What one AI coding task really costs (Oct 2026)
Short answer: an everyday task, like adding a feature across a few files, is usually under $1.50 at API prices, by our estimate. A hard task is a different story. Artificial Analysis ran 290 tough benchmark tasks through each coding agent and measured $11.42 a task for Claude Code with Opus 5.5, $13.30 with Sonnet 5.5, $11.73 for Antigravity CLI with Gemini 4 Argon and $1.50 for Codex with GPT-6.1 Sol. The thing that moves the bill most isn't the price per token. It's how many tokens the agent burns to finish.
Benchmark figures from Artificial Analysis's Coding Agent Index v2.0 and API prices from each vendor's pricing page, all checked Oct 29, 2026. Anything we worked out ourselves is labeled as an estimate.
Measured: cost per task on hard benchmark tasks
These are Artificial Analysis's numbers, not ours. Each task was tried three times, and cost is the average at pay-per-token API prices, counting cache reads and cache writes. It is not what a Claude, ChatGPT or Google plan charges you.
| Agent and model (effort) | Index score | Cost per task | Tokens per task | Time per task | Works out to, per 1M tokens | List price per 1M tokens (in / out) |
|---|---|---|---|---|---|---|
| Claude Code, Claude Opus 5.5 (max) | 62 | $11.42 | 14.5M | 1.0 h | $0.79 | $4 / $20 |
| Claude Code, Claude Sonnet 5.5 (max) | 59 | $13.30 | 24.1M | 1.3 h | $0.55 | $2 / $10 |
| Codex, GPT-6 Astra (max) | 55 | $7.04 | 3.2M | 24.8 min | $2.20 | $10 / $50 |
| Codex, GPT-6.1 Sol (max) | 54 | $1.50 | 3.9M | 23.7 min | $0.38 | $2 / $10 |
| Antigravity CLI, Gemini 4 Argon (high) | 53 | $11.73 | 13.9M | 34.7 min | $0.84 | $2 / $10 through Nov 30, then $4 / $20 |
| Codex, DeepSeek V4.1 Flash (max) | 47 | $0.39 | 27.2M | 38.1 min | $0.01 | Not on our price list |
| Claude Code, Claude Haiku 5.5 (max) | 37 | $2.67 | 16.1M | 1.2 h | $0.17 | $0.10 / $0.50 up to 100K-token prompts, $0.50 / $2.50 above |
| Codex, GPT-6 Luna (max) | 31 | $0.16 | 8.9M | 21.3 min | $0.02 | $0.10 / $0.50 |
Index score, cost, tokens and time from Artificial Analysis: Coding Agent Index, Claude Code vs Codex and Antigravity CLI vs Claude Code, checked Oct 29, 2026. "Works out to" is our math: cost per task divided by tokens per task. List prices from Claude API pricing, OpenAI API pricing and Gemini API pricing. The Claude Code runs use Anthropic's default fallback, so a different model can finish an attempt after a safety refusal.
What the tasks are: 100 real software tasks on open-source repos from Cognition's FrontierCode v1.1, graded on whether a maintainer would merge the change; 66 terminal tasks from Terminal-Bench 4.0; and 124 questions about code repos from Scale AI's SWE-Atlas-QnA, regraded by Artificial Analysis. So think "a gnarly ticket," not "rename this variable."
Why our estimate says $0.69 and the benchmark says $11.42
Our usage limits page prices a medium task (a feature across 3 to 5 files, with tests) at about $0.69 on Sonnet 5.5 and $1.37 on Opus 5.5. That's not wrong, it's a smaller task. Here's the gap, piece by piece:
- Task size, by far the biggest part. Our medium task assumes about 1.5M tokens. The benchmark tasks averaged 3.2M to 27.2M depending on the setup, and 14.5M for Claude Code with Opus 5.5. That's about 10 times our medium task.
- Per token, the math lines up. Claude Code with Opus 5.5 came out at $0.79 per million tokens on the benchmark. Our estimate works out to $0.90, a bit higher because we assumed 90% of input is cached and Artificial Analysis measured 95%. Codex with GPT-6.1 Sol: $0.38 measured vs $0.45 in our math. So the per-token pricing isn't the problem. The token count is.
- Max effort. Most setups ran at max or high effort, which means more thinking tokens. Your default may be lower.
- Long, chatty runs. Claude Code with Opus 5.5 averaged 146 turns and an hour per task. Codex with GPT-6 Astra averaged 39 turns and 25 minutes. Every turn resends the context.
