Where this stands on 2026-11-05: a seat plus a meter, as in July, but the seat now buys less. At GitHub it buys access to Copilot and nothing else. At Microsoft it buys last generation’s models, and the newest models and the agents are metered. The meter counts the vendor’s own credit. How much usage a plan includes is still set by the gap between a model’s list price and what the vendor pays for it. Two vendors now draw that line on their own pages: Microsoft keeps the flagships out of its seat, and Anthropic gives its own flagship a separate weekly limit inside its $200 plan. The outcome unit this thread tracked in September still has one vendor, and its rate is a credit rate behind a login.

The seat becomes an access fee

On October 28 GitHub added Copilot to GitHub Enterprise (opens in a new tab) for every licensed user, and the same changelog says what that buys. The access “does not include monthly AI credits per seat”. “AI usage is metered from the first request against your enterprise budget”, and “there is no included usage allowance to track or top up”. Code completion keeps working with no budget. Chat, code review and the cloud agent stop. “An exhausted budget never removes access.” It is open to account-team customers now, with self-serve dates “in the coming weeks”, and a $10 Copilot Standalone seat for users without an Enterprise licence also carries no monthly AI credits. Until then every Copilot seat on this thread came with an allowance equal to its price. This one comes with none. Microsoft’s partner notice (opens in a new tab) of November 4 describes it as “usage rights at no extra cost”, with usage “billed through AI Credits”.

Microsoft’s own Copilot draws the line by model rather than by allowance. Nicole Herskowitz’s account of the pricing model (opens in a new tab) (September 25) puts “everyday AI” in the per-user licence under fair-use limits. That includes automatic model selection, GPT-5.6 and Sonnet 5, with Opus 5 “included with limits”. Cowork, Code, Autopilot and “frontier models like Fable and Astra” run on usage-based billing in Copilot Credits, which “builds on the USL, which is required”. So the seat buys last generation’s models and the meter buys this generation’s. That is the cost rule below, drawn by a seller that buys both labs’ models. The same November notice adds unpublished fair-use limits on Copilot Chat for users without a licence from November 2 (November 9 for organisations over 2,000 seats), and tells partners to treat users hitting them as a reason to sell the seat.

The case against reading this as the seat shrinking is that the seat still costs what it did ($21 a user (opens in a new tab) on GitHub Enterprise). Microsoft argues that metering every task would make companies ration AI, which is its reason for keeping a fixed price for everyday use. What moved is what the fixed price buys. Both moves come from one company, so this is one vendor’s direction until a second, unrelated vendor ships a seat with nothing included.

The included usage follows the vendor’s cost

OpenCode Go (opens in a new tab) costs $10 a month ($40 for Go Plus) and states its included usage “as monthly dollar amounts”, model by model, across about thirty models. Five of them come from frontier labs: Claude Haiku 5.5, GPT 6 Luna, GPT 5.6 Luna, Grok 4.7 and Gemini 4 Argon. There is no Opus, Sonnet, Fable or GPT 6 Sol, as of November 5. All five sit at the bottom of the range: $15 a month each on Go, except Argon at $30 while a 50% discount runs through November 30, against $100 for DeepSeek V4.1 Flash (raised from $80 since October 28). Open-weight models share that floor (Kimi K3, the newly added Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4 Pro are also $15). The page gives its reason in OpenCode’s own words. “For most models, we make this work through bulk discounts and reserved GPU capacity.” For the models with the lower allowance, it says, OpenCode has not yet negotiated a discount or found a cheaper host, because the model is new or its public price is already discounted: “you still get a little more than if you paid the model providers directly; this is why their included monthly usage is lower.” The page does not say which models those are. That is a vendor’s account of its own costs, which nobody outside can check. But it is the mechanism the October plans imply, stated as policy, and it covers the case they did not: a lab’s model inside the plan, priced a little above list rather than at it. The page gives no date for any of these additions. Haiku 5.5 came out on October 7, so it joined after that.

The fence is not the licence, and it is not the price per token either. Vercel’s Token Pass docs (opens in a new tab) now list all three of its featured models with the list rates the $80 credit is drawn at. Kimi K3 is $3.00 per million input tokens and $15.00 per million output, GLM 5.3 is $1.40/$4.40, and DeepSeek V4.1 Flash is $0.15/$0.60. An open-weight model at $3/$15 costs more per token than Haiku 5.5’s $0.10/$0.50 (for prompts up to 100,000 tokens). Vercel discounts it anyway, because it buys that inference from Wafer for less than list. What decides the allowance is the size of that spread. It is wide where many hosts compete to serve the same open weights, and narrow where one lab sells its own model.

