In four days from April 20 to 23, three of the largest AI companies publicly reset how they charge. GitHub paused new signups for Copilot Pro, Pro+, and Student plans and tightened usage limits while preparing a June move to token-based billing. Anthropic briefly removed Claude Code from its $20 Pro plan on April 21 before walking it back as a 2% test. Leaked Microsoft documents published April 22 show Copilot Business moving to $19/month with $30 in pooled AI credits and Enterprise to $39/month with $70, starting June.
The number that matters: weekly cost of running GitHub Copilot has nearly doubled since January. Agentic workflows consume 10× or more tokens than per-request subscription economics anticipated.
Our read: this is not a dev-tools story. It’s the clearest signal yet that every flat-rate AI subscription in your B2B stack is on borrowed time. HubSpot Breeze shifted to outcome-based pricing. ChatGPT Ads moved CPM to CPC. Adobe launched CX Enterprise on credits. The pattern is confirmed across the four biggest AI vendors selling to B2B, and procurement doesn’t have a playbook. Brinker’s 2026 supergraphic adds a market-wide cross-section to the same pattern: 1,367 products exited the martech landscape this year while only 1,488 entered, a 9.7% gross churn rate that makes the procurement playbook gap structural rather than vendor-specific.
Key Takeaways
- Between April 20-23, GitHub paused Copilot individual signups, Anthropic tested removing Claude Code from the $20 Pro plan, and leaked Microsoft docs detailed a June move to token-based Copilot billing.
- Agentic workflows consume 10× or more tokens than per-request plans assumed. GitHub’s weekly Copilot run cost nearly doubled since January.
- Proposed Copilot Business: $19/month with $30 pooled credits (promo), settling to $19 for $19 in tokens. Enterprise: $39/month with $70, then $39 for $39.
- The pattern matches ChatGPT Ads, HubSpot Breeze, and Adobe CX Enterprise. Every major AI vendor selling to B2B is repricing in Q2 2026.
- Three procurement actions: audit token consumption before renewal, pull forward renewal conversations one quarter, add model-deprecation clauses to RFPs.
What Actually Happened in Four Days
April 20: GitHub paused new signups for Copilot Pro ($10/month), Pro+ ($39/month), and Student plans. Existing customers stayed, but weekly token caps tightened and Opus 4.5/4.6 came out of the Pro tier. GitHub cited agentic workflows fundamentally changing Copilot’s compute demands.
April 21: Anthropic’s pricing page briefly showed an X next to Claude Code under the $20 Pro plan, with checkmarks only on Max 5× ($100) and Max 20× ($200). Head of Growth Amol Avasare confirmed on X it was a 2% test, though the pricing page change had rolled out globally. A day later, Ed Zitron published leaked Microsoft documents detailing Copilot’s full transition: token-based billing launches in June. Business at $19/month with $30 pooled credits during promo, Enterprise at $39/month with $70. After promo, $19 for $19 in tokens and $39 for $39 in tokens.
Why the Shift Is Happening
A $10 Copilot Pro subscription used to cover 300 API requests per month against smaller models. That same user running Claude Opus 4.7 in Copilot CLI on an agentic task can generate 100,000+ output tokens in a single session. At $25 per million output tokens, infrastructure cost per user now matches or exceeds the subscription price.
Anthropic’s position, paraphrased from its head of growth: Max was built a year ago for heavy chat usage. Claude Code, Cowork, and long-running async agents didn’t exist then. Engagement per subscriber is way up, and current plans weren’t built for it.
The response is convergent. Token-based billing, outcome-based (HubSpot Breeze in April), CPC (ChatGPT Ads), and credit-based (Adobe CX Enterprise) are the same bet: align vendor revenue with buyer compute cost, stop subsidizing heavy users, let light users pay less. The AgentCore Payments rail AWS launched with Coinbase + Stripe on May 7 takes the same logic to the per-call micropayment scale: agents pay for resources at fractions of a cent, which means the next pricing-model conversation B2B procurement has to absorb is one where the spender is the agent itself, not the buyer.
