How Much Does an AI Coding Agent Cost Per Month? 2026 Total Cost of Subscriptions, Tokens, API, and Servers for Claude Code, Codex, Gemini, and Cursor

Another $20 line shows up on the card. A teammate says an agent is “just one subscription.” Finance asks what to budget per head. Then you open the usage page: the same $20 leaves one person with leftover quota and another buying overage — plus a machine that never sleeps.

In 2026 an AI coding agent is rarely a single charge. Subscription, tokens, API, and the server are four ledgers. How windows and weekly caps work is in the Claude Code 2026 usage limits guide; product capability is in the Claude Code vs Codex vs Gemini 3.8 Flash comparison. This piece only adds up the money.

What actually shows up on the bill

Do not start with “whose Pro is cheaper.” Four products’ $20 do not buy the same thing.

Claude Code and Codex behave like “a window plus a weekly bucket”: you stop first, you do not always pay extra. Cursor behaves like a dollar credit pool: picking a frontier model, Max mode, or a huge context carves the pool; empty it and on-demand billing can start. On Gemini, split the invoice: model list price is not the same line as an agent seat.

Figures follow September 2026 public list prices. FX, tax, and promos move — check the vendor page before you buy.

What the sticker covers — and what it does not

How you read it matters more than the sticker:

• Claude Code: no personal plan under $20. Max only multiplies the same windows by 5 or 20 — it does not remove the ceiling. Team Standard (~$20/seat) does not include Claude Code; CLI for everyone means Team Premium (~$100/seat, often 5-seat minimum). Details stay in the limits guide.
• Codex: still the easiest way to get a full terminal agent under twenty dollars. If you already pay ChatGPT Plus, Codex is usually not a second line item.
• Gemini: cheap Flash is not a cheap developer seat. Antigravity, Gemini Enterprise Agent Platform, and Code Assist seats are extra. Wiring Flash into a multi-backend CLI such as Codex is the usual “cheap brain, mature shell” move.
• Cursor: Pro’s $20 buys IDE orchestration and credits, not unlimited Opus. On the same Pro plan, Auto can last a month — or a week of hand-picked flagship + Max dumps you into on-demand. For Cursor-only math, see the Cursor Pro real-cost analysis.

Tokens and API: the same task, several multiples

The subscription is the ceiling; tokens are the floor. API list prices (per million tokens, September 2026):

A mid-size agent turn at 80k input + 8k output is roughly $0.09 on Flash, $0.24 on Sonnet 5, $0.48 on Sol, $0.60 on Opus 5. Cache hits often drop input another order of magnitude.

A full day is easier to feel. Anthropic has published a ballpark: Claude Code developers average about $6/day, with 90% under $12. Over 22 workdays that is $130–260 on the model alone — the shadow cost if you priced everything at API rates. Max 20x at $200/month can be a bargain for someone who fills the window (as if buying $600–1,500 of tokens at a flat fee) and a markup for someone who codes a few hours a week.

Cursor overage is a different pipe: if on-demand stays on after credits hit zero, you keep paying API rates for the model you picked. The community pattern is Pro drifting from $20 to $40–80 — not because the model got pricier, but because huge context and Max mode punched through the pool.

# Same “fix the test + open a PR”
Flash / Terra / Auto     → scout, write a repro script
Sonnet / Sol             → change the public API, land the diff
Opus / high effort       → cross-module refactors only

Why a server lands on the monthly bill

Tokens are metered in the cloud. The repo, tests, MCP, and sandbox run on the machine in front of you. When that machine sleeps, three things happen together: the session dies, the sandbox clears, Prompt Cache expires. The next “continue” is billed as full input. The money you saved by not upgrading a plan may not cover those two warm-ups.

More teams in 2026 put an “always-on agent box” on its own line. Order-of-magnitude only (not a quote):

The machine is not a fourth “AI subscription,” but it decides whether the first three are wasted. Unified memory on Apple Silicon is a good fit for indexing a large repo; macOS ships Unix, Homebrew, Docker, and SSH, so Claude Code / Codex terminal loops skip a WSL layer. Idle power is about 4W — closer to a background worker than a gaming laptop left on all night.

Three usage levels, added up

These are stacked monthly estimates, not recommended stacks. FX and overage are treated as zero; the machine is either always-on or not.

Easy ways to double-count:

1. Already on ChatGPT Plus — adding $20 “for Codex” is the same seat twice.
2. Cursor Ultra + Claude Max 20x are both $200 and overlap a lot. Unless IDE and terminal must both run flat-out, raise only one rail first.
3. Do not default the whole team to Premium. Reviewers and PMs can sit on light seats; people who write code take Claude Code / Cursor Teams.
4. Uncapped API is how accident bills happen. Public write-ups have put per-engineer Claude Code API spend in the hundreds or thousands per month. A subscription’s job is to cut that tail.

How to keep the month from running away

1. Pick a primary shell, then buy quota. Long terminal jobs → Claude Code or Codex; in-editor work → Cursor. Owning both is fine; taking both to 20× is not.
2. Default cheap. /model sonnet, Cursor Auto, Flash effort=low should be daily; Opus / Sol / high effort are exceptions.
3. /clear before a new task. /compact is itself a large request. Tool logs, full-table dumps, and 6,000-line files burn more than the model sticker.
4. Cap credits and API keys. Turn off blind Cursor on-demand; set a monthly API budget. Hitting a subscription wall pauses you. Hitting an API wall bills you.
5. Keep the machine awake. Closing the lid for an hour to save electricity can expire Prompt Cache and recharge the next turn at full input — the worst “save power, burn tokens” trade.

# In-session: cheap model + low effort, then switch
/model sonnet
/effort low
/rename api-auth-fix
/clear

FAQ

Is a single $20 plan enough for a full-time job?
Yes, if you mostly complete, ask about one file, and run an agent now and then. Multi-session days, large-repo sweeps, and subagents will empty a Pro 5-hour window or credit pool first. Stepping to Max 5x / Pro+ usually beats stacking two $20 plans.

When credits run out, buy more or switch to API?
Want a predictable number? Upgrade the subscription. Want spikes, non-interactive CI, and your own budget cap? Use the API and set a monthly ceiling. Usage credits on Pro often beat opening a second subscription for rare overflow.

Can Gemini 3.8 Flash be “free Claude Code”?
No. It is a model. The terminal loop, permissions, and MCP still need a shell. Individual Code Assist has tightened; do not budget from 2025-era free-tier memories.

Should every teammate get Max / Ultra?
No. See who actually runs long agents. Two Max 5x seats ($200) often beat one Max 20x for parallel work — you bought two independent windows.