AI usage graduates onto its own dedicated meter, decoupled from the base plan
On 2026-07-06 three platform vendors pulled AI usage onto a dedicated billing currency that no longer inherits the base product's price. Snowflake Cortex split AI features onto an edition-independent 'AI Credit' ($2.00 global / $2.20 regional), so identical AI usage costs the same on Business Critical as on Standard. Vercel shipped Vercel Agent as a stand-alone $0.25/1M-token line rather than folding it into the $20 seat. Anthropic added a Microsoft Foundry Claude Consumption Unit meter ($0.01/CCU). All three price AI as a thin fee over pass-through, on its own axis — routing, model, or marketplace — rather than the base seat or edition.
What's happening — and why
What's happening: instead of bundling AI into the per-seat price or letting it inherit the account's edition/plan rate, some platform vendors are giving AI usage a billing currency of its own. Snowflake made Cortex AI bill in an 'AI Credit' priced only by cross-region routing — flat across every Snowflake edition — while the warehouse compute that runs each AI query still bills at the edition rate. Vercel launched Vercel Agent as a token-metered line item ($0.25 per 1M tokens plus pass-through model cost) rather than sweetening the $20 platform seat with 'free' agentic Code Review. Anthropic added Microsoft Foundry as a Claude Consumption Unit marketplace meter, converting token usage to CCU at $0.01 each.
Why: AI inference is becoming the largest and most volatile cost-of-goods line in a software P&L. When AI is folded into a seat or an edition, a model-price cut can't reach the buyer without re-rating the whole plan, and a usage spike quietly crushes the margin on a bundled tier. A dedicated AI meter lets the AI price float on its own axis, so a cheaper route or model flows straight through — and it makes AI spend separately forecastable and auditable, which is exactly what a FinOps function needs.
How it works
Evidence over time
14 supporting · 7 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| Snowflake Cortex | Jul 2026 | Split every Cortex AI feature (AI Functions, Cortex Search / Batch Search, Cortex Agents, Snowflake Intelligence, Cortex REST API, AI Parse Doc) onto a new edition-independent 'AI Credit' priced only by cross-region routing — $2.00 global / $2.20 regional — so identical AI usage now costs the same across Standard/Enterprise/Business Critical. Warehouse compute and storage stay on edition-priced Platform Credits ($2/$3/$4). Capacity discounts explicitly do NOT apply to AI Credits; only automatic ACV-based discounts do. |
| Vercel | Jul 2026 | Launched Vercel Agent (BETA) as a stand-alone AI SKU billed $0.25 per 1M tokens plus pass-through model cost on Pro (Custom on Enterprise) — a thin token-metering fee, not a per-seat or per-action charge, on the same zero-markup rails as v0 and AI Gateway. Refused to bundle agentic Code Review / Investigations into the $20 platform seat, keeping AI as its own line item. |
| Anthropic | Jul 2026 | Added Microsoft Foundry to Claude Platform (alongside AWS) as a Claude Consumption Unit marketplace meter — token usage rated in USD then converted to CCU at $0.01 each (100 CCU = $1.00) and billed in arrears via the hyperscaler marketplace, a distinct AI-usage currency layered over the direct per-token API. |
| Frase | Jun 2026 | Rebuilt packaging as a 'content operating system' with per-plan AI-generation meters distinct from the base plan fee and overage switched to opt-in PAYG — an app-layer step toward separating the AI meter from the seat, though still inside the plan rather than a wholly independent currency. |
| Weights & Biases | Jul 2026 | Added ARIA as a distinct 'Token based pricing' product line on its pricing compare table — free for a limited time on Free and Pro, included on Enterprise — separate from the existing seat, storage ($0.03/GB), and Weave ingestion ($0.10/MB) meters. AI given its own metered line on an MLOps platform, decoupled from the base tier meters. |
