AI coding tool usage limits and real cost per task (Oct 2026)

Plans, published limits and what happens at the limit

"Not published" means the vendor doesn't give a number. We don't guess one.

Sources, all checked Oct 23, 2026: claude.com/pricing · Claude Code costs and limits · Claude API pricing · Codex pricing · OpenAI API pricing · Cursor models and pricing · GitHub Copilot plans · Copilot usage-based billing · Copilot model pricing · Gemini CLI quotas · Gemini API pricing · Gemini Code Assist pricing · Antigravity plans · Antigravity models. Google lists Code Assist as an hourly fee; we converted at 8,760 hours a year divided by 12. US prices, before tax.

Which reset rhythm fits how you work

• You code in bursts every day: five-hour windows (Claude Pro, Codex Plus, Antigravity on Google AI Pro) suit you. Hit the wall, take a break, come back. The weekly cap is what stops all-day, every-day use.
• You have a few heavy days a week: Cursor's weekly pool lets you burn most of it on Tuesday. One heavy day can't take your whole month, but it can take your week.
• You want a dollar number: GitHub Copilot is the only one that tells you exactly what you get ($15 of usage on Pro, $70 on Pro+, $200 on Max). The catch: it resets once a month, so a big first week can leave you dry until the 1st. And a third of Pro's credits are "flex," which GitHub says can change "as the economics of AI evolve."
• You're on a company license: Code Assist's 1,500 or 2,000 requests a day is the most predictable quota here, but it counts requests, not tokens, and Google hasn't said what happens when you run out.

Estimated cost per task Estimate

This section is our estimate, not a vendor number. No vendor publishes a per-task price. We took three task sizes, assumed token counts, and priced them at each model's published API rate. That's also what you pay once you go past a plan's limit on Claude (usage credits), Codex (API key), Cursor (third-party models), Copilot (AI credits at provider list price) and Antigravity (Google AI credits).

Why input is so large: an agent resends the conversation, files and tool output on every turn, so input dominates and most of it is a cache hit. Formula: (uncached input x input price) + (cached input x cache-read price) + (output x output price). We left out cache-write charges (Anthropic charges extra to write the cache), long-context surcharges and fast modes, so real costs can be higher.

Prices from Claude API pricing, OpenAI API pricing, Gemini API pricing and Cursor models and pricing, checked Oct 23, 2026. GitHub's Copilot model pricing lists the same per-token rates for these Claude, GPT and Gemini models. Opus and Sonnet 5.5 cache reads are 5% of the input price; Grok 4.7's are 25%.

What that means for each plan Estimate

• Copilot Pro ($10, 1,500 credits): about 22 medium tasks a month on Claude Sonnet 5.5 (68.5 credits each). Opus 5.5 isn't on Pro.
• Copilot Pro+ ($39, 7,000 credits): about 51 medium Opus 5.5 tasks, or about 100 on Sonnet 5.5. Max ($100, 20,000 credits): about 146 on Opus.
• Claude Code: Anthropic says the average enterprise cost is about $13 per developer per active day and 90% of users stay under $30. On our assumptions, $13 is about 19 medium Sonnet tasks or 9 on Opus. Past your Pro or Max limit, usage credits bill at these API rates.
• Codex: OpenAI says "a typical GPT-5.6 Sol task may use 2-15 credits," but it doesn't publish one price for credits, so we can't turn that into dollars. On an API key, GPT-6.1 Sol costs the same as Sonnet 5.5 on our assumptions.
• Cursor: Grok 4.7 has a cheaper output price than Sonnet 5.5, but its cache reads cost 5x as much ($0.50 vs $0.10), so in long, cache-heavy agent loops it comes out about 60% more expensive per task. It's inside your weekly pool, so you feel it as hitting the pool sooner, not as a bill.
• Antigravity: Gemini 4 Argon is the same price as Sonnet 5.5 until Nov 30, then doubles to Opus 5.5 pricing on Dec 1. If you pay for Google AI credit overages, expect the same per-task cost to double.

The honest takeaway: past the limit, the same model costs about the same everywhere. What you're really comparing is how big the included allowance is and how often it refills, and only GitHub and OpenAI put numbers on that.

How to stretch your limits

Everywhere

• Start a fresh session for each task. Every turn resends the whole conversation, so a long, unrelated history is the biggest hidden cost.
• Use the cheaper model for easy work. Opus 5.5 costs 2x Sonnet 5.5 per task on our estimate; Gemini 3.8 Flash and GPT-6 Luna cost a fraction.
• Keep instruction files short. CLAUDE.md, AGENTS.md and GEMINI.md load into every request.
• Turn off MCP servers you aren't using. Their tool definitions take up context on every turn.
• Be specific. "Fix the null check in auth.ts line 40" reads far fewer files than "fix the login bug."

