The $20 Era of AI Coding Is Ending

This week Anthropic pulled Claude Code from the $20 Pro plan for new signups, and GitHub paused new Copilot signups (even the $10 tier), both within 48 hours and both for the same reason.

Anthropic moved Claude Code to their Max tier at $100/month. The developer community pushed back hard, thousands of developers within hours, and Anthropic walked back the public pricing page, calling it "a 2% test on new signups." Existing Pro subscribers still have access.

GitHub went further and paused all individual paid plans. Pro at $10, Pro+ at $39, both say "temporarily unavailable" right now. GitHub VP Joe Binder: "It's now common for a handful of requests to incur costs that exceed the plan price."

I've been using Claude Code daily for months, so when both announcements landed in the same week I looked at the actual numbers.

Various measurements show that 60 to 87% of tokens in coding agent sessions go to searching and reading code, not writing it. Only 5-15% of tokens go to actual code generation. A one-line typo fix can use up to 21,000 tokens. Ever watched your agent work? It spends most of its time grepping files and re-sending context, and the actual coding is just a small fraction.

These companies knew agent compute was expensive, but the growing wave of vibe coding and open-source coding agents brought a massive number of new users who run agents all day. That volume is what broke the economics.

Claude Pro at $20 was originally priced for chat, and Copilot at $10 was priced for autocomplete. Agent features got added to these plans without adjusting the price. The problem isn't just the cost, it's the variance: a light user and a heavy vibe coder on the same $20 plan can consume 100x different amounts of compute, and flat-rate pricing can't survive that spread.

Here's what each option actually costs right now:

Claude Code on Pro: $20/month, but Anthropic already tested pulling it for new users Claude Code on Max: $100/month, fixed ceiling, the "safe" option Claude Code on API: no ceiling, and from personal experience it goes up very fast GitHub Copilot Pro/Pro+: $10-39/month but you can't sign up right now OpenAI Codex: included in ChatGPT free and Plus ($20) tiers with usage limits, still accepting signups Cursor: $20/month + usage-based credits for agent mode Windsurf: $20/month Pro with daily quotas, heavy users report $25-55/month effective cost Gemini CLI: free with Flash models (1,000 req/day), Pro models require a paid plan Google Antigravity: currently free in public preview with rate limits, paid tiers expected Local models: no monthly cost, but you need a serious GPU and you won't get Opus-level quality

If you're choosing a coding agent today:

Light use (a few sessions per day): Codex free, Gemini CLI free, or Copilot free still work Daily use (10+ sessions): Claude Pro $20, Cursor $20, or Windsurf $20, but track your usage closely Heavy use (agents running all day): Claude Max at $100 gives you the highest predictable ceiling. API sounds flexible but the costs add up faster than you expect Maximum control: local models give you no surprise bills, but the quality gap with Opus or Sonnet is real

The $20/month era for AI coding agents is ending. Worth checking where you stand before your next billing cycle.

What's your actual monthly spend on AI coding tools right now?