Every "which coding agent is best" post I read ranks accuracy, plugin support, and vibes. Almost none of them show a monthly bill. This is the post I wanted to read before I picked one.
For the 31 days of June 2026 I ran all three official coding agents on the same laptop: Claude Code (Anthropic), Codex (OpenAI, inside ChatGPT), and Cursor. Same repos, same tasks, same me. I kept receipts. I logged sessions. I ran a local Qwen 3.5 35B on an RTX 4070 in parallel to see where owning silicon paid back.
Short answer: one of them is cheapest for me right now, but not by the margin the marketing pages suggest, and if my usage shape changes even a little the ranking flips. That is the real headline. The bill depends on what you do more than on which logo you pick.
What "one month" was, in detail
The month I logged:
- 22 working days at the keyboard.
- Roughly 6 hours a day inside an agent, mixed reading + writing.
- Two projects: one TypeScript SaaS with ~180 files, one Python data pipeline with ~60.
- Three "hard" refactors (multi-file, week-scale). The rest was normal feature and bugfix work.
Usage shape matters more than the number on the pricing page. A part-time hobbyist and a full-time engineer buying the same $200 tier are buying two different products at the same price. The line I care about is not $/month, it is $/hour-of-agent-time.
The three cost shapes
Every coding agent bills you as one of three shapes, and June 2026 has not changed that.
- Subscription with usage multiplier. Flat fee, soft cap on requests, throttle when you're over. Claude Code Pro / Max, ChatGPT Plus / Pro, Cursor Pro / Ultra all fit here.
- Metered API. Per token. No monthly ceiling unless you set one.
- Local GPU. You bought the card. Electricity + amortization. Zero variable cost per token but a fixed capacity ceiling.
The break-even between these three is what most posts skip. My June numbers, in one table:
Claude Code: $200/month Max, and I hit the throttle twice
I ran Claude Code on the Max 20x plan at $200/month, per Anthropic's current pricing page. The Pro tier at $20/month exists but is calibrated for a few focused sessions a day, not an agent-driven workflow.
What the month bought me:
- Claude Code was my primary tool for the two hard refactors. Long-context is where it earns the bill.
- I hit the 5-hour rolling limit twice on Max, both times during a deep multi-file refactor where I was running Sonnet 4.6 in the background across 3 subagents. That is Anthropic's cost tell: subagents multiply token spend, and the Max tier feels the cap when you fan out.
- Everything else fit comfortably.
If I had run the same month on the metered API at Sonnet 4.6 pricing of $3/M input and $15/M output (the introductory $2/$10 promo runs through August 31, 2026, so I'm quoting the standard rate), the honest estimate for my token volume landed somewhere in the $260-$380 range. The subscription won by ~$60-$180, at the cost of a rate limit I can predict but not remove.
Specific lesson: for a heavy user, Max 20x beats the metered API only until you fan out subagents. Anthropic's own engineering blog notes that multi-agent runs use roughly 15x the tokens of a single-agent chat. That's real. If I ran three subagents on Sonnet as my normal shape, the metered path would probably beat Max.
Codex: bundled inside ChatGPT Pro at $100, and the metering just changed
Codex is the odd one. It does not have a standalone subscription. It rides inside your ChatGPT plan.
I ran ChatGPT Pro at $100/month, the tier OpenAI added in April 2026 to sit between the $20 Plus and the $200 Pro-20x. Two things the marketing page does not emphasize:
- On April 2, 2026, Codex pricing moved from per-message to API-token-equivalent metering. If you were used to the old plan, your "same amount of work" started drawing down credits at a different rate. Nobody's monthly bill stayed the same, they just moved.
- The average Codex developer sits at roughly $100-$200/month across all instances they're running. That is the same number I hit. Bundling with ChatGPT means I paid one bill, but the token math is not hidden. It is just billed as one line item.
Where Codex earned the seat for me: quick front-end scaffolding, one-shot script generation, "explain this stack trace" flows where I want an answer inside a browser tab I already have open. Where it did not: long-lived agent sessions in a real repo. That is where Claude Code held the line.
Cursor: $60/month Pro+, and I stopped forcing frontier models
Cursor's June 2025 credit model change means the plan you buy is really a credit pool. Hobby is $0, Pro is $20, Pro+ is $60, Ultra is $200. I paid for Pro+ this month.
The move Cursor pushed on me was Auto mode. On any paid plan Auto is unlimited: it picks a model for you and does not touch your credits. If you leave Auto on for tab completion and short edits and only reach for Claude Sonnet or GPT-4o for the actual reasoning work, the credit pool lasts.