The model owner draws the same line from the other side. SemiAnalysis (opens in a new tab) (October 5) measured every Claude and ChatGPT subscription tier and priced the allowances at API rates. On a $200 plan, about $2,485 of Fable 5.1 used only half the limit. OpenAI’s $200 plan ran out after about $2,897 of GPT-6 Astra, after OpenAI halved that allowance and added a $500 tier. In SemiAnalysis’s reading, subscriptions are about a tenth of Anthropic’s revenue but can take more than 40% of its inference compute, and Anthropic lowers the API-equivalent value of its pricier models within a plan. That is one firm’s measurement, from the free half of a paywalled post, and OpenAI’s own pricing page could not be checked against it. It is still the only independent figure on this thread for what an owner’s plan gives away. It says the owner’s subsidy is largest where inference is cheapest for the owner, the same fence a reseller draws.

Anthropic’s own pages now state that fence, where until now only SemiAnalysis had measured it. The $200 plan was renamed from Max 20x to Ultra (press reports date it to October 30), at the same price. Its help page (opens in a new tab) says “Ultra includes 10 times the Pro plan’s usage every week”, with no five-hour limit once an account is verified, and that usage is also capped by a separate weekly limit for Fable. The pricing page (opens in a new tab) adds “$200 a month in API credits, up from $100 on Max”. The old name said 20x and the new pages say 10 times Pro’s weekly usage; neither says whether the allowance changed, so a cut is possible and unproven. The flagship has its own ration inside the owner’s plan, and Microsoft keeps it out of the seat altogether.

The best counter-example found is LLM Gateway’s DevPass (opens in a new tab). At $29, $79 or $179 a month it is “worth 2x what you pay in model usage, metered at provider list rates”, across its whole catalogue, frontier flagships included. That would be a discount on models the vendor buys at list, except for its fence. Any model priced at $5 or more per million input tokens, or $15 or more per million output, draws on a weekly cap of 10% to 15% of the plan’s credits. A reseller that buys at list can only use price to stand in for cost, so the expensive tokens are rationed rather than excluded. The page also lists a model line-up a generation old and carries no date, so it cannot say what the plan covers today.

One aside from the same table reopens a question this thread closed in August. OpenCode Go carries Meta’s Muse Spark 1.3 Contributor at a $60 allowance, described as “heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models”. The data-for-discount trade is not a second vendor’s pricing axis yet. It is Meta’s tier, resold inside someone else’s subscription.

The October specimens

DigitalOcean’s Agent Droplets (2026-10-01) read, on the thread’s first pass, as a plan over the raw meters: $50 a month at 15% off usage, $200 at 20%, no seat charge. The pricing page (opens in a new tab) is narrower than that, and was on launch day: the discount “applies to inference on DigitalOcean-hosted open models” - Kimi K3 and GLM 5.3 are the examples - while “frontier models like Claude and GPT are available pay-as-you-go at standard rates”. The customer also chooses whether spending stops when the plan’s allowance is used or continues at list. The thread’s earlier reading of the 15-20% as a discount on usage is corrected here.

The X Subscription (2026-10-12) is the same shape at a different scale. xAI’s announcement (opens in a new tab) folds Grok, Grok Bot, Cursor and X Premium into one ladder - $0, $8, $30, $100, $200 - with “a generous pool of Grok model usage shared across Grok and Cursor”, Premium at 4x Plus usage, Super at 15x, Ultra at 50x, and “at least twice as much coding usage with our models as the comparable Cursor plan”. The pool covers Grok models. Cursor’s pricing page (opens in a new tab) now carries the ladder, says subscriptions are “only sold directly through X Subscription Corp”, and states that the pool refreshes weekly. It also states the other half. The $8 Plus tier “includes only Grok models”. From Premium up, third-party models are available at API list price and are “not included in the monthly prices shown”, which the help page (opens in a new tab) describes as “billed at API list prices from the first use and paid from your prepaid credit”. Cursor’s $20 Pro is closed to new signups; current subscribers can keep it or switch (Cursor’s forum post of the same day). The “previously $90/month combined” line is xAI’s sum of three products; a Cursor Pro user who ran Claude inside the old plan now pays $30 for pooled Grok usage and list for Claude from the first token, which is the complaint that fills the forum thread.