| Jul 2026 | Added the AlphaEvolve agent to Vertex AI as a stand-alone billable line priced as the base Gemini model rate plus a separate 2x agent surcharge (3x all-in), itemized apart from the underlying token rate. The agent gets its own price line/multiplier independent of the base model meter — a hyperscaler decoupling the agent from the model it runs on. | |
| Linear | Jul 2026 | The app-layer/seat-based confirmation this trend explicitly asked for, and the clearest case of a seat being unbundled feature by feature. 'Linear Agent automations (beta)' — previously listed in the Business plan card and the AI-and-agent-workflows feature matrix as included free in the seat — was renamed 'Loops' and now carries the same '** Requires AI credits' footnote as Coding Sessions, making it the SECOND AI capability Linear meters outside the flat per-seat price. Seat prices unchanged (Free $0, Basic $10/user/mo, Business $16/user/mo, both billed yearly). Code Intelligence (beta) and Triage Intelligence remain bundled at no extra charge, so one pricing page now carries both the decoupled and the bundled treatment. |
| Jasper | Jul 2026 | A pure per-seat vendor with NO meter of any kind — no tokens, no generations, no credits — added an AI-usage meter on top of the seat and published the rate card. The pricing FAQ's new entry states that 'some usage-based features, like the GEO Hub and Agents, run on credits, which sit alongside seat-based pricing on the Business plan', with admins able to view and govern workspace credit allocation. The public Jasper Rate Card (stamped 'Updated 6.9.2026') itemizes 10 credits per Grid row, 10 per GEO Hub query or page run, 40 per Research or Translation Agent run, 100 per Optimization Agent run, and 1-40 credits per API/MCP call; Prompts and standard Agents are unlimited and standard Chat/Canvas content creation is explicitly not credit-consuming. Pro held at $59/seat/mo yearly ($69 monthly). Metering reintroduced three years after Jasper killed its launch word caps. |
| Freshworks | Jul 2026 | A seat-based CRM incumbent publishing a per-unit AI rate for the first time, keeping the AI meter strictly separate from the seat: the Freddy AI Agent add-on is listed at $49 per 100 bot sessions (~$0.49 per session), offered identically on Growth, Pro and Enterprise, after two years of quoting session packs through sales. A bot session is any unique end-user engagement. Seat prices did not move ($9/$39/$59 per user/mo billed annually; $11/$47/$71 monthly), and Configure-Price-Quote was separately priced at $19/user/month — so AI and CPQ both sit as their own metered lines above an unchanged seat, while the $0 three-user Free plan was removed. |
| Modal | Jul 2026 | A platform-layer adopter giving AI inference its own unit rather than its own tier: Modal launched an OpenAI-compatible Shared API metered by TOKEN, promoted on its pricing page ('Kimi K3 is live. Try the new Shared API with token-based pricing'), running alongside — not instead of — the existing per-GPU/CPU/memory-second rate card. It is Modal's first departure from pure per-second compute billing since founding, giving customers a choice between dedicated per-second capacity and shared per-token inference for supported models, with Starter's $30/month free credit applying to Shared API usage too. Per-token rates were not yet published on the pricing page or billing docs at capture. |
| Braintrust | Jul 2026 | Meter proliferation on a platform that already ran three: data retention became Braintrust's FOURTH meter at $0.50 per GB per month on Pro past a 30-day included window, shown both as a '+ $0.50/GB/mo' badge on the plan card and in a comparison row renamed from 'Default retention' to 'Included data retention'. Retention had been a plain feature gate (14 days Starter / 30 days Pro / custom Enterprise) with no published way to buy more short of an Enterprise contract. It joins model/Topics token credits ($10 included on Starter, $249 on Pro, then $0.06/mtok input and $0.40/mtok output), processed data (1 GB / 5 GB included, then +$4/GB and +$3/GB) and scores (10K / 50K included, then $2.50 and $1.50 per 1,000). Tier fees unchanged. |
