GitHub Copilot has confirmed that all plans will move to usage-based billing on June 1, 2026. Under the change, Microsoft will no longer rely on a fixed pool of request counts across models. Instead, pricing will be tied more directly to token usage, with the source article noting that a $19 monthly plan would buy $19 worth of tokens.
The piece, written by Ed Zitron and repackaged by BlockTempo, treats the pricing update as a sign of strain across generative AI. Its argument is blunt: subscription pricing has hidden the real cost of inference for years, and Copilot’s change makes that mismatch harder to ignore. It cites an Wall Street Journal report from October 2023 saying Copilot charged individual users $10 per month, while GitHub was losing more than $20 per user each month on average during the first months of that year, with some users costing the company $80 a month.
Why the old request model is being replaced
GitHub’s explanation, as quoted in the article, is that Copilot has changed from an in-editor assistant into an agent-style coding platform capable of long, multi-step sessions across a codebase. That shift raises compute and reasoning demand. The article pushes back on the framing and says the core issue is not product evolution but economics: users have been allowed to consume more compute than their subscription price could reasonably support.
According to the source, Copilot’s previous structure included 300 premium requests per month alongside “unlimited chat requests” on cheaper models such as GPT-5 mini. It also says Microsoft had allowed broader model access until May 2025. In the author’s view, systems built around requests and soft limits make it difficult for users to see how many tokens they burn and what a given task actually costs.
The criticism extends beyond Copilot
The article broadens the claim to the wider AI market, naming OpenAI, Anthropic and Perplexity as examples of services built on flat monthly subscriptions while underlying inference costs remain variable. It says Anthropic had at one stage allowed users to burn as much as $8 in compute for every $1 of subscription revenue, and argues OpenAI has operated under a similar imbalance, though with less visibility into the exact numbers.
The central thesis is simple: for services built on large language models, charging without tracking actual token consumption means subsidizing heavy usage. Once pricing starts to reflect real costs, customer expectations change fast. That is the backdrop the author uses to explain the backlash from Copilot users who now see the service as materially different from what they originally signed up for.
Enterprise spending is also under scrutiny
The piece also points to the enterprise side. It references updated Claude Code documentation from Anthropic stating that in enterprise deployments, average cost is about $13 per active developer day, or roughly $150 to $250 per month, while 90% of users stay below $30 per active day. Using a 21-workday month, the article calculates an average monthly cost of about $273 per developer, or $3,276 per year. At $30 per active day, that rises to about $630 per month and $7,560 per year.
From there, the argument turns to return on investment. The article questions whether companies have clearly measured what they gain from these tools once token usage stops being buried inside a broad software budget and starts showing up as a visible operating expense. Copilot’s pricing shift is presented as one example of that wider reckoning: AI usage may still be attractive, but the bill is becoming harder to disguise.