Inference reprices by the GPU while apps fold the agent into the tier

Cohere rebuilt Model Vault's rate card around hardware instead of models. Our capture shows the old table pricing dedicated hosting per model - Embed 5, Rerank 5 and Parse 5 - at $3-7 an hour or $2,000-4,300 a month per instance ^[18]. The new table prices GPU types instead, each with an on-demand hourly rate and a promotional rate running through March 31, 2027; the A100 40GB lists at $3.25 on demand and $2.25 on promotion. The page now states that pricing follows the GPUs a model runs on and the replica count ^[1], and the docs say the same ^[2]. The page also prices the confidential-compute rows above the standard rows for the same GPU, so isolation carries a visible premium.

Together AI moved the other way on its flagship. Our capture has Qwen3.7-Max rising from $1.50/$4.50 to $2.50/$7.50 per 1M tokens, with Kimi K3.1 Pro and Ripple-65B Preview added and six models retired - Qwen3.8-2.4T, Inkling, Qwen3.5-397B, Cogito, Llama 3 8B Lite and Qwen3 235B ^[19]. The increase is a reversal rather than a first move: Together's own changelog last recorded a cut on this model, effective September 22, 2026 ^[5], and has logged nothing about it since. The page moved and the changelog did not, which is how a list-price rise reaches customers without an announcement.

Catchpoint is the week's clearest bundling move, and the page we track is now a LogicMonitor page: the former Catchpoint capabilities, Synthetics and Internet Sonar, are standard in every tier ^[9]. Edwin AI stopped being an add-on. Our capture shows the separate Edwin AI, LM Uptime and SaaS Monitoring add-ons dropped from the table and AI Credits bundled per Hybrid Unit at 20, 75 and 150 across Essentials, Advanced and Signature ^[20][8]; Synthetics and Internet Sonar now include 100, 250 and 500 points by plan. List prices moved with the packaging: Essentials $16 to $19, Advanced $27 to $37, Signature $53 down to $50 and renamed off Edwin AI, and the free trial cut from 15 days to 14.

Aircall cut the meter and tightened the terms in the same change. Our capture shows AI Voice Agents pay-as-you-go falling from $0.49 to a flat $0.29 a minute, the standalone minimum easing from $100 to $99 a month, and the bundle ladder re-cut from 500/2,500/10,000 minutes to 1,000/3,500/10,000 - now requiring an annual commitment ^[21]. The page carries both halves: $0.29 with no commitment on pay-as-you-go ^[11], and bundles billed monthly on an annual term ^[12]. A lower rate bought with a longer commitment is not a discount; it is a re-pricing of flexibility.

Appcues removed a tier and revealed a price at once. Our capture shows the Start plan discontinued, leaving Grow and Enterprise, with Grow disclosed at $9,000 a year and capped at 50,000 MAUs where the tier previously showed no price at all ^[22]. The same change set added a plain-language interface, access to a multi-agent system, AI insights and agentic campaign, guide and report building - on both plans, not the upper one. The page is explicit that nothing is held back: every feature ships ungated with every integration from day one ^[14]. Here the AI is the reason to buy the plan rather than a line item on top of it.

OpenAI made the smallest move and the most legible one. Our capture shows the GPT-5.6 Sol promotional window converted from an open-ended "at least through November 21, 2026" to a firm end date of November 30, 2026 ^[23]. Nothing about the rate changed; the optionality did. OpenAI still uses the open-ended form elsewhere on the same page - GPT-6 Astra for Law is promotional "at least through December 31, 2026" ^[16] - which makes the Sol edit a deliberate narrowing rather than a template change.

The official framing, where there is one, is about packaging rather than price. LogicMonitor says the point is to stop selling AI separately: its announcement puts Edwin AI and Digital Experience in every tier under one licensing model, with included allowances that scale by tier ^[8]. It also says why it can afford to. The company reports selling more than 826,000 Hybrid Units since launch and that more than 90% of new customers this year chose platform pricing ^[17]. Bundling a credit allowance into a unit the customer already buys is cheaper than selling a second product to the same account.

At the infrastructure layer the framing is the opposite of packaging - it is metering, described in hardware terms. Cohere now explains Model Vault as a choice of model, a choice of GPU, and a replica range, with an hourly rate for each GPU those replicas run on ^[2]. Aircall sits between the two and says so on its page: a small allowance of agent minutes comes with the seat, and everything past it is metered per minute or pre-bought in a bundle ^[11]. Three vendors, three different answers to the question of what the customer is actually renting.

Coverage tracked the application-layer move quickly and the infrastructure layer not at all. Trade press picked up the LogicMonitor packaging within days and read it the right way, as AI moving into the standard offer rather than being sold beside it ^[10]. On the model layer, practitioner tooling beat the press: the LiteLLM cost registry was corrected to the page's $2.50/$7.50 on the explicit grounds that the pricing page wins when a vendor's own sources disagree ^[6], and a proxy-side issue had already logged Together's two official pages contradicting each other ^[7]. Nobody wrote an article; the people who bill against the number noticed anyway.

Elsewhere the web is still describing products that no longer exist. Pricing guides present Model Vault as a flat rent on a dedicated single-tenant instance ^[3], and as the one track publishing rates for embedding and reranking models but not for Command A ^[4]. Both were true of our previous capture and neither is true of the page now, which prices GPUs and is indifferent to which model occupies them. Directories still list Appcues as three MAU-tiered plans called Start, Grow and Enterprise ^[15], a structure the page has retired. And Aircall's old $0.49 rate, with its $0.39 step above 2,551 minutes, is still quoted in competitor teardowns ^[13] days after the page moved to a flat $0.29.

The week's two directions are the AI monetization picture in miniature. Across the trailing 13 weeks the Monitor classes 53% of AI monetization moves as bundled into existing tiers, against 29% metered as usage or credits, 9% as a new or restructured tier, and 8% as a separate paid add-on ^[24]. Catchpoint and Appcues are the 53%; Cohere and Together AI are the 29%. The separate add-on, which is the shape most people expected AI pricing to take, is the smallest bucket on the board, and no lead company made one this week.

Credits are how a bundle acquires a limit. The credits theme now carries 459 moves against 259 in the prior period, across 241 companies, 192 of them pricing credits for the first time, with 135 expanding an allowance and 64 shrinking one ^[25]. Agents follow the same curve, 452 moves against 275 prior across 282 companies, and free AI allowances 163 against 99 ^[26]. The 20, 75 and 150 credit ladder at LogicMonitor is that pattern at full maturity: the AI is included, and the allowance is the price.

Volume is still climbing. AI monetization moves ran 203, 222 and 241 over the last three weeks, and quantified price increases reached 22 this week against 13 the week before ^[27]. Over the trailing window increases outnumber decreases 118 to 27, with 71 carrying a number, and the hike share of quantified moves is 81% against 74% in the prior period ^[28]. Half of everything we capture now touches AI, 50% against 48% prior, while the separate add-on share sits at 7% ^[29]. This week supplied 129 AI-relevant diffs out of 348 changes ^[29]. The direction is not in question; the mechanism still is.