Vercel’s Token Pass (2026-10-15) states the pattern as a product. For $30 a month (opens in a new tab) a Pro team gets $80 of monthly credit on GLM 5.3 and Kimi K3, “more models to come”, unused credit does not roll over, and there are no automatic overage charges - requests stop. The docs (opens in a new tab) add what the changelog left out: Token Pass is early access with a waitlist, Enterprise teams cannot buy it, the requests carry zero data retention, and the models are “currently served by Wafer”, a third party. Vercel neither hosts nor owns what it discounts; it buys open-weight inference at a price it can afford to halve. A third model, DeepSeek V4.1 Flash, joined on October 20 (opens in a new tab), open-weight again, and is on the sign-up page (opens in a new tab) and in the docs. That is a 62.5% discount on open-weight models and nothing on any other, from the same company whose October gateway index (opens in a new tab) has open-weight models at more than seven in ten gateway tokens and nearly one in four dollars, and the average price per token down 47% since December (Vercel’s own routed traffic, estimated at list prices).

The owner’s end of the same rule came from a model vendor. Anthropic’s Max, Ultra and Team plans now include monthly credits for its own API: $100 on Max, $200 on Ultra, and $20 or $100 per Team seat pooled to a $500 cap (opens in a new tab), spent on the Claude API, Managed Agents and the Agent SDK, not on Claude Code or extra usage in the apps, expiring each cycle with no rollover and shown in the Console as “Promotional credits”. The Claude Platform release notes (opens in a new tab) date it October 7. This thread previously said October 15, which is a different entry. Neither the doc nor Anthropic’s credit terms (opens in a new tab) name an end date for the programme. A vendor that makes the model can give its inference away inside a seat, and SemiAnalysis’s figures say how much it does.

Read together with OpenCode Go, these plans draw one line. A vendor discounts inference in proportion to what it can buy or make below list. That is a lot for open weights that many hosts compete to serve, a lot for the owner on its own cheaper models, and a little for a lab’s model bought from the lab. Warp, which resells everything at API rates, open-weight models included (Kimi K3 and GLM sit on its list beside Claude and GPT), discounts nothing by model and prices its margin as a markup instead. Either way the token stays on the backend, and the invoice shows credits, pooled usage or a monthly allowance. For the reader the advice from September stands with one clause added: publish one worked task at list, state every cut as its effect on that bill, and state which models the cut covers and how much of the allowance each one gets. OpenCode Go’s per-model dollar table is the first pricing page on this thread that does all of that.

The outcome is a credit rate behind a login

The September reading was that the meter’s unit was moving to the outcome. On 2026-10-28 Sourcegraph’s pricing page (opens in a new tab) still lists the Enterprise plan “starting at $16K” and “includes credits for AI features”, with Agentic Batch Changes listed as a feature and no rate. The credits themselves are described more fully now: included per user, pooled across the org, with “No monthly credit expiry”, unlike Anthropic’s or Vercel’s, which lapse each cycle. The September 14 changelog (opens in a new tab) says how the merge is actually billed: “you pay per changeset merged into your codebase” and “final credit rates are published in Enterprise Portal”. The outcome is denominated in the vendor’s credits, priced where a prospect cannot read it. No second coding agent bills per merge; the adjacent public units are effort units - CodeRabbit’s $0.40 per agent minute and $0.25 per reviewed file.

Scott Woody’s essay on Stripe’s blog (opens in a new tab) (2026-10-01) draws the distinction this thread had been blurring: output pricing charges for a verifiable action (a resolved ticket, a merged changeset); outcome pricing charges for the customer’s result, and he calls it impractical outside high-ticket enterprise deals. His progression runs tokens, then unified credits, then output, with “roughly one in six (and growing)” Stripe users past key revenue milestones on hybrid pricing as of August 2026 (Stripe’s book; Metronome sells the billing layer). The independent numbers point the same way at a slower pace. High Alpha’s (opens in a new tab) 2025 benchmark has companies that monetise AI at 53% subscription, 31% hybrid, 11% usage and 5% outcome; Kyle Poyar’s (opens in a new tab) 230-company survey (April to May 2026, Metronome-supported) has AI credits at 29% with another 33% planning them within a year. Stripe’s own survey of 200 SaaS platform leaders (opens in a new tab) (March 2026, half of them Stripe users, published October 19) has seats as the primary model falling from 16% to a planned 11%. One in four plan to “keep per-seat fees, then charge separately for agent work based on credits, usage, or outcomes”. The unit that is spreading is the credit.