- Cache writes. Artificial Analysis counts them. We left them out, so our numbers run a little low on Claude.
- API price, not plan price. On Claude Pro or Max, ChatGPT Plus or Google AI Pro you don't pay this per task. It comes out of your allowance instead.
Our take: keep using $0.10 to $1.40 as the cost of an everyday task, and budget roughly $1.50 to $13 at API prices for a hard, hour-long one, depending a lot on the agent and model.
Three things that surprised us
The cheaper Claude model cost more
Claude Sonnet 5.5 is half the price of Opus 5.5 per token. But on these tasks Claude Code with Sonnet 5.5 used 24.1M tokens to Opus's 14.5M, took longer (1.3 hours vs 1.0) and cost more ($13.30 vs $11.42), with a lower score. On hard work, Opus can be the cheaper pick.
Same list price, 9x the bill
GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2 in and $10 out per million tokens. Codex with Sol finished tasks on 3.9M tokens for $1.50. Claude Code with Sonnet used 24.1M and cost $13.30. That's different agents as well as different models, so it isn't a pure model test, but it shows how much the agent's habits matter.
Haiku 5.5 isn't as cheap as it looks
Claude Haiku 5.5 lists at $0.10 in and $0.50 out, the same as GPT-6 Luna. But Anthropic bills any request with a prompt over 100,000 tokens at 5x that ($0.50 / $2.50), and agent sessions pass 100K fast. On the benchmark, Claude Code with Haiku 5.5 cost $2.67 a task. Codex with Luna cost $0.16.
Haiku's two-tier pricing is on Claude API pricing. OpenAI also doubles GPT-6 input prices above 272K tokens (OpenAI API pricing).
A note on Gemini 4 Argon's $11.73
Google's API price for Gemini 4 Argon is $2 in and $10 out per million tokens through Nov 30, 2026, then $4 and $20 from Dec 1. Artificial Analysis says it uses current API pricing, but its $11.73 is almost exactly double the $5.86 an earlier snapshot of the index showed for the same 13.9M tokens, which matches the Dec 1 price. Our read: through Nov 30, the same run would cost about half on the API. Artificial Analysis doesn't spell out which price it used, so treat that as our inference.
In Antigravity itself you don't pay per token. Google AI Pro ($19.99) includes Argon, but on Pro it has its own weekly-only quota, with no five-hour refresh. See Cursor vs Antigravity for the rollout and limits.
Sources: Gemini API pricing, Google's Gemini 4 Argon announcement, and the earlier index snapshot as archived by aicharts.io (retrieved Oct 27, 2026), all checked Oct 29, 2026.
What that means for your plan Estimate
- Paying per token (your own key in Cline, Kilo Code, OpenCode or Aider, Claude usage credits past your limit, or a Codex API key): $20 covers less than 2 benchmark-size Opus 5.5 tasks in Claude Code, or about 13 GPT-6.1 Sol tasks in Codex. Everyday tasks cost far less, about $0.69 each on Sonnet 5.5 by our estimate.
- Claude Code for work: across enterprise deployments, Anthropic says the average Claude Code cost is about $13 per developer per active day, and 90% of users stay under $30. $13 is about one benchmark-size Opus task, or about 19 of our medium Sonnet tasks.
- GitHub Copilot: AI credits are $0.01 each at the provider's list price. Pro+ has 7,000 credits ($70 of usage). If a Copilot run burned as many tokens as Claude Code did on the benchmark, that's about 6 hard Opus 5.5 tasks. Artificial Analysis didn't test the same setup in Copilot, so this is a rough guide only.
- Usage-priced plans like Warp Build or Amp Megawatt include $20 of usage at API rates, so the same math applies: lots of small tasks, or one or two huge ones.
Plan prices and credits from our usage limits page and Claude Code alternatives, checked Oct 23 to 28, 2026. Anthropic's daily cost figure is from Claude Code costs docs.
How to cut your cost per task
- Don't default to max effort. On the benchmark, Codex with GPT-6.1 Terra scored 48 at xhigh for $0.69 a task and 48 at max for $1.24. Same score, almost double the cost. At low effort it was $0.20 with a score of 35.
- On hard tasks, try the bigger model. Opus 5.5 finished in fewer tokens than Sonnet 5.5 and cost less per task.
- Start fresh for each task. Every turn resends the conversation, so long sessions with old history pile up input tokens.