What This Means for B2B Procurement
B2B SaaS procurement spent a decade optimizing seat-based pricing: volume discounts, multi-year commits, seat-count flexibility. Usage-based AI pricing breaks most of those levers.
Budget forecasting gets harder. A RevOps team with 50 flat-rate seats has a predictable line item. The same team on tokens can see consumption swing 3-5× between a quiet month and a quarter-end push. Vendor evaluation needs consumption trajectory, not just features. Contract language needs rebuilding — most AI contracts from 2024-2025 don’t contemplate usage-cap enforcement, model deprecation, or pricing-floor renegotiation.
The real risk isn’t that AI pricing is going up. It’s that a 12-month contract signed on flat-rate terms in Q1 can be subject to mid-term repricing, model removal, or rate-limit tightening — which is what GitHub just did to existing Pro subscribers, with 30 days’ refund notice.
What B2B Buyers Should Do in the Next 60 Days
Three concrete moves:
Audit actual token consumption before renewal. Most AI vendors now provide usage dashboards. Pull 30-day consumption per user and compare to the new per-token math. If you’re flat-rate and your team runs agents heavily, usage-based could increase cost 2-3×. If your team is light, it could save money. You can’t negotiate without the number.
Pull forward renewal conversations by one quarter. If your HubSpot, Salesforce, Microsoft 365, GitHub Copilot, or Anthropic Enterprise contract renews in Q3-Q4, start now. Mid-cycle repricing is the scenario this week made live. Teams that renegotiated Salesforce Headless 360 and HubSpot AEO terms in April already have stronger positions than teams waiting until September. Mailchimp customers facing the same Q3 renewal pressure on the SaaS side have a tighter window: Intuit’s May 20 confirmation that Mailchimp is now run for cash flow rather than growth turns the renewal conversation into a stay-evaluate-migrate decision with a 90-day clock attached.
Add three questions to every AI vendor RFP. What’s your roadmap for pricing model changes? What’s the notice period and refund policy if billing changes mid-contract? What usage caps or model-deprecation clauses are in the standard agreement? Answers tell you which vendors are preparing to move existing customers.
B2B AI procurement is where cloud procurement was in 2012 — pricing unstable, contracts vendor-favorable, buyers haven’t organized counter-leverage. Teams treating Q2 2026 as a procurement-discipline moment will spend the rest of the year negotiating from a stronger position. DALL-E 3’s May 12 retirement is part of the same forced-rebuild pattern hitting B2B image-gen workflows. The Anthropic-Google $40B deal that closed days later reshaped the supply-side picture this piece described from the demand side, making the Q2 procurement window even more material. Deel folding Sastrify into Deel IT on May 5 closes the loop on the procurement-tooling side, putting the consumption-data and benchmarked-contract layer Sastrify built inside the same workforce-management surface where mid-market companies already approve software purchases.
Frequently Asked Questions
Token-based billing charges customers based on input tokens (what you feed the model) and output tokens (what the model generates, including reasoning) rather than a flat fee or fixed request count. Claude Opus 4.7, for example, costs roughly $5 per million input tokens and $25 per million output tokens. A single long agentic session can generate hundreds of thousands of output tokens.
Existing Copilot Pro and Pro+ subscribers keep their plans, but usage limits have tightened and Opus 4.5/4.6 are being removed from the Pro tier. GitHub is offering a refund window through May 20 for subscribers who cancel due to these changes. Full token-based billing is scheduled for June 2026 and will affect Business and Enterprise customers first.
A $20 Pro user running Claude Code on Opus 4.7 can consume compute worth 10× or more the subscription price, per Anthropic’s head of growth. The test aimed to determine whether pushing heavy-usage developers to Max 5× ($100/month) or Max 20× ($200/month) could align pricing with consumption. Anthropic walked back the rollout but confirmed pricing changes are under consideration.
Probably not on pricing alone, since the shift is industry-wide. Switching from GitHub Copilot to a competitor still offering flat-rate pricing is a temporary advantage — that competitor will likely follow the same economics within 6-12 months. The more durable move is to renegotiate contract terms, audit actual consumption, and build token-consumption monitoring into your vendor evaluation process.