Elena Verna, who leads growth at Lovable and says so, named the cost of that on 2026-10-16: credits, tokens and compute units (opens in a new tab) are “arcade money”, and the worst case is a multiplier with no baseline - GrokBot’s “4x” and “15x” of an unstated unit, which is now also Cursor’s pricing page. Her remedy is legibility rather than a different unit, which is the thread’s too.

The meter by default, postponed twice

Microsoft means to make the meter the default at a larger seller than any other on this thread, and keeps moving the date. Its October Partner Center announcement (opens in a new tab) (posted October 1, updated October 14) switched pay-as-you-go on by default for new Microsoft 365 Copilot Business licences bought through partners from December 1, a month later than first announced. On November 3 the November page (opens in a new tab) moved it again: “We heard your feedback and are adjusting the effective dates”. The rollout now begins in “Microsoft’s FY27 Q3 (January-March 2027)”. The terms did not move. A new purchase also switches the meter on for the tenant’s existing Copilot Business licences, and “disabling usage-based billing might limit access to AI experiences”. The default spending limit is still “4,000 Copilot Credits per user per month, which can be adjusted by admins”, or up to 400,000 credits a month for 100 users, Microsoft’s own example. The meter has slipped twice before billing anyone, and the stated reason is partners asking for time. The billed experiences include Copilot Cowork, Work IQ APIs and “GitHub Copilot Harness”. The same page retires AI Builder capacity add-ons from November 1 and tells partners to “plan to use Copilot Credits” instead. Microsoft’s billing docs (opens in a new tab) call those credits “a common currency for eligible Microsoft services”. A spending policy covers future services automatically, because “Auto-apply new services” is on by default. Neither page states a price per credit. The seat stays and the meter is on unless an admin turns it off, in the vendor’s own currency, and it widens to whatever ships next.

The thread had the Copilot test half-described. GitHub announced on 2026-04-27 (opens in a new tab) that every plan moved from premium requests to GitHub AI Credits on June 1, priced on tokens at each model’s API rate, with each plan’s included credit equal to its price (Pro $10, Pro+ $39, Business $19 and Enterprise $39 per seat), promotional boosts for existing Business and Enterprise customers through August, and the fallback models removed. Today’s docs (opens in a new tab) split each allowance into base and flex credits - Pro 1,000 plus 500, Pro+ 3,900 plus 3,100, and a $100 Copilot Max at 10,000 plus 10,000 - at a cent a credit. The August 28 changelog (opens in a new tab) is narrower than “every seat”: prepaid seats from September 1 for new Business and Enterprise customers and October 1 for existing ones apply to accounts paying by credit card or PayPal, prices unchanged, and “additional payment may be required for you and your users to continue using Copilot” past the included usage. The bill-shock stories that exist describe June’s switch and the September 1 promo lapse. The one public admin report since prepaid seats began runs the other way. On GitHub’s community forum (opens in a new tab) on October 8 and 9, a Business admin says the included pool stayed at 3,800 credits after the org went from two seats to five, leaving users blocked, with a support ticket open and no staff reply. That is the meter under-delivering, not over-billing. The loop that waited for a bill-shock story from prepaid seats closed empty on October 31: searches on November 5 found the June and July coverage and nothing from an admin since October 1. The meter was not shown to be dishonestly priced, and the answer GitHub gave instead was a seat with nothing included at all. One billing story did land in the window, and it is not a pricing one: an Anthropic Console top-up of 5.04 charged as $504 through a locale bug, on Hacker News (opens in a new tab) on October 16, acknowledged by Anthropic.

