Every AI coding tool launches with a simple pricing page: $20/month, unlimited everything. Six months later, the page has changed. Now there are tiers. Usage caps. Overage charges. “Credits.”
This pattern is so consistent it’s predictable. Here’s exactly how it plays out, what it costs you, and how to avoid getting caught.
The Bait and Switch
The playbook is identical across every major player:
Cursor launched at $20/month for unlimited completions. Today: Free, Pro $20, Pro+ $60, Ultra $200. Each tier has different usage limits and model access. The “unlimited” language quietly moved from the main page to an asterisked footnote.
GitHub Copilot launched at $10/month for unlimited completions. Today: Free with 2K completions, Pro $10, Pro+ $39, Business $19, Enterprise $39. In June 2026, they migrated to “AI credits” — a system where different operations consume different credit amounts, making actual costs opaque.
Claude Code launched at $20/month with Claude Pro. Today: Pro $20, Max $100, Max Premium $200. The Pro tier now shares usage limits between chat and Claude Code, meaning a heavy coding session eats into your chat quota.
Every single one followed the same playbook: launch with an attractive flat price, build dependency on the tool, then introduce usage-based pricing once switching costs are high.
The Token Economics They Don’t Explain
The reason usage-based pricing is inevitable: token costs are real and variable. Here’s what a typical coding session actually costs in raw API terms:
| Activity | Approximate tokens | Cost (Claude Sonnet 4 at $3/M input) |
|---|---|---|
| Simple completion | 2K-5K | $0.006-$0.015 |
| Code review of 200-line PR | 15K-25K | $0.045-$0.075 |
| Multi-file refactor | 50K-100K | $0.15-$0.30 |
| Full codebase context analysis | 100K-200K | $0.30-$0.60 |
| Extended agentic session (8h) | 500K-2M | $1.50-$6.00 |
A developer running Claude Code for 8 hours a day consumes roughly $3-6 in raw API costs. A developer using it for an hour consumes $0.50-1.00. That’s a 5x difference in cost between your heaviest and lightest users — on the same $20/month subscription.
The flat-price model forces light users to subsidize heavy users. Companies tolerate this during the growth phase when user acquisition matters more than revenue per user. Once they need to show unit economics to investors, usage-based pricing arrives.
The Overage Trap
The sneakiest version of this isn’t a price increase — it’s the overage charge. Here’s how it works:
- You pay $20/month for “500 premium requests”
- Each request can be a completion, a chat message, or a review
- Premium requests cost more when you use expensive models (Claude Opus, GPT-4.5)
- You get a surprise bill for $30-80 in overages at the end of the month
GitHub Copilot’s credit system is the latest example. Credits aren’t fungible — using Copilot Chat consumes credits at a different rate than code completions. The $10/month Pro plan gives you a pool of credits, but the actual number of operations you can perform depends on which features you use and which models you select.
The fix? Set hard spending limits. Every provider that offers usage-based pricing also offers budget caps. Set them at 120% of what you expect to spend. Yes, you’ll occasionally hit the cap mid-day. That’s better than a surprise $200 bill.
Annual Cost Projections
Here’s what each tool actually costs over 12 months for a developer using it daily:
| Tool | Monthly | Annual | Effective $/day |
|---|---|---|---|
| GitHub Copilot Pro | $10 | $120 | $0.47 |
| GitHub Copilot Pro+ | $39 | $468 | $1.84 |
| Cursor Pro | $20 | $240 | $0.94 |
| Cursor Ultra | $200 | $2,400 | $9.43 |
| Claude Code (Max) | $100 | $1,200 | $4.71 |
| Hermes Agent (BYO keys) | $0 + API | ~$180-$360 | $0.70-$1.41 |
| Oh My Pi (BYO keys) | $0 + API | ~$24-$360 | $0.09-$1.41 |
| Gitlawb Zero (BYO keys) | $0 + API | ~$24-$360 | $0.09-$1.41 |
The subscription tools have predictable monthly costs but unpredictable annual costs as tiers shift. The open-source tools have unpredictable monthly costs but usually cheaper annual costs because you’re buying tokens at wholesale rates with no middleman markup.
Who’s Doing Pricing Right
The honest pricing tools share one trait: no VC pressure to grow revenue quarter-over-quarter.
Hermes Agent — free, open source, you pay for your own API keys. Your cost scales with your usage because you’re buying tokens directly from providers. No subscription tier to manage. No overage surprises.
Oh My Pi — free tool, BYO API keys. $2-30/month in API costs depending on how aggressively you use it. Same token economics, no platform markup.
Gitlawb Zero — free, open source, BYO keys. No subscription at all.
The pattern is clear: open-source tools have honest pricing because their business model isn’t your subscription revenue. Commercial tools have complex pricing because they need to show growing average revenue per user to investors.
The Migration Playbook
If you’re currently paying for a commercial tool and considering switching to an OSS alternative with BYO API keys:
- Start with parallel usage. Run both tools for two weeks. Compare output quality on your actual work, not benchmark results.
- Estimate your token consumption. Run the OSS tool for a week, check your API dashboard for actual usage. Most people overestimate by 2-3x because they assume agentic mode costs more than completions.
- Set up budget alerts. Every major API provider (Anthropic, OpenAI, Google) offers spending alerts. Set them at $50/month initially.
- Migrate incrementally. Switch one workflow at a time — code review first, then debugging, then refactoring. Don’t try to replace everything in a weekend.
The biggest friction point isn’t the tool switch — it’s the model selection. Each provider’s models have different strengths. Claude excels at long-context analysis. GPT-4.5 handles structured data well. Gemini has the best context window for large codebases. Open-source models (Llama, Qwen, DeepSeek) are getting competitive for routine completions.
The Prediction
In 12 months, there will be no “unlimited” AI coding tool subscriptions. Every commercial tool will have usage-based pricing. The only question is whether they grandfather existing users or surprise everyone with the change at once.
The smart move: learn to use open-source tools with your own API keys now. The subscription that seems cheap today will look expensive in a year. And switching costs only go up with every month of lock-in.
Related: Best Coding Agents 2026 Decision Guide · What Developers Actually Say About Claude Code
Related articles
- Claude Code Alternatives in 2026: 12 Options Compared
- Coding Agents in 2026: Three Hard Lessons HN Developers Learned the Expensive Way
- What Developers Think About Coding Agents — The Skill Atrophy Crisis, Interface Wars, and Trust Divide From 500 HN Comments
Sometimes Claude writes better code. Sometimes GPT does. aiFiesta lets you compare both instantly in one chat for $12/mo — plus 7 more premium models.