- Watch long prompts on small models. Haiku 5.5's price jumps 5x past 100K tokens, and GPT-6 input doubles past 272K.
- Use a plan for heavy days. If you regularly run hour-long agent sessions, a subscription allowance is usually cheaper than paying API rates per task.
Terra figures from Artificial Analysis: Claude Code vs Codex, checked Oct 29, 2026.
Ready to try one?
Every tool here can be cancelled monthly, so the fastest way to decide is a real task on each.
Claude Code
OpenAI Codex
Google Antigravity
GitHub Copilot
FAQ
How much does one AI coding task cost?
At API prices, an everyday task like a feature across 3 to 5 files is about $0.69 on Claude Sonnet 5.5 or GPT-6.1 Sol and about $1.37 on Claude Opus 5.5, by our estimate. Hard tasks cost much more: on Artificial Analysis's Coding Agent Index (checked Oct 29, 2026), Claude Code with Opus 5.5 averaged $11.42 a task, Claude Code with Sonnet 5.5 $13.30, Antigravity CLI with Gemini 4 Argon $11.73 and Codex with GPT-6.1 Sol $1.50.
Why does Artificial Analysis say a Claude Code task costs $11.42?
Because its tasks are big. They are hard benchmark tasks (real repo changes graded on whether a maintainer would merge them, terminal tasks and code Q&A), run at max effort, and Claude Code with Opus 5.5 averaged 14.5 million tokens, 146 turns and about an hour per task. That is about 10 times our medium task. Per token, the cost is close to our math: $0.79 per million tokens measured vs $0.90 in our estimate. The figure is at API prices, not a Claude plan price.
Is Claude Sonnet 5.5 cheaper than Opus 5.5 per task?
Per token, yes, it's half the price. Per hard task, not on Artificial Analysis's benchmark: Claude Code with Sonnet 5.5 used 24.1 million tokens and cost $13.30 a task, while Opus 5.5 used 14.5 million and cost $11.42, with a higher score.
Which coding agent is cheapest per task?
Among strong setups on Artificial Analysis's index, Codex with GPT-6.1 Sol: a score of 54 for $1.50 a task. Cheaper options score lower, such as Codex with DeepSeek V4.1 Flash (47, $0.39) or GPT-6 Luna (31, $0.16). Checked Oct 29, 2026.
Do I pay these prices on a Claude, ChatGPT or Google AI subscription?
No. Those figures are pay-per-token API prices. On Claude Pro or Max, ChatGPT Plus or Google AI Pro, usage comes out of your plan's allowance. You pay API rates when you use your own API key (for example in Cline, Kilo Code, OpenCode or Aider), buy Claude usage credits past your limit, or use a Codex API key.
How much does Gemini 4 Argon cost per task?
Artificial Analysis lists $11.73 a task for Antigravity CLI with Gemini 4 Argon (high), on 13.9 million tokens. That looks like Google's standard price of $4 in and $20 out per million tokens, which starts Dec 1, 2026; through Nov 30 the API price is half that. In Antigravity, Google AI Pro ($19.99) includes Argon on a separate weekly-only quota.
Ship your app: the rest of the stack
Picked your AI coding tool? These are the usual next costs. Free tiers first, and prices checked Oct 20, 2026 on each vendor's pricing page.
Code review
- CodeRabbitAI review on every pull request, in GitHub or GitLabFree for public repos; Essentials $24/user/mo billed annually ($30 monthly)
Hosting
- VercelNext.js and front-end appsHobby free (personal, non-commercial); Pro $20/mo per developer seat
- NetlifyStatic sites and front ends, unlimited Pro membersFree; Personal $9/mo; Pro $20/mo with unlimited members
- RailwayBack ends, workers and databases with no server setupFree trial $5 credit; Hobby $5/mo incl. $5 usage; Pro $20/mo
- RenderHeroku-style web services, cron jobs and PostgresHobby $0 + compute; Pro $25/mo + compute
Database and auth
Domain name
- Cloudflare RegistrarDomains at cost with no markupWholesale price per TLD
- PorkbunCheap domains with free WHOIS privacyVaries by TLD
- NamecheapPopular registrar with hosting and email add-onsVaries by TLD
Editorial disclosure
Rankings are not for sale. Sponsored placements and affiliate links are always labeled, and they never change our recommendations, quiz results, prices, or comparison data. None of the links on this page are affiliate links: no vendor here pays us, and pricing source links are never affiliate links. Editorial policy.