One more correction. Gemini 4 Argon’s $2/$10 was described here as a cut with no end date. The Gemini API pricing page (opens in a new tab) states one: $2.00 and $10.00 per million tokens through November 30, 2026, $4.00 and $20.00 from December 1. A promotion with a published end date is a term, and Argon is not Google’s only one. The same page doubles Gemini 3.8 Flash and 3.6 Flash, from $0.75/$3.75 to $1.50/$7.50, on January 1, 2027, and three more models step up the same day. The page does not call any of them promotions. Google states several models’ prices as dated rates, and a buyer quoting a Gemini bill needs the date beside the rate.

How the thread got here

The answer we had in July was tidy. AI features in developer tools price as a per-seat licence plus a metered consumption line - the seat pays for access, the meter pays for tokens. It is the shape the category converged on without much argument, and the guide carries it in §01.

August broke the tidiness in one move. Meta launched Muse Code (opens in a new tab) with standard pay-as-you-go pricing at $1.25/$4.25 per million input/output tokens, and alongside it a “contributor” tier whose prompts and completions “may be used to improve our products”, rate-limited over a rolling five-hour window and available in select countries. Meta’s post states no contributor price. The $0.10/$0.20 figure - roughly 12x cheaper on input and 21x on output - is reported by Channel Insider (opens in a new tab) (August 6) from Meta’s pricing listing, which this thread has not been able to open. That is not a volume discount and it is not a seat. It is a third axis: the customer pays with data, and the discount puts a public market rate on what that data is worth. The immediate consequence is concrete enough to act on, and it is already a claim: a training-data policy is now a pricing decision, so it belongs on the pricing page rather than buried in a DPA.

Meta ran both directions at once: the same month it open-sourced Muse Glimmer - 30B parameters, Apache 2.0, on a single consumer GPU with no network call. The careful reading was that the data trade is a beta acquisition mechanism rather than a settled axis, and the incumbent clock supported it: the loop gave incumbents until the end of August to ship or rule out a data-for-discount tier, and it ran out on 2026-08-31 with nothing dated either way. The contributor tier survived a model revision (Muse Spark 1.3, September 2) with the same spread, which weakens “quietly disappears at GA” without settling it.

The week of September 7 answered the thread’s first open loop, once. Atlassian added two meters beside its Rovo AI credits - automation steps and AI agent resolutions - announced September 1 with billing from December 3, admin usage views and limit controls shipped at once, and the allowance (400 steps per Jira Standard user a month) and overage ($0.50 per 1,000 steps) published in its support docs. An automation step is not an AI feature: the first non-generative unit to reach a usage meter since this thread opened, and it says the meter has become a billing habit rather than a pass-through of model cost. One vendor is not a pattern, so the loop stays open for a second; none was found through 2026-10-28. The sequencing is the part worth keeping: the cap shipped a quarter before the meter.

The unit moved to the outcome, for a week

The week of September 14 changed what the meter counts. Sourcegraph took Agentic Batch Changes to general availability on September 16 billed per changeset merged into the customer’s codebase - not per token, seat or attempt, and nothing for a pull request the team rejects - and said plainly which default it was rejecting: every other coding agent charges for effort, per token with a fee on top, whether or not anything lands. The reason it moved then is that the token stopped being a price a vendor owns: Vercel’s September index put the average price per token down 23.2% in August, and UC Berkeley and Arena’s HarnessTax study ran GPT-5.6 Sol through three coding-agent harnesses at $1.540 (Claude Code) and $0.441 (Pi) per resolved task - a 71% gap - with no quality difference surviving correction (Pan, Yang, Arabzadeh, Chiang, Stoica and Zaharia; published at harnesstax.github.io, not on arXiv). Two things kept this from being the answer then and still do: outcome billing means the vendor eats every failed attempt while the customer defines “merged”, which suits an enterprise contract and not a $20 seat, and the rate is not public. What October added is that it is not a rate in dollars at all.

The rate stopped being the claim

The week of September 21 showed what a vendor says when its rate will not hold still. Anthropic’s Opus 5.5 (September 22) cut list 20% to $4/$20 per million tokens and led with “at default settings it will cost 40% less than Opus 5 on typical workloads”, with every benchmark table stated as a cost per task. OpenAI’s GPT-6 Sol and Luna, the same day, halved the GPT-5.6 rates and pitched $0.27 per task - a comparison that sets two effort levels against each other. Two cuts of 20% and 50% in one day, on top of a token price already falling monthly, mean the rate cannot carry a positioning claim; the bill for a defined job can, and both vendors moved to it. The product side moved the same way from below: Vercel listing billable duration and CPU minutes on every deployment, DigitalOcean’s Managed Agents billing per second of active CPU with idle waits at $0. A bill is still the number the vendor defines - “typical workloads” appears on no pricing page - so the per-task figure is a rate with better copy until the workload ships beside it.

