GitHub Copilot is moving to usage-based billing

Original Article ↗ Hacker News Discussion ↗

GitHub’s move to usage‑based billing for Copilot, including steep new “model multipliers” (e.g. 6–9x for mid‑tier models and 27x for Claude Opus on some annual plans), is widely seen as the end of heavily subsidized AI coding assistance. Commenters argue that per‑request pricing was never sustainable once long‑running agentic sessions became common, but many feel blindsided by how sharply effective costs are rising, especially for users locked into annual plans. The change is pushing developers toward direct API access, OpenRouter and other harnesses, or local/open‑weight models, and raising broader concerns about vendor lock‑in, “enshittification,” and whether AI coding tools will remain affordable for individuals and small teams.

Pricing changes and mechanics

  • Copilot is moving from per-request subscription to usage-based, token-priced “AI Credits” that roughly mirror underlying API costs.
  • Plan fees (e.g., $10 Pro, $39 Pro+) stay nominally the same, but now just prepay that dollar amount in credits each month.
  • Annual Pro/Pro+ subscribers keep the old “premium request” model until renewal, but model multipliers jump sharply (e.g., Sonnet 4.6 from 1×→9×, GPT‑5.4 from 1×→6×, Opus from 3×→27×), effectively slashing allowed high-end usage.
  • Autocomplete and basic “next edit suggestions” remain unlimited within plans; agentic chat, code review, and containers consume credits (and GitHub Actions minutes for reviews).
  • New models and features will not be added to legacy annual plans; users can get prorated refunds or convert to monthly.

User sentiment and behavior

  • Many individual users say the deal went from “incredible value” to “not worth it overnight” and plan to cancel or switch.
  • Heavy agentic users report they were effectively getting hundreds of dollars’ worth of Opus/Sonnet tokens monthly for $10–$40; they see this as a massive (often 10–100×) effective price hike.
  • Some will keep Copilot solely for autocomplete and VS Code integration, wishing for a cheaper autocomplete-only tier.

Economics of inference and “enshittification”

  • Broad agreement that flat per-request pricing was unsustainable once “one request” could trigger hours-long, multi-agent runs.
  • One camp frames this as classic “enshittification”/bait‑and‑switch after a subsidized land‑grab; another says it’s just the end of loss-leading and alignment with real compute costs.
  • Concern that token anxiety will make users self-censor use of stronger models, degrading experience and adoption.

Alternatives and competition

  • Many mention moving to:
    • Direct APIs (Anthropic, OpenAI, DeepSeek, Kimi, etc.).
    • Routers like OpenRouter (with ~5–5.5% markup but unified API, failover, easier billing).
    • Other IDE/agent tools (Claude Code, Codex CLI, Cursor, Windsurf, Cline, OpenCode).
  • Enterprises may stick with Copilot due to Microsoft ecosystem lock‑in and data-governance approvals, even at higher effective cost.

Legal and contract concerns

  • Some argue mid-term changes for annual plans (especially 6–9× multipliers) may violate consumer-protection laws in regions like the EU/Australia; others note that prorated refunds are offered and ToS often allow such changes.

Local models and longer‑term outlook

  • Growing interest in local/open-weight models (Qwen, Gemma, DeepSeek, etc.) via tools like Ollama, llama.cpp, Unsloth, especially to avoid variable bills and privacy risks.
  • Debate over whether cloud inference costs will keep rising or be undercut by increasingly capable, cheap-to-run open models on consumer hardware.