I did the opposite for the first two weeks and paid for it. I forced Claude Sonnet everywhere, my credits drained by day 18, and I bought a top-up. The last two weeks I flipped the pattern: Auto for autocomplete, frontier models only when I explicitly asked. Same Pro+ pool covered the rest of the month with room to spare.
The Cursor-specific insight nobody puts on the pricing page: Pro at $20 is fine if you use Auto for most things. Pro+ at $60 is where you land if you cannot help yourself and keep clicking Sonnet. Ultra at $200 is for someone running agent mode all day; that was not me this month.
The local GPU: RTX 4070 + Qwen 3.5 35B, the breakeven that is not what you think
I keep an RTX 4070 running Qwen 3.5 35B locally, and I ran it as a fourth "agent" in parallel to see which jobs made sense to send there.
The naive breakeven math is easy and misleading. A $600 card amortized over 24 months is $25/month. Electricity at moderate use is another $10-$15. On paper: $35-$40/month for unlimited tokens.
Here is what the naive math misses:
- The 4070 cannot run the frontier models. It runs the small ones, competently but at ~40% of the quality of Sonnet 4.6 on the tasks I care about. For me, that eliminated it from primary agent duty.
- Where it did earn its slot: bulk classification, refactor precheck, and offline batch jobs. I ran a 3,000-file "which files touch payment logic" scan on Qwen overnight; the same job on Claude API would have been $8-$12. Twenty of those a month is where the card breaks even.
- Zero of my long-context refactor sessions belonged on local. The context window on Qwen 3.5 35B is not comparable to Sonnet 4.6.
Local is not a subscription replacement. It is a batch-workload absorber. Buying a 4070 to save money on your primary agent is buying the wrong tool for the wrong job.
The 31-day scoreboard
Here is what landed on the credit card and what I got for it.
| Tool | Plan | Monthly cost | Where it earned it | Where it did not |
|---|---|---|---|---|
| Claude Code | Max 20x | $200 | Long refactors, subagent runs | Rate-limited twice on fan-out |
| Codex | ChatGPT Pro $100 | $100 | Quick scaffolding, browser-tab flow | Multi-file agent sessions |
| Cursor | Pro+ | $60 | Autocomplete + Auto mode | Credit drain if you force Sonnet |
| Local (Qwen 3.5 35B) | RTX 4070 amortized | ~$35 | Bulk scans, batch classification | Long context, frontier quality |
Total this month: $395. Yes, all four. The overlap is not wasted; each tool covered a workload the others charged 3-5x more to do.
Now the honest bit. If I had to pick one and drop the rest:
- If my month were 90% multi-file agent refactors and I could tolerate a rate limit, I would keep Claude Code Max 20x and drop the other three. Roughly $200/month.
- If my month were 90% "one file at a time, browser tab open," I would keep ChatGPT Pro at $100 and use its Codex allocation. Roughly $100/month.
- If my month were 90% autocomplete-first coding with a couple of hard problems a week, I would keep Cursor Pro at $20 with Auto default and pay-as-I-go for Sonnet. Roughly $20-$40/month.
The reason I keep all three is that none of my months look like "90% one shape." Yours probably do not either. But if you have to pick, pick against your dominant workload, not against the benchmark you saw on Twitter.
What I will drop for July
Two changes I'm making for the July run.
Cap the Claude Code subagent fanout at two. I hit the throttle twice this month, both times because I told it to run three subagents in parallel. Anthropic's own doc reports 15x token spend on multi-agent runs; I don't need three, two solves most of the problem, and I stay under the rate limit.
Stop reaching for Sonnet in Cursor. Pro at $20 with Auto default was clearly the right tier for me. The $40/month I saved on Cursor pays for something more useful than a habit.
If I write this post again at the end of July I'll compare the two months side by side. The Twitter version of "which agent is cheapest?" is the wrong question. The version worth answering: what is my usage shape, and which of these four bills matches it?
That answer changed for me twice this month. It will probably change again. But I'll have receipts.
The book I extracted the receipt-first playbook from
If you liked the receipt-first framing, I turned the general "estimate cost before shipping the prototype" mindset into a full playbook in Claude Code Mastery. It covers the subagent fanout math, when Skills earn their token cost, and how to keep a heavy month from turning into a surprise. Same posture as this post: real numbers, then decisions.
Sources
- Claude Code / Anthropic API pricing: claude.com/pricing.
- ChatGPT / Codex pricing: chatgpt.com/pricing and developers.openai.com/codex/pricing.
- Cursor pricing: cursor.com/pricing.
- Anthropic multi-agent token math: Anthropic Engineering — Multi-Agent Research System.
- Simon Willison publishes ongoing coding-agent usage-shape breakdowns worth reading alongside: simonwillison.net.
Top comments (0)