The week AI vendors stopped averaging the heavy user into one price

OpenAI split ChatGPT Business into two seat types. Our capture for Aug 24 → Aug 31 shows the Standard seat holding at $20/mo annual and $25/mo monthly, and a new Premium seat at $100/mo annual and $125/mo monthly carrying 5x usage and no 5-hour limit ^[5]. The same restructure capped Business at 200 employees, where the plan had been unbounded ^[5]. Coverage puts the Premium rate at the same two numbers ^[2], and a practitioner note dates the cap to the first day of our window ^[4]. The flat seat did not get cheaper; a second, richer seat was added above it.

Microsoft moved the other way. Our capture shows Voice Live Pro's Audio-Custom cached input falling from $40 to $0.40 per m tokens, which puts it on the flat $0.40 cached-input rate every other row in the Pro tier already used ^[9]. The same diff adds GPT Realtime 2 to Voice Live Pro with its own Output 2 pricing column ^[9]. Microsoft's docs describe that series as speech-to-speech models with built-in reasoning ^[7], and confirm that cached audio and text context is itself billed ^[6]. So: a correction on one row, and a reasoning model arriving as a new output dimension on the page.

Cohere added Parse 5 to Model Vault, its single-tenant inference product, at $4.00/hr or $2,500/mo for a Medium instance and $7.00/hr or $4,300/mo for XL, alongside the existing Embed and Rerank rows ^[14]. Cohere describes Parse as a cost-effective vision language model for processing large volumes of enterprise documents ^[10], and coverage of the launch notes the metered API route is priced per thousand pages, with Model Vault held out for higher-volume deployment ^[12]. The page itself says pricing depends on the models you select and each model's performance tier ^[11]. Buying the same parser by the hour is a second price for one model, aimed at steady volume rather than spiky demand.

Vercel repriced the agent plumbing. Our capture shows Vercel Connect renamed to Token Requests, the included quota cut from 5K to 500 requests per month, and the overage rate raised tenfold, from $3 per 10K requests to $3.00 per 1,000; a Triggers line arrived with 1,000 included and $0.95 per 1,000 beyond that ^[18]. Vercel's changelog states the Pro rates in the same terms ^[16], and its docs define a trigger as a forwarded third-party webhook, counting one per destination ^[15]. Two metered dimensions replaced one line item, and the free allowance on each got smaller.

Intercom let a discount lapse, which is its own kind of increase. Our capture shows the Essential plan's 'New customer offer: Save 35%' banner removed and the seat price reverting from $19/seat/mo to the standard $29/seat/mo ^[21]. The page now carries no new-customer banner, states that you pay per teammate based on your plan ^[19], and prices Fin per outcome rather than per seat ^[19]. Trade coverage dated the end of that promo to late June 2026 ^[20], two months before our capture recorded the page catching up to it.

The official framing is capacity, not price. OpenAI's announcement presents Premium as giving the most active teammates 5x more usage than Standard seats and removing the five-hour usage limit ^[1]; that page refuses automated fetches, so the wording is corroborated by coverage carrying the same two rates ^[2]. Coverage of the launch is blunter about the cause: agent-based AI burns through far more tokens ^[3]. Vercel frames its own change as pricing based on token requests and trigger events ^[16]. Cohere frames Model Vault as a managed service where pricing depends on the models you select and each model's performance tier ^[11]. None of the four described a rise.

The web is behind the pages in two places this week. Microsoft's own Voice Live documentation still names an older set of models for the Pro pricing category ^[6], while the pricing page our capture read already has GPT Realtime 2 in that tier ^[9]; on Microsoft's Q&A forum a customer trying to budget a Pro deployment wrote that they could not find the latest pricing of Azure Voice Live API ^[8]. On Intercom, trade coverage dated the end of the 35% new-customer promo to late June 2026 ^[20], while our capture only saw the banner leave the page in this window ^[21].

Practitioners are doing the arithmetic the vendors left out. A review of Parse 5 warns that if your volume is lumpy you should do the math before committing to a dedicated instance, because idle GPU time erases the per-page savings fast ^[13]. Coverage of the same launch put it behind the frontier models on a parsing benchmark while judging that it wins on cost per page ^[12]. A Vercel teardown notes that the platform fee is $20 a month, which buys one deploying team seat and comes with $20 in monthly credit ^[17], so a metered line like Token Requests draws that credit down first. On OpenAI, a practitioner note calls the 200-seat combined limit an important planning boundary ^[4].

Zoom out to the AI Monetization Monitor and the credit is the fastest-moving construct in the index. Across 2026W24–2026W36 we count 293 credit-priced AI changes against 168 in the prior thirteen weeks, spread over 155 companies, of which 128 introduced credits for the first time; agent-related changes rose to 299 from 244 ^[22]. Against that, how AI actually gets charged is still dominated by inclusion: 790 changes, or 55%, bundle AI into existing tiers, while 401, or 28%, meter usage, credits or tokens, 156, or 11%, create a new or restructured tier, and 96, or 7%, sell a separate paid add-on ^[23].

This week's leads sit in the smaller buckets, and they do what bundling cannot: charge the heavy user more without repricing everyone. That shows up in the price index, where the app basket sits at 101.1 and the infra basket at 95.2, a spread of 5.9 points ^[24]. Infrastructure keeps getting cheaper per unit — Microsoft's cached-input correction is a small instance of it — while application pricing drifts up. The gap is why the seat gets split rather than raised: a vendor whose own input costs are falling cannot easily justify a broad increase, but it can sell capacity to the people consuming it.

Volume, though, is thin. This week's 93 tracked changes are the lowest of the thirteen weeks in the window, down from 137 the week before; of the 148 changes carrying events in our capture, 53 were AI-relevant, 20 on application pages and 4 on infrastructure pages ^[25]. Across the same window 88 changes raised prices against 28 that cut them ^[24]. Read together, the week looks less like a wave than a settling: fewer pages moved, and the ones that did moved structure — seat types, metered dimensions, instance sizes — rather than a headline number.