GitHub just announced that Copilot is moving to usage-based billing on June 1. The Claude Opus 4.7 multiplier jumps from 7.5× to 27×. GPT-5.4 goes from 1× to 6×. GitHub's own admission: "The subscription model is no longer sustainable." A single agentic coding session burns 500K+ tokens — $5-15 worth of compute on a $10/month Pro plan. The unlimited AI coding era is definitively, permanently over.
The New Pricing Reality
Here's what changes on June 1:
| Model | Old Multiplier | New Multiplier | Change | |-------|---------------|---------------|--------| | GPT-5.4 | 1× | 6× | 6× more expensive | | Claude Opus 4.7 | 7.5× | 27× | 3.6× more expensive | | GPT-5.5 | New | 12× | New baseline | | Claude Sonnet 4 | New | 3× | Budget option |
What does this mean in practice? Under the old model, a Copilot Pro user paying $10/month could use Claude Opus 4.7 for coding tasks. Each request consumed tokens at 7.5× the base rate, but the flat subscription absorbed the cost.
Under the new model, that same user gets a monthly token allowance. Once they exceed it, they pay per token. And with Opus 4.7 at 27× the base rate, power users will burn through their allowance in days.
The Agentic Session Problem
Here's why the economics broke:
A traditional autocomplete suggestion uses ~500 tokens. At $10/month, a developer making 1,000 suggestions/day costs GitHub roughly $0.50/day. Sustainable.
An agentic coding session — where the agent reads files, plans changes, writes code, runs tests, reviews errors, and iterates — uses 500,000+ tokens per session. At frontier model pricing, that's $5-15 per session. A developer running 3-4 agentic sessions per day costs GitHub $15-60/day. On a $10/month subscription.
The math doesn't work. It never did. GitHub was losing money on every power user who discovered agentic workflows.
Why Every Vendor Is Making This Move
GitHub isn't the first. They're not even the most dramatic:
Anthropic
- Pulled Claude Code from Pro tier
- Reduced effective compute by ~40% for heavy users
- Max tier subscribers now cost more than they pay
OpenAI
- Launched GPT-5.5 at 2× the price of GPT-5.4
- No longer offering "unlimited" anything at flat rates
The Pattern
Every vendor followed the same playbook:
- Launch generous flat pricing to capture market share
- Add agentic features that 10-100× per-user costs
- Subsidize power users with casual user revenue
- Hit the compute-cost wall when agentic usage scales
- Switch to usage-based pricing (or quietly reduce quality)
GitHub is just the latest to reach step 5. Every other vendor offering flat-rate AI subscriptions will follow.
What This Means for Development Teams
1. Budget for Variable AI Costs
AI coding tools are no longer a fixed line item. They're a variable cost that scales with usage. Budget accordingly — and set team-wide limits to prevent surprises.
2. Choose Models Strategically
Not every coding task needs Opus 4.7 at 27× the cost. Use tiered model selection:
- Autocomplete: Base model (1×) — fast, cheap, good enough
- Code review: Claude Sonnet 4 (3×) — solid analysis at reasonable cost
- Complex refactoring: GPT-5.4 (6×) — capable without premium pricing
- Critical architecture decisions: Opus 4.7 (27×) — save for when it matters
3. Monitor Usage Per Developer
Agentic workflows are addictive. A developer who discovers they can describe a feature and have an agent build it will use 10-50× more tokens than a developer using autocomplete. Track per-developer usage and set alerts.
4. Evaluate Self-Hosted Alternatives
With usage-based pricing, self-hosted open-weight models become more attractive:
- DeepSeek V4-Flash at $0.14/M tokens — 27× cheaper than Opus 4.7
- Xiaomi MiMo V2.5 — MIT-licensed, 1T params, agent-optimized
- mini-swe-agent — 100 lines of Python, >74% SWE-bench
5. Build Abstraction Layers
Don't lock your development workflow into GitHub's billing model. Build tool-agnostic coding infrastructure that can switch between providers based on cost and capability.
Honest caveat: Usage-based pricing is actually more fair than flat-rate pricing. It aligns costs with value received. Teams that use AI heavily pay more — but they're also getting more value. The problem isn't the pricing model; it's the sticker shock of transitioning from artificially cheap flat rates to actual cost. And for teams that were barely using agentic features, the new model might actually be cheaper.
The Financial Impact
Before: Flat-Rate Copilot Pro
| Item | Cost | |------|------| | 10 developers × $10/month | $100/month | | Unlimited agentic sessions | "Free" (subsidized) | | Total | $100/month ($1,200/year) |
After: Usage-Based Copilot
| Item | Cost | |------|------| | 10 developers × $10/month base | $100/month | | ~50 agentic sessions/day × $10 avg | ~$1,500/month | | ~500 autocomplete/day × $0.001 | ~$15/month | | Total | ~$1,615/month ($19,380/year) |
Cost increase: 16× for a team using agentic workflows.
Optimized: Tiered Model Strategy
| Item | Cost | |------|------| | 10 developers × $10/month base | $100/month | | 5 Opus 4.7 sessions/day × $15 | ~$225/month | | 20 Sonnet 4 sessions/day × $3 | ~$60/month | | 25 base model sessions/day × $0.50 | ~$12.50/month | | Total | ~$397.50/month ($4,770/year) |
Optimized cost: 75% less than uncontrolled usage, 4× more than old flat rate.
The lesson: you can't avoid paying more for agentic AI. But you can control how much more by being strategic about model selection.
Closing Thoughts
The era of "$10/month for unlimited AI coding" is over. It was never sustainable — it was a customer acquisition strategy that vendors could afford while AI coding was niche. Now that every developer wants agentic workflows, the math doesn't work.
This isn't bad news. Flat-rate pricing created perverse incentives: developers overused expensive models for simple tasks, and vendors silently degraded quality to manage costs. Usage-based pricing is honest — you pay for what you use.
The teams that adapt fastest will be the ones that build model selection discipline: using the cheapest model that gets the job done, reserving frontier models for tasks that genuinely need them. The teams that don't adapt will get a very large bill in July.
Welcome to pay-per-use AI. It's more honest. It's also more expensive. Plan accordingly.
Need help optimizing AI coding costs? Book a Developer Tool Cost Audit — we'll analyze your team's AI usage patterns and build a tiered model strategy that balances cost and capability.