A one-line seat price hold survived all three GitHub Copilot repricing moves in June 2026 and blocked none of them, because the vendor changed the meter instead of the price

Between June 1 and June 30, 2026, GitHub moved chat and agent features to usage-based AI Credits, launched Cowork at $1 to $7-plus per task on top of the $30 seat, and expired volume discounts. Sticker price never moved. A price hold that only names the seat rate is a clause that pays out nothing, and the fix is four limbs, not one sentence.

Prepared by Redress Compliance · August 26, 2026 · GenAI and agent licensing advisory. Enterprise renewal and rate-card engagements, 2025 to 2026.

The seat rate is the only number most price holds name, and it is now the least interesting number in the contract.

GitHub's June 2026 sequence proves the point with unusual clarity: chat and agent features moved to usage-based AI Credits on June 1, Cowork launched with per-task billing on June 16, and volume discounts expired June 30.

A clause reading "seat price is held at $30 per user per month for the term" was true, enforceable, and completely inert through all three. The vendor never touched the surface the buyer had protected.

Treat a GenAI contract as carrying four independently movable price surfaces plus a fifth soft one: seat rate, per-task or per-action rate, per-token rate, bundled inclusions, and the discount that sits on top of all four. Hold one, and you have left four routes open.

That is the same structural problem covered in the broader GenAI and agent redline set, applied to price rather than rights.

Here is the language, limb by limb. Limb 1 (seat and substrate): "The net per-unit fees for each SKU listed in Exhibit A shall not increase during the Initial Term or any Renewal Term.

This hold applies to the net price after all discounts stated in Exhibit A, and extends to any license or subscription that Vendor requires as a prerequisite to use of the listed SKUs." That last sentence is the whole point.

Without it, the seat is held at $30 while the E3 underneath goes from $36 to $39 and your all-in moves anyway.

Limb 2 (credits and conversion): "Vendor shall not increase the unit price of Credits, nor increase the number of Credits consumed per unit of work, above the rates set out in Exhibit A.

For clarity, this includes but is not limited to Credits per standard action, per voice action, per task, and per agent invocation." This is the limb the vendor will resist hardest, and the resistance is the tell.

GitHub kept credits at $0.01 while task cost varied with model selection, context retrieved, tool calls, and runtime, all of which the vendor controls and the buyer cannot see. Holding $0.01 and leaving conversion floating protects nothing.

Limb 3 (tokens and successors): "Token rates for each capability class in Exhibit A shall not increase.

Where Vendor retires a model, the replacement model made available to Customer for the same capability class shall be charged at no more than the retired model's rate for the remainder of the Term." Bind rates to capability class, never to a model name.

GPT-5.4 and GPT-5.4 mini retired in Codex on August 31, 2026 and o3 left ChatGPT on August 26, 2026. A hold naming those models expired with them.

Also fix the internal ratios: cached input at one tenth fresh, output at six times fresh, because those ratios are where a nominally flat card gets expensive.

Limb 4 (no degradation of inclusions): "Functionality included in Customer's subscription tier as of the Effective Date shall remain included in that tier for the Term at no additional charge, notwithstanding any repackaging, tier restructuring, or SKU end-of-sale by Vendor." ServiceNow's April 9.

2026 restructure, with legacy SKUs going end of sale July 1 and legacy pricing explicitly unreinstatable, is the case study.

All four limbs attach to Exhibit A: Rate Card, dated and version-stamped, listing SKU, list price, discount, net price, credit conversion ratios, and token rates by capability class. The exhibit is not administrative housekeeping.

OpenAI states plainly that the signed agreement controls and the public card is a list-price map, which is the vendor conceding the exhibit is the only durable record.

Renewal teams working through the 2026 GenAI renewal cycle should treat a missing or undated exhibit as a failed hold regardless of how the clause reads.

The June 2026 Copilot sequence is not three separate pricing actions. It is one architecture, executed three ways. On June 1 chat and agent features moved behind AI Credits at a cent each.

On June 16 Cowork arrived at roughly $1 to $3 per light task, $4 to $7 medium, and $7 and up for heavy work, stacked on the $30 seat. On June 30 volume discounts expired, returning below-list agreements to $30 at their next renewal.

A price hold naming the seat rate survives all three because in each case the seat rate is exactly the number the vendor did not need to move. That is not bad luck in drafting. It is the predictable result of negotiating against a meter the vendor designed you to negotiate against.

Seat economics and inference economics behave differently, and buyers who treat them the same lose money.

Under seat pricing, price and revenue moved together: the vendor's marginal cost of an additional user was near zero, so the unit price was the whole commercial conversation and a hold on it held everything.

Under inference pricing the vendor carries real variable cost per unit of work, and that cost moves with model choice, context retrieved, tool calls made, and runtime, none of which the buyer can observe on an invoice.