| Glean | Aug 2026 | The mechanism this trend hypothesised, demonstrated. Glean's Model Hub Usage rate card (docs.glean.com, stamped 2026-07-30) cut GPT 5.6 Terra 20% across every dimension (input $2.50 to $2.00, cache write $3.125 to $2.50, cache read $0.25 to $0.20, output $15.00 to $12.00) and GPT 5.6 Luna 80% ($1.00 to $0.20, $1.25 to $0.25, $0.10 to $0.02, $6.00 to $1.20) — the first published price decreases on that card since Glean started disclosing dollar rates in July 2026. The credit-denominated side moved in lockstep: the Enterprise Flex model-tier docs (footer redated 2026-08-04) moved Luna from Premium to Standard, so it now falls inside the included 100/user/week Thinking and Adaptive Reasoning allowance instead of always consuming FlexCredits. No seat price, plan structure or Enterprise Flex commitment changed. A cheaper model reached the AI meter without re-rating the account — which is precisely the unit-economics argument this trend was built on, observed rather than inferred. Glean restructured the same card again on 2026-08-11, splitting GPT Realtime 1.5/2/2.1 into per-modality rows and adding GPT-4o Mini TTS and Deepgram Nova-3 rates. |
| Framer | Aug 2026 | New app-layer adopter: Framer introduced fixed monthly Agents AI-credit pools — 500 / 1,000 / 3,000 credits by tier — to meter its agentic site-generation features, layered on a site-plan ladder (Free / Basic $10/mo yearly / Pro $30/mo yearly / Enterprise) and per-seat Workspace plans that were not repriced. It dropped the intermediate Scale plan in the same restructure. A website builder giving AI its own credit pool rather than folding agent generation into the site plan, following the Oct-Nov 2025 simplification that removed the Mini tier. |
| Mintlify | Aug 2026 | The AI meter carved out of the free tier entirely, then priced. Mintlify reversed its free-Starter-plus-Enterprise structure into three tiers: free Starter now ships only the base documentation platform (custom domain, web editor, MCP server, API playground, 5 editor seats) with the AI assistant, writing agent, agent skills and automations REMOVED and its previous 5,000-credit allotment gone from the feature table; a reintroduced Pro tier at $450/mo billed annually or $540/mo monthly carries those AI features with 10,000 AI credits/month at $0.01/credit overage; Enterprise stays custom-quoted with committed-volume credit discounts. The credit meter and its $0.01 peg survived intact — what moved is which plan the meter attaches to. Reconfirmed on the 2026-08-11 live capture. |
Counterexamples
- Wispr Flow · Aug 2026 — The bundling force at full strength: a whole second product folded into the seat at no price. Wispr shipped Wispr Notetaker — meeting transcription, speaker ID, cross-meeting Q&A, calendar and Slack connections, MCP export into Claude and ChatGPT — as a new top-level nav item and help-center collection, included in Free and Pro (Mac only) with Enterprise marked 'coming soon'. No new SKU, no seat-price change, no separate meter. The site nav moved from 'Dictation | Business | Pricing' to 'Dictation | Notetaker | Business | Pricing' and the footer now lists two products. Same shape as Anthropic's 2026-07-23 Cowork/Design/Science bundling, one tier down the market.
- Diffbot · Aug 2026 — A fifth product added to an existing meter rather than given its own. Diffbot's Web Search API (1 credit per query, the same rate as a page extraction) was bundled into the existing 'all APIs included' entitlement alongside Extract, Natural Language and Knowledge Graph on every plan including Free, with Crawl remaining the only feature gated to the $899/mo Plus tier and above. Headline prices ($0/$299/$899/custom), credit allotments and per-credit overage rates were all unchanged. FLORA did the same on the same date, bundling a new Fashion Studio into every plan at no incremental seat price with generation metered from each plan's existing usage pool. One currency, more products, no new line — the exact inverse of carving AI onto its own meter.