The margin became a number and the rate became a dial

The second week of October gave the thread a vendor stating its resale margin on the page. Warp Factories’ pricing (opens in a new tab) has Pay as you go at $0 a month with “Usage at 20% above API rates”, Build at $20 and Business at $50 a user with “Usage at API rates”, Enterprise custom, under a footer that factory usage is “billed separately based on inference, compute, and agent usage” (whether the 20% covers compute is not said; unchanged on 2026-10-28). The arithmetic the page invites: the subscription is the price of not paying the markup, and at list $20 a month pays for itself past about $100 of model usage (our arithmetic). Warp buys every model it sells at API rates, so its lever is the markup. The October vendors make, host or buy some of their inference below list, so theirs is the discount. Same decision, opposite sign.

The rate underneath moved from a number to a setting the same week. Claude Haiku 5.5 (opens in a new tab) (October 7) lists $0.10/$0.50 per million tokens up to a 100,000-token context and $0.50/$2.50 above it, with an effort setting from Low to Max so the buyer decides “whether to optimize for cost or intelligence”. No per-level price is stated, so the dial moves tokens spent rather than the rate. A markup on a rate the buyer can turn, or a pooled allowance of an unstated unit, is a bill nobody can quote in advance; the worked task at list belongs beside the rate, the markup and the model list.

Why it moved now

The pricing question got sharp because the product question went flat. Three terminal coding agents launched inside forty-eight hours in early August - Muse Code, Memcode, Clark Code - and none competed on what the agent can do: Meta on price-as-data-trade, Memcode on repository memory, Clark on an Apache-2.0 licence. When nobody is arguing about capability, the terms sheet becomes the product. October sharpened it again from the other side. When the model is a commodity whose price fell 47% in ten months, a vendor can only afford to discount inference it gets for less than list, and how much less differs model by model. That is why the plan now says which models it covers and how much of each, and why the reader should check.

What would settle it

5 open loops

  1. 01 By end of 2026

    Does a *non-generative* feature get metered? Answered once on 2026-09-01 - Atlassian meters automation steps at $0.50 per 1,000 above a per-seat allowance, billing from December 3 - so the meter is a billing habit for at least one vendor. A second vendor metering a non-AI unit makes it the pattern; none by the end of the year leaves Atlassian as one company's billing reform. No second vendor found through 2026-11-05.

  2. 02 By end of 2026

    Does the outcome unit become public pricing - Sourcegraph's per-merge rate on its pricing page, or a second coding agent billing per merged changeset? One vendor with a hidden rate is a sales motion described in outcome language; a published number or a second vendor makes it the fourth axis. As of 2026-11-05 the rate is a credit rate "published in Enterprise Portal", not on the page.

  3. 03 By end of January 2027

    Does a flagship frontier model (Opus 5.5, Fable 5.1, GPT-6 or GPT-6.1 Sol, Astra, Gemini 4 Carbon) enter a reseller's fixed-price plan at well above list value - on OpenCode Go's model table, Token Pass's featured models or DigitalOcean's discount scope? So far resellers carry only the labs' small models, at the bottom of their allowances. A flagship above list would mean a lab is selling resellers a real discount on its best model, and the narrow half of the rule would fall.

  4. 04 By end of April 2027

    When Microsoft switches pay-as-you-go on by default for new Microsoft 365 Copilot Business licences bought through partners (moved on 2026-11-03 from December 1 to January-March 2027, the second slip; default cap 4,000 Copilot Credits per user a month), does an admin publish a first bill that surprised them? The date moved with Microsoft's. A third slip, or the default dropped, says the default-on meter did not survive its partners.

  5. 05 By end of February 2027

    Does a seat with no included AI usage spread beyond GitHub? GitHub Enterprise has included Copilot with no AI credits since 2026-10-28, for account-team customers, with self-serve dates still to come. GitHub adding credits back at self-serve, or a second vendor outside Microsoft shipping the same shape, would settle whether the seat is becoming an access fee.

Each is checkable by a date. If one goes quiet, that is itself an answer.