A vendor that hands you an absolute freeze on the work-to-billing conversion has handed you an open-ended cost exposure on its own P&L. It will not do it, and it is not being tactical when it refuses.

That distinction is worth money if you use it. The vendor's resistance to a conversion freeze is honest; its resistance to a cost-per-unit-of-outcome cap is not.

In our experience across these negotiations, teams that walk in demanding a frozen credits-per-task ratio get a flat no and spend their capital getting nowhere.

Teams that ask for a fixed unit price on credits or tokens plus a bounded annual drift on the conversion ratio, typically 10 percent, get it, because the vendor keeps its ability to reprice work as its cost base moves while surrendering the ability to reprice you by five times overnight.

Three channels do the actual repricing, and they are the same mechanism wearing different clothes. Model retirement is the first: GPT-5.4 and GPT-5.4 mini leave Codex on August 31, 2026, o3 leaves ChatGPT on August 26, 2026, and GPT-4o went in February 2026.

A rate held against a model name expires with the model, which is why holds must bind to a capability class with a successor cap.

Bundling is the second: ServiceNow replaced five tiers with three on April 9, 2026, made Now Assist, Moveworks and AI Control Tower non-optional, and put legacy SKUs end of sale on July 1 with no reinstatement, while Google runs consumption charges past subscription quota on Gemini Enterprise.

Discount expiry is the third, and it is the crudest: nothing about the price changed on June 30, only the number you actually paid.

Each channel exploits the same gap. The buyer negotiates the unit of billing. The vendor controls the unit of outcome and the mapping between them.

Hold one and you have held nothing, which is why the drafting standard in GenAI and agent contract redlines treats price protection as a floor and a ceiling on delivered work, not as a line item on a rate card.

The practical rewrite is short. Stop drafting a hold on cost per unit of billing. Draft a hold on cost per unit of outcome: this class of task, this volume, this total cost, for this term, whatever the meter is called on the day the invoice prints.

That framing forces the vendor to price its own repricing rights, and it converts an argument you lose about models into an argument you win about arithmetic.

Four objections will arrive in a predictable order, and each has a trade that gets you most of the protection without asking the vendor for something it genuinely cannot give. Treat them as pricing conversations, not drafting ones.

The same discipline applies to the reservation mechanics discussed in the Azure OpenAI negotiation guide: what you cannot freeze, you cap. | Vendor objection | The counter | What it is worth against June 2026 | | "We cannot hold token rates, models change." | Bind rates to a capability class.

Not a model name, with a successor cap at or below the retired rate. | Covers the August 26 and August 31.

2026 retirements and the February 2026 GPT-4o precedent at zero cost to the vendor's roadmap. | | "Credits are already fixed at one cent." | Then holding the credit-to-task conversion costs you nothing.

If not, cap conversion drift at 10 percent per year. | Neutralizes the June 1 AI Credits move and the $1 to $7-plus Cowork task spread, the only variable the buyer cannot see. | | "Bundled AI is a value increase.

Not a price increase." | No-degradation limb plus a right to decline the bundled tier at prior net price for the balance of the term. | Directly answers the ServiceNow April 9, 2026 pattern where legacy SKUs went end of sale July 1 with no reinstatement. | | "The discount is a promotion.

Not a price." | Hold net effective price per seat, not list, with the discount stated as a contractual rate for the full term. | Blocks the June 30, 2026 volume discount expiry that returned below-list Copilot seats to $30 at renewal. |

Sequence matters. Concede the model-name freeze early and loudly, because you were never getting it and the vendor will read the concession as good faith. Trade it for the successor cap and the drift ceiling.

Then hold the net-price and no-degradation limbs to the end, when the deal is otherwise closed and the cost of reopening exceeds the cost of agreeing.

Walk in with a target sheet, not a principle. The seat rate is the easy limb and the one the vendor will concede first, so do not spend goodwill there: get $30 flat for the full term in writing and move on. The limb that actually bleeds money is underneath it.

Microsoft moved E3 from $36 to $39 and E5 from $57 to $60 on July 1, 2026, which means an all-in Copilot seat went to $69 on E3 and $90 on E5 without the Copilot line item changing at all.

Bind the qualifying base license to CPI or 3 percent, whichever is lower, or your held $30 sits on a floor that rises 8.3 percent underneath it.

On the consumption side, hold the credit unit at $0.01 and, more importantly, hold the credit-to-task conversion flat through year one with a 10 percent cumulative cap on drift thereafter, because task cost is driven by model selection, context retrieval, tool calls and runtime.

All of which the vendor controls and none of which you can audit.

For tokens, hold by capability class rather than model name and preserve the two structural ratios: cached input at one tenth of fresh, output at six times fresh, carried onto any successor model. Freeze inclusions at signature scope so bundled AI cannot be quietly promoted into a higher tier.

Our experience across these deals is that buyers who ask for all four limbs get three; buyers who ask for one get one.