- Anthropic · Jul 2026 — The strongest bundling counter-current in the corpus, from its largest vendor: Claude Pro ($20/mo), Max (from $100/mo) and Team ($20-$100/seat/mo) were expanded to bundle Claude Cowork, Claude Design and Claude Science alongside Claude Code — three additional products folded into the subscription at UNCHANGED prices, where the page had previously foregrounded Claude Code alone. Headline subscription prices and the Opus/Haiku API rates did not move. The Enterprise tier added a 500k-token context window on the default model plus a self-serve/sales-assisted split and a university-wide Education plan. Where Linear and Jasper carved AI out of the seat, Anthropic pushed three more products into it.
- Sequence · Jul 2026 — Shipped an entire agent layer with no meter and no new SKU: named agents for AR, reconciliation, invoice review, quote approvals, payment reminders and contract intake — each configured in plain English with explicit tools, app scopes and guardrails, reviewed in a new Watchtower console — plus Sequence MCP for Claude/Cursor/ChatGPT. None of it is separately priced. Growth remains $799/month for startups under $1m annual revenue, Core and Scale remain bespoke platform fees on projected billed revenue, and the paid add-ons are still the quote builder, custom reporting dashboard and revenue recognition. AI folded entirely into an unchanged platform fee.
- Shortwave · Jun 2026 — Went the opposite way: raised every paid seat ~20-29% but still sells AI capacity ONLY by moving up a tier — no separate AI meter, no metered overage. The daily AI quota, search depth, and token multipliers are bundled into the seat price, so AI stays folded into the base plan.
- GitHub Copilot · Jun 2026 — Bundles most AI into the per-seat price (Business $19, Enterprise $39) with premium-request overage layered on top rather than a wholly decoupled AI currency — the dominant SaaS pattern of AI-inside-the-seat that this trend runs against.
- Mistral AI · Jul 2026 — Kept AI usage and consumer subscription cleanly on two surfaces but did NOT introduce a decoupled-from-edition AI currency — it simply raised developer-API token rates (Small 4 to $0.15/$0.6, OCR 4 to $4/1K pages) while holding the Vibe consumer ladder flat. A separate API vs consumer split is not the same as an edition-independent AI meter.
Trivia
-
Snowflake's 2026-07-06 AI Credit split has a precise, counterintuitive consequence: before the change, running the identical Cortex AI query cost exactly 2x more on a Business Critical account ($4/credit) than on Standard ($2/credit) purely because of the edition rate. After the split, the AI half of the bill is flat ($2.00/AI Credit) on every edition — but the warehouse compute that executes each AI query still varies by edition, so Snowflake decoupled only half the bill. It is the corpus's cleanest example of a vendor un-bundling AI from its own tiering structure.
-
Vercel (2026-07-06) priced its agent against the grain of its category: where the corpus's other coding-agents reach for per-seat or per-action pricing (which quietly re-introduces inference margin), Vercel Agent bills $0.25 per 1M tokens plus pass-through — a thin metering fee on the same zero-markup rails as v0 and AI Gateway. It refused to sweeten the $20 seat with 'free' agentic Code Review, giving the agent its own token-metered line so its inference cost is never cross-subsidised by the platform fee.
-
All three same-day adopters route the AI meter differently — Snowflake by cross-region routing ($2.00 global vs $2.20 regional), Anthropic by marketplace unit (CCU at $0.01, 100 CCU = $1.00), Vercel by a flat per-token platform fee — but converge on the same structural move: AI stops inheriting the base plan's price and gets a currency of its own.
-
The signal graduated from emerging to sharpens in exactly one capture cycle: logged 2026-07-07 with three same-day (2026-07-06) adopters, it added two more on 2026-07-14 — Weights & Biases' ARIA 'token based pricing' line and Google's AlphaEvolve agent SKU on Vertex AI — an eight-day round-trip from directional cluster to five-vendor, two-cycle structural move, one of the fastest emerging→sharpens graduations in the corpus.
