GitHub Copilot Usage-Based Billing Goes Live in 2026: Token Pricing Explained
TL;DR
- GitHub retired Copilot’s premium request model and launched token-based billing that charges users based on actual AI model consumption
- Plan prices unchanged ($10-$100/month for individuals, $19-$39 for business), but credits now deplete based on token usage across input, output, and cached data
- Model choice matters financially — GPT-5.5 costs 20x more per token than GPT-5 mini, turning model selection into a cost management decision
- Some power users report projected bills jumping from $39 to $847 for identical usage patterns under the new system
What Happened
GitHub officially switched all Copilot subscribers to usage-based billing as of this week, replacing the premium request unit system with GitHub AI Credits tied directly to token consumption. The change affects all plan tiers, though annual subscribers remain on the old model until their current term expires.
The core shift: instead of a monthly allowance of premium requests, users now receive a dollar-denominated credit pool that depletes based on which AI models they use and how much they use them. A Pro user paying $10/month now gets $15 in total credits ($10 base + $5 “flex allotment”), while the new Copilot Max tier at $100/month includes $200 in credits.
Each interaction with Copilot now has a visible cost measured in tokens — roughly 3-4 characters of text. Input tokens (your prompts), output tokens (Copilot’s responses), and cached tokens (stored context) all consume credits at published per-model rates. GitHub published a full pricing breakdown showing massive variance: GPT-5 mini costs $0.25 per million input tokens while GPT-5.5 costs $5.00 — a 20x difference for choosing a more capable model.
Why It Matters
This fundamentally changes how development teams budget for AI assistance. What was previously a predictable monthly expense is now a variable cost that scales with usage patterns and model selection.
The economic reality behind the change is straightforward: GitHub was subsidizing power users who ran long agentic coding sessions across entire repositories. As Joe Binder, VP of Product at Microsoft, acknowledged, the escalating inference costs made the flat-rate model “unsustainable” as Copilot evolved from simple autocomplete to autonomous multi-file operations.
For enterprise teams, this introduces a FinOps discipline that didn’t exist before. Admins must now monitor which developers consume the most credits, which models they’re selecting, and whether those choices align with the actual complexity of their tasks. Reaching for Claude Opus 4.8 for every query is no longer cost-neutral — it’s 5x more expensive than Claude Haiku 4.5 per million tokens.
Key Details
Individual Plan Pricing (Monthly)
| Plan | Price | Total Credits | Base Credits | Flex Allotment |
|---|---|---|---|---|
| Pro | $10 | $15 | $10 | $5 |
| Pro+ | $39 | $70 | $39 | $31 |
| Max | $100 | $200 | $100 | $100 |
Business/Enterprise Pricing
- Business: $19/user/month with $19 in credits per seat (no flex allotment)
- Enterprise: $39/user/month with $39 in credits per seat (no flex allotment)
- Promotional boost through August: Business gets $30/user, Enterprise gets $70/user
Model Pricing Comparison (Per Million Input Tokens)
- GPT-5 mini: $0.25
- Claude Haiku 4.5: $1.00
- GPT-5.5: $5.00
- Claude Opus 4.8: $5.00
What Doesn’t Consume Credits
- Code completions
- Next edit suggestions
- Code review (now uses GitHub Actions minutes instead)
Budget Controls for Enterprise
GitHub introduced four layered controls:
- Universal user-level budget (always enforced, hard stop)
- Individual user-level budget overrides
- Cost center budgets (only active after shared pool exhausted)
- Enterprise-wide budget (caps metered overage, not total spend)
Critical detail: The enterprise budget doesn’t cap your total bill — it only limits metered charges after your included credits run out. An organization with 400 Business seats ($7,600 in license fees) and a $5,000 enterprise budget faces a maximum bill of $12,600, not $5,000.
Implications
The backlash on Reddit and developer forums reveals a fundamental tension: users signed up for what they understood as unlimited usage within their tier, and now face consumption-based limits that can be exceeded in hours of heavy work.
One Pro+ subscriber reported their projected bill jumping from $39 to $847 for identical usage — work that previously consumed a fraction of their 1,500 monthly premium requests now burned through 200 credits in a single complex prompt. Another user noted they were exploring direct OpenAI and Anthropic subscriptions as alternatives, potentially cutting GitHub out of the value chain entirely.
For GitHub’s business model, this is a calculated risk. The company is betting that transparency around model costs — plus tools like auto-routing to cheaper models for simpler tasks — will keep most users within their included credits while heavy users either upgrade to Max or accept metered charges. The “flex allotment” mechanism provides GitHub a valve to adjust included credits as AI infrastructure costs evolve, without changing sticker prices.
The broader industry implication: usage-based AI billing is becoming the standard. OpenAI, Anthropic, and now GitHub have all moved away from flat-rate unlimited models. The era of subsidized AI experimentation is ending, replaced by cost structures that reflect actual computational expense.
Our Take
GitHub made the only decision that made economic sense, but the execution raises questions about whether they understood their own power users.
The math is brutal: a developer running sustained agentic sessions with GPT-5.5 or Claude Opus 4.8 can easily burn through $70 in credits in days, not weeks. The Max tier at $200/month helps, but it’s still a 10x price increase over Pro — a sticker shock that will drive some teams to evaluate whether they’re getting 10x the value.
What to watch: Whether GitHub adjusts the flex allotment upward in response to usage data, and how quickly competitors like Cursor and Codeium capitalize on pricing arbitrage opportunities. If a power user can get equivalent functionality for less by stitching together direct API access to Claude or GPT models, GitHub’s platform value proposition weakens.
The enterprise budget system is unnecessarily complex — the “lowest remaining headroom wins” rule will absolutely catch admins off guard when individual user budgets don’t align with organization-level caps. This feels like a system designed by engineers who understand tokens intimately but haven’t spent time with finance teams managing SaaS sprawl.
Bottom line: If you’re a light Copilot user doing basic completions, nothing changed. If you’re running multi-file refactors or autonomous debugging sessions, you now have a FinOps problem. Start tracking which models you’re using and whether you actually need the expensive ones for every task.