GitHub Copilot's new usage-based pricing model went live on June 1, 2026, and subscribers are watching their monthly AI allotments evaporate in hours. The system, announced in April, replaces the old request-and-premium-request tiers with credits priced at $0.01 each. The $10/month Pro plan now ships with 1,500 credits, Pro+ at $39/month includes 7,000 credits, and the $100/month Copilot Max plan grants 20,000 credits. Users posting to Reddit, Bluesky, and X are reporting that previously routine workflows now consume a fifth to all of a month's quota in a single session.
GitHub's stated rationale is that the old flat-request model treated a quick chat question and a multi-hour autonomous coding session as equivalent, forcing the company to absorb much of the escalating inference cost behind heavy usage. The new model passes that cost through directly, with each prompt's credit consumption set by input and output token counts at the underlying model's rates. That makes pricing extremely sensitive to model selection.
The spread between models is stark. One million output tokens from OpenAI's GPT-5.4 nano runs $1.25 on Copilot, while the same volume on the frontier GPT-5.5 model costs $30 — a 24x difference. Users relying on Copilot's Auto mode to pick the best available model are reporting that it sometimes escalates to expensive models for trivial queries, blowing through credits with no warning.
Key facts
- 01GitHub Copilot switched from request-based to usage-based pricing on June 1, 2026, with each credit priced at $0.01.
- 02The $10/month Pro plan includes 1,500 credits, Pro+ at $39 includes 7,000, and the $100 Copilot Max plan includes 20,000.
- 03One user reported burning 840 credits in a single cautious day testing Claude Sonnet 4.6 through Copilot.
- 04Model choice swings cost wildly: 1M output tokens cost $1.25 on GPT-5.4 nano versus $30 on GPT-5.5.
- 05One Reddit user reported running Deepseek inside GitHub's VSCode environment for roughly 7 cents per 15 million tokens.
Real-world consumption figures from users tell the story. A simple build-a-Minesweeper test on Claude Haiku 4.5 burned about 94 credits. One complex prompt ran 171 credits. A short series of prompts hit 700 credits. A couple of Copilot-led commits on an existing codebase consumed 5,000 credits — a third of the entire Pro+ monthly allotment. Even smaller actions surprised users: a run-of-the-mill query reportedly cost 15 credits, and generating a small plan ran 100.
The reaction from power users has been blunt. One subscriber testing Claude Sonnet 4.6 watched 840 credits drain on the first day despite limiting their usage, with no complex work yet attempted. Another reported burning through 21% of their Pro plan's monthly credits in a single day and said they were considering switching tools. One organization reported using all 8,000 of its monthly AI credits in a day — a pace that does not survive contact with a 30-day billing cycle.
“Even though I was super cautious on the first day, trying it out with a limited number of uses, it still consumed 840 credits”— Anonymous Copilot user, GitHub Copilot subscriber
Not everyone is bouncing. Coder Henri Kinnunen reported spending only 161 credits in a productive day with GPT 5.3-Codex by sticking to focused, deliberate edits. On Bluesky, coder Neil Hewitt pointed out that continuing a three-day-old chat session resends the entire history as input tokens on every turn, noting that input tokens use credits and that this is not rocket science. The implication is that disciplined prompt hygiene — short sessions, cheaper models, narrow context windows — can keep usage in line with the new caps.
The cost-estimator tool GitHub built to show users what their old usage would have looked like under the new pricing has produced bills in the thousands of dollars for some power users, illustrating how heavily Copilot was subsidizing inference under the previous model. That subsidy is now ending, and the bill is being itemized for the first time. Users who had treated Copilot as effectively unmetered are discovering they were running enterprise-scale inference on a $10 plan.
Some Copilot subscribers are migrating to services with looser limits or to self-hosted setups. One Reddit user described wiring Deepseek into a GitHub VSCode environment for roughly 7 cents per 15 million tokens — a rate that makes the credit math on premium frontier models look indefensible for routine coding. The broader competitive pressure here is that token efficiency, not just raw capability, becomes a procurement criterion when customers see the meter.
The skeptic's read on this rollout is that GitHub has botched the messaging. Users are not, on the whole, objecting to paying for what they use — they are objecting to the lack of intuition about what each action will cost, especially in Auto mode. A credit system pegged to underlying token rates is functionally a variable price list the user cannot see in advance, and one bad model selection can cost a fifth of a monthly plan. Better real-time cost visibility, hard caps, and model-budget controls would defuse most of the complaints without changing the underlying economics.
The wider implication is that the era of flat-fee, all-you-can-eat AI coding assistants is ending. Inference costs at the frontier are real, and someone has to carry them. GitHub is the first major coding-assistant vendor to push the full cost onto the user, but Cursor, Replit, and others running on the same upstream OpenAI and Anthropic rates face the same arithmetic. Expect usage-based pricing to spread across the category over the next two quarters — and expect cheaper, more token-efficient models like Deepseek and Haiku tiers to gain share at the expense of frontier defaults, regardless of who has the better benchmark scores.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
