The three biggest AI coding tools are all switching to metered billing within three weeks of each other. Here is what the numbers actually look like.
May 2026 is the month the deal died.
For two years, AI coding tools were ridiculously cheap. Twenty bucks a month for a tool that could write, debug, and refactor code across your entire codebase. It was obvious this could not last. The compute costs alone were staggering. But the party went on long enough that engineering teams started building real workflows around these prices.
Now the bill is coming due, all at once, from every direction.
The Copilot pricing shock
On June 1, GitHub Copilot replaces its flat-rate Premium Request Units with token-based AI Credits. The headline prices look reasonable: Pro gets $10 in monthly credits, Pro+ gets $39, Business gets $19, Enterprise gets $39. Code completions stay free.
The fine print is brutal.
Users on Reddit started posting screenshots of their projected bills under the new model. One user went from $39/month to a projected $5,851.77/month. Another went from $39 to $942. The thread titles say it all: "100% sure I am out, GitHub just turned my $39/month Copilot into $942/month overnight."
The math is simple once you see it. GPT-5.5 has a 57x multiplier under the new system. A Pro+ subscriber at $39/month in credits gets roughly 144 Opus-level requests per month, or about 4.8 per day. A single agentic coding session of two hours can burn 50 to 200 inferences. One session can eat a week of quota.
GitHub removed the safety net too. Previously, when you ran out of Premium Request Units, you fell back to a cheaper model. Now, when your credits are gone, you pay out of pocket. There is no fallback.
Multiple HackerNews users reported projected costs of $3,000 to $5,000/month under the new billing model, up from roughly $40/month. One company dropped Copilot Business for OpenAI Business with Codex before the change takes effect.
Anthropic's four-hour subscription
Anthropic is making the same pivot, just with better marketing. On June 15, claude -p and Agent SDK usage get removed from subscription limits and billed against a monthly credit at API rates: $20 for Pro, $100 for Max 5x, $200 for Max 20x.
Anthropic calls this a "new SDK credit." The community calls it a "Trojan Horse" and a "massive nerf disguised as a feature."
The community did the math. A $200/month Max 20x credit gets exhausted in roughly four hours of normal Opus usage on a single session. One user burned through 45% of their five-hour rate limit on just three questions using Opus 4.7/4.8.
This came on top of another incident. Anthropic tested removing Claude Code from the $20/month Pro plan entirely. When screenshots leaked, they called it a "test" and walked it back. The HackerNews thread hit 683 points. Simon Willison commented: "If you don't want things like this spreading through screenshots of X and Reddit, don't run 'tests' like this in the first place."
The Opus 4.8 launch did not help. The model was built on Opus 4.7, which the community widely rejected as a downgrade from 4.6. The Reddit launch thread had 2,600 upvotes and 786 comments, mostly skeptical. The top comment with 686 upvotes: "Just hoping 4.8 behaves more like 4.6."
The $200/month comparison that should embarrass Cursor
At the same $200/month price point, the difference between AI coding tools is staggering.
A side-by-side comparison by developer Andrew Shu found that Claude Code Max 20x gives you roughly 678 agent-hours of frontier model usage per month. Cursor Ultra gives you approximately 18. That is a 38x gap.
Cursor also burns 7 to 9x more "usage" per prompt than VS Code with the same model. After three chat prompts in Cursor, usage is already at 20%. In VS Code with the same Sonnet 4.6 model, the meter barely moves. Nobody can explain why.
Cursor did launch Composer 2.5, a proprietary model built on Moonshot's Kimi K2.5 checkpoint, at $0.50/$2.50 per million tokens. It now handles 87% of agent-hours on the Ultra plan. But you are paying $200/month for a tool that mostly uses a budget model and gives you less than a day of frontier access.
SpaceX announced an option to acquire Cursor for $60 billion. The company is at $300 million in annual revenue and raising $2 billion at $50 billion valuation. A US Congressional probe into Cursor's use of Chinese AI models adds regulatory uncertainty.
The Chinese models that match GPT-5 at 1/10th the price
Here is where it gets interesting for anyone willing to look past the geopolitical concerns.
DeepSeek made its 75% V4-Pro discount permanent on May 22. The price is now $0.435 per million input tokens and $0.87 per million output tokens. For comparison, Claude Opus 4.8 is $5/$25. DeepSeek is 11.5x cheaper on input and 28.7x cheaper on output.
A NIST CAISI evaluation rated DeepSeek V4-Pro "on par with GPT-5." A United States government agency certified parity between a Chinese open-weight model and OpenAI's flagship.
The HackerNews community built DeepClaude, a project that routes Claude Code through DeepSeek's API. Users report quality "close to Opus 4.5" at less than half of Anthropic's Haiku pricing. The project received 678 upvotes on HackerNews. One user who ran it for a week on non-confidential projects said: "I honestly can't tell the difference."
Moonshot's K2.6 model won the AI Coding Contest outright with 22 match points, beating GPT-5.5 (16 points), Claude Opus 4.7 (12 points), and Gemini Pro 3.1 (9 points). It costs $0.95/$4.00 per million tokens, roughly 5x cheaper than Opus 4.8. Its SWE-Bench Pro score of 58.6% beats GPT-5.4 and crushes Claude Opus 4.6.
The catch is always the same. No enterprise features. No SSO, no audit logs, no IP indemnity. Data is processed in China under the National Intelligence Law. DeepSeek has no opt-out for training on your API data. For solo developers and small teams working on non-sensitive code, these models are a no-brainer. For anyone with compliance requirements, they are not an option.
What to actually do about this
The era of cheap, flat-rate AI coding is over. Here is what that means in practice.
Budget for real costs. A team of 10 developers using Copilot Business at $19/user/month is paying $190/month today. Under the new AI Credits model with typical agentic usage, that same team could easily hit $3,000 to $5,000/month. Run the numbers before June 1.
Stop using frontier models for everything. Cursor's own data shows that only 13% of agent-hours actually need a frontier model. The other 87% run fine on cheaper alternatives. Augment's Prism system routes each turn to the cheapest model that can handle it and claims 33% lower cost than Claude Code at the same quality. Model routing is the new hotness, and it works.
Lock in promotional rates. Copilot Business has a promotional $30/month credit rate through August. Enterprise has $70/month through August. OpenAI Pro promotional multipliers expire May 31, cutting capacity by 50%. Whatever deal you are on, check the expiration date.
Evaluate the cheap models. If you do not have compliance constraints, try DeepSeek V4-Pro or Moonshot K2.6 for non-sensitive work. The quality is genuinely close to frontier, and the price difference is not incremental. It is an order of magnitude.
Consider self-hosting. Mistral Medium 3.5 at $1.50/$7.50 per million tokens is open weights, 128B dense, scores 77.6% on SWE-Bench Verified, and can run on your own hardware. The Mistral Vibe Pro plan at $14.99/month is the cheapest paid subscription that includes a full coding agent.
The math has changed. The tools have not caught up to the prices. Plan accordingly.
Data sourced from the AI Coding Agents Monthly Report for May 2026, covering 17 suppliers with verified pricing, benchmarks, and community signals.