Price the walk-away before you table any of this, because a hold clause without a comparator is a request. ChatGPT Enterprise reports at $45 to $75 per seat, roughly $60 average, with a 150-seat minimum and a realistic entry near $108,000 per year.

Gemini Enterprise starts from $21 with consumption beyond quota. Microsoft's own E7 at $99 all-in bundles E5, Copilot, Entra Suite and Agent 365 and is not moving on price, which makes it a genuine internal alternative to a $90 E5-plus-Copilot stack that has three unheld meters attached.

Say those numbers out loud in the room. The renewal strategy work should already have modeled which of those you can actually execute.

The pattern is consistent enough to cite as precedent rather than anecdote.

ServiceNow collapsed five legacy tiers into three on April 9, 2026, bundled Now Assist, Moveworks, Workflow Data Fabric and AI Control Tower into every tier, put legacy SKUs end-of-sale July 1, 2026, and stated plainly that legacy pricing cannot be reinstated afterward.

Google withdrew the $20 and $30 Gemini add-ons, made AI Ultra Access unpurchasable, and repositioned AI Expanded Access at roughly $20 per user with an annual commitment. Salesforce runs two incompatible meters and lets the buyer pick the wrong one.

OpenAI retired GPT-4o in February 2026, then scheduled o3 out of ChatGPT on August 26, 2026 and GPT-5.4 out of Codex on August 31, 2026, which turns a roadmap calendar into a repricing calendar for anyone whose rate card names a model.

Across these five vendors, roughly three quarters of the 2026 price movement arrived through packaging, tier collapse, meter substitution and model retirement, not through a rate card line going up.

That is the argument to put in front of the account team: you are not asking for protection against a hypothetical, you are asking for protection against what their peers did last quarter. The same logic drives the redline set for agent contracts.

Related advisory: the full topic guide.

Does a seat price hold protect me if the vendor adds per-task billing?

No. GitHub Copilot Cowork launched on June 16, 2026 with per-task charges of roughly $1 to $3 for light tasks, $4 to $7 for medium and $7 or more for heavy, all sitting on top of the unchanged $30 seat.

A hold that names only the seat SKU leaves the vendor free to introduce a second meter for work the seat previously covered. The fix is a limb that holds the scope of the seat, not just its price.

How do I hold a token price when models get retired every few months?

Bind the rate to a capability class rather than a model name, and add a successor-rate rule stating that any model designated by the vendor as the replacement for a retired model is billed at no more than the retired model's rate for the remainder of the term.

GPT-5.4 and 5.4 mini retire from Codex on August 31, 2026 and o3 leaves ChatGPT on August 26, 2026, so a rate tied to a model name has a shelf life measured in months. Also preserve structural ratios: on the GPT-5.6 card, cached input is one tenth of fresh input and output is six times fresh input.

What is a credit-to-task conversion hold and why does it matter more than the credit price?

Copilot Credits are priced at $0.01 each, which sounds like a firm number. But task cost depends on the model used, the context retrieved, the tool calls made and the runtime, all of which the vendor controls and the buyer cannot see.

The vendor can hold $0.01 indefinitely and still double your bill by increasing the credits a given task consumes, so the enforceable hold is on how many credits a defined unit of work consumes.

Can I get a price hold on bundled AI features?

You can get a no-degradation limb, which is different and often more valuable.

When ServiceNow collapsed five tiers into three on April 9, 2026 and bundled Now Assist, Moveworks, Workflow Data Fabric and AI Control Tower into every tier, capabilities including Change, Problem, Major Incident and Process Mining moved to Advanced or above.

The clause to draft states that every feature available to you at signature remains available at your contracted tier for the term, regardless of repackaging.

Is a discount expiry the same as a price increase?

Commercially yes, contractually no, and that gap is where buyers lose money. Copilot volume discounts expired June 30, 2026 and agreements priced below list return to $30 at their next renewal, with no list price change involved.

Draft the hold against net effective price per seat per month, defined as the amount actually invoiced divided by billed seats, rather than against list price.

What should I attach as an exhibit to make the clause enforceable?

A version-stamped, dated copy of the vendor's current rate card, covering seat SKUs, credit or action rates, token rates by class, and the list of features included in your tier.

OpenAI states explicitly that the signed agreement controls and the public card is a list-price map rather than a replica of every invoice, which is exactly the argument for pinning today's card into the contract.

Without the exhibit, the hold refers to a document the vendor can rewrite unilaterally.

How long should a GenAI price hold run?

Match it to the full initial term and add a defined renewal cap so the hold does not simply expire into an unprotected repricing. In a market where three pricing changes landed inside thirty days in June 2026, a twelve-month hold is a twelve-month reprieve, not protection.

Pair the hold with an uplift cap at renewal and a right to renegotiate on a triggering event such as meter substitution or tier consolidation.