Claude Code

• Use /clear between tasks and /compact in long ones. Check usage with /usage.
• Default to Sonnet; switch to Opus only for hard problems. Lower /effort for routine work.
• Keep CLAUDE.md under 200 lines and move rarely needed instructions into skills, per Anthropic's docs.
• Use plan mode before big changes, and subagents or hooks to keep noisy output out of the main context. Agent teams use about 7x the tokens of a single session, Anthropic says.

OpenAI Codex

• Check what's left with /status. Keep AGENTS.md small and point Codex at specific files.
• Fast mode uses 2.5x your included usage and Ultrafast 8x, per OpenAI. Leave them off unless you need the speed.
• GPT-6 Luna's published range (350 to 3,000 messages per five hours on Plus) is about 20x GPT-6.1 Sol's, so use Luna for simple edits.

Cursor

• Composer 2.5 costs about a third of Grok 4.7 per task on our estimate, so it stretches the weekly pool furthest.
• Set a monthly cap on prepaid credit before you use Claude or GPT models on Premium and up.

GitHub Copilot

• Use auto model selection: GitHub gives a 10% discount on it (30% on Max through Oct 28, 2026).
• Completions and next edit suggestions are unlimited on paid plans and don't use credits, so lean on them for small edits.
• Set a dollar budget so going over is a choice, and remember the reset is the 1st at 00:00 UTC, not your billing date.

Gemini CLI and Antigravity

• In Gemini CLI, /stats model shows what you've used.
• In Antigravity, Gemini and Claude/GPT models draw from separate pools. When one runs dry, switch families instead of waiting.
• Leave AI Credit Overages on Never unless you want to pay per token.

Tips from Claude Code costs docs, Codex pricing, Cursor docs, Copilot billing docs, Gemini CLI quotas and Antigravity models, checked Oct 23, 2026.

FAQ

Which AI coding plan gives the most usage for $20?

Nobody publishes a like-for-like number, so there is no honest single winner. OpenAI is the only vendor that publishes ranges for its $20 plan: ChatGPT Plus gets an estimated 15 to 160 GPT-6.1 Sol local messages per five hours in Codex. GitHub publishes a dollar budget instead: Copilot Pro at $10 includes 1,500 AI credits (worth $15 at $0.01 each) per month. Anthropic (Claude Pro $20) and Google (Google AI Pro $19.99 in Antigravity) publish only the reset rhythm, not the size of the allowance. Cursor publishes multipliers (Premium is 4x Plus) but not the size of the Plus pool.

What happens when I hit my Claude Code limit?

Claude Code tells you whether you hit the five-hour session limit, the weekly limit, or a per-model (Opus or Sonnet) limit. You can wait for the reset, upgrade to Max, or turn on usage credits, which are billed at standard API rates. You can see your usage with /usage in Claude Code or under Settings > Usage.

Do Codex limits stop me mid-task?

OpenAI says that when you reach a limit the current turn can finish. After that you can wait for the reset, buy ChatGPT credits on Plus and Pro, or switch Codex to an API key and pay API rates. OpenAI says the price of credits can vary by account, region and plan, so it doesn't publish one price. Check your usage with /status.

When do GitHub Copilot AI credits reset?

At 00:00 UTC on the first day of every calendar month, whatever your billing date is. Unused credits don't roll over. If you run out you can upgrade (you pay only the price difference), set a dollar budget for additional usage at $0.01 per credit, or wait for the reset.

How much does one AI coding task cost?

It depends on the model and the size of the task, and no vendor publishes a per-task price. Our estimate, at published API prices, for a medium task (1.5 million input tokens, 90% of them cached, and 25,000 output tokens) is about $0.69 on Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon (before Dec 1, 2026), about $1.37 on Claude Opus 5.5, and about $1.13 on Grok 4.7 in Cursor. Anthropic says the average Claude Code cost is about $13 per developer per active day. These are estimates; your token counts will differ.

Does Copilot Pro include Claude Opus 5.5?

No. GitHub's plans page lists Claude Opus 5.5 on Copilot Pro+, Max, Business and Enterprise, not on Copilot Pro. Copilot Pro at $10 includes Claude Sonnet 5.5, Sonnet 5 and Haiku 5.5, among others.

Related guides

• AI coding tool pricingAll 40 plans + cost calculator
• Claude Code vs OpenAI CodexSame prices, different limits and uptime
• GitHub Copilot vs Claude Code$10 credits vs $20 flat
• Gemini CLI vs Claude CodeGemini CLI is API-key only for individuals now
• Cursor vs Google Antigravity$19.99 Gemini + Claude vs $30 + API
• Claude Code vs CursorTerminal agent vs AI IDE

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.