cache. A cache read costs a tenth of the input price, or less on Opus 5.5 and Fable 5.1.¹⁷ Claude Code keeps a one-hour cache on subscriptions, which drops to five minutes once you draw on usage credits. API keys default to five minutes. The first message after a longer break reprocesses the whole context.¹ A developer on API billing who steps away for ten minutes pays to rebuild the cache when they come back. Codex credit billing has no separate charge for cache writes.¹⁰

Subagent ceilings. Subagents and agent teams send their own requests on top of the main conversation. Anthropic says agent teams use about 7 times the tokens of a standard session when teammates run in plan mode.¹ Each Claude Code subagent definition can set its own model, effort and maxTurns, so routine searches can run on a smaller model with a turn limit.³² For scripted runs, --max-budget-usd stops a print-mode run at a dollar amount. Subagent spend counts toward it, and no new subagents start once it is reached.³³ Codex caps concurrent spawned agents with agents.max_concurrent_threads_per_session.³⁰

A frontier-model sub-budget. Claude Fable 5.1 lists at 2.5 times the per-token price of Opus 5.5, and GPT-6 Astra Ultrafast uses a ChatGPT allowance at 8 times the standard rate.¹⁷¹¹ Databricks puts its premium tier at 2 to 3 times the cost of the next one down, and gives it a fixed share of each user’s monthly budget.⁶ Anthropic does something similar on Max, where Fable can take at most half of the weekly limit.¹⁵

Context hygiene. Claude Code sends the conversation with every request, so stale context costs money on every turn. Anthropic’s docs recommend /clear between unrelated tasks, a CLAUDE.md under 200 lines with workflow instructions moved into skills, and turning off MCP servers you aren’t using. They also suggest CLI tools such as gh in place of MCP servers, and hooks that cut test output down before Claude reads it.¹ Scheduled tasks send the full context each time they fire. OpenAI’s Codex guidance is the same. It suggests nesting AGENTS.md files, limiting MCP servers and giving the agent only the files it needs.¹⁰

To see which of these matter for your team, log tokens per model and per kind of work. Why our LLM pipeline reported 0 dropped shows a ledger that records input, cached, output and reasoning tokens for every call, which is the data a monthly spend review needs.

How the figures were gathered

• Prices, limits and controls come from each vendor’s pricing pages, help centre articles and documentation as they stood on 5 October 2026. Prices are in US dollars before tax, with monthly billing unless stated.
• The spend figures from Anthropic, Cursor and Databricks are self-reported. The 23% and 6% figures come from Computer Weekly’s report of Gartner Peer Insights research, and neither it nor Gartner’s press release gives a sample size.
• Real-SWE’s costs are Specific Labs’ estimates over 10 tasks with eight attempts each. Artificial Analysis’s cost per task is for its general test suite, and we use it only for the ratio between effort levels.
• The session table and the worked budgets are our arithmetic on those inputs, with the assumptions labelled. We have not measured our own spend per engineer for this post.
• Not covered: GitHub Copilot, Gemini CLI plans, enterprise discounts, taxes and regional pricing.

A spend policy template for coding agents

Copy this, replace the values in brackets, and review it after the first month.

1. Seats. Engineers who use an agent on most working days get [seat type]. Everyone else uses [lighter seat or API key]. Seat types are reviewed each quarter against usage.
2. Defaults. The default model is [model] at [effort level], enforced in managed settings. Higher effort is a per-task choice.
3. Hard limits per person. Each engineer has a monthly limit of [amount] and a daily limit of [amount]. A team lead can raise the daily limit for the rest of that day.
4. Frontier budget. Up to [share] of each monthly limit can go on the most expensive models and speed modes, such as Fable, GPT-6 Astra and Ultrafast.
5. Automation. CI jobs, scheduled agents and scripts run on API keys in their own workspace or project, with a hard monthly limit and a budget per run.
6. No unlimited settings. An unlimited credit limit or auto-reload needs sign-off from [role].
7. Alerts. The engineer and the budget owner get alerts at 50%, 75% and 90% of every limit.
8. Monthly review. Spend is reviewed by engineer, model and kind of work. Anyone over [amount] talks with their lead about what the spend produced before the limit changes.

Frequently asked questions