-
Google's AlphaEvolve (2026-07-14) is the corpus's first case of the separate-AI-meter expressed as a MULTIPLIER rather than a flat fee: the agent is billed as the base Gemini token rate plus a 2× agent surcharge (3× all-in), so unlike Snowflake's flat $2.00/AI-Credit or Vercel's $0.25/1M, the decoupled AI line here scales with the underlying model price instead of floating free of it.
-
This trend wrote its own kill condition and then failed to be killed. The 2026-07-15 file said the pattern would be "a data-warehouse / dev-platform quirk" unless a third and fourth cycle added app-layer or seat-based vendors. Within a fortnight three arrived: Linear moved a second AI feature behind purchased credits at unchanged $10/$16 seats (2026-07-29), Jasper added a credit meter to a previously meter-free per-seat product (2026-07-22), and Freshworks published a per-unit AI rate on a seat-based CRM (2026-07-21).
-
Linear (2026-07-29) shows the seat being unbundled one feature at a time rather than all at once. "Linear Agent automations (beta)" was included free in the $16 Business seat; renamed "Loops", it now carries the same "Requires AI credits" footnote as Coding Sessions — making it the second metered AI capability — while Code Intelligence and Triage Intelligence remain bundled at no extra charge. The same pricing page therefore now contains both halves of the argument: two AI features on their own meter and two folded into the seat.
-
Jasper (2026-07-22) reintroduced metering three years after abolishing it. The company killed the word caps it launched with, ran a pure per-seat subscription with no meter of any kind — no tokens, no generations, no credits — and has now added a credit meter to its Business plan with a public rate card: 10 credits per Grid row, 10 per GEO Hub query or page run, 40 per Research or Translation Agent run, 100 per Optimization Agent run, 1-40 per API/MCP call. Standard Chat and Canvas content creation is explicitly excluded, so only the agentic and GEO surfaces are metered.
-
Modal (2026-07-29) made its first departure from per-second billing since founding: a new OpenAI-compatible Shared API metered by TOKEN, launched with Moonshot's Kimi K3, running alongside — not instead of — the existing GPU/CPU/memory-second rate card, with Starter's $30/month free credit applying to both. A pure per-second compute vendor now gives customers a choice of meter for the same workload, which is the decoupling logic applied to the unit rather than to the tier.
-
Anthropic is the counter-current at the largest scale in the corpus, and it is winning on its own surface: on 2026-07-23 it expanded Claude Pro ($20/mo), Max (from $100/mo) and Team ($20-$100/seat/mo) to bundle Claude Cowork, Claude Design and Claude Science alongside Claude Code — three additional products folded into unchanged subscription prices. Where Linear and Jasper carved AI out of the seat, Anthropic pushed three more products into it.
For buyers
If a platform you use puts AI on a separate meter, treat AI spend as a distinct forecasting and audit line — it will move on its own cadence (a model-price cut or a routing change can drop it without your base plan changing). Read the routing knob: Snowflake's AI Credit is $2.00 for global cross-region routing but $2.20 when you lock to a region for data residency, so a compliance requirement is a ~10% AI-cost premium. Check which discounts apply — Snowflake's capacity discounts explicitly do NOT apply to AI Credits (only automatic ACV-based ones do), so pre-paid commits won't cut the AI half. And confirm the markup story: Vercel's Agent and v0 pass model cost through at a thin fee, so a model-price drop reaches you automatically, whereas a seat that bundles AI captures that drop as vendor margin.
For vendors
Running this play needs a metering layer that can rate AI usage on a different axis than the base product — by cross-region routing, by model, or by a marketplace consumption unit — and reconcile both onto one invoice. The reusable move is to make the AI price edition/plan-independent so a cheaper model or route flows straight to the buyer without re-rating the account, protecting the margin on your base tiers from inference volatility. The tradeoff is ARPU: bundling AI into the seat captures inference upside inside the subscription and is simpler to sell (the Shortwave / GitHub Copilot path), while a separate meter trades that capture for cleaner unit economics and a FinOps-friendly bill. The tell that you need the split is when AI cost-of-goods becomes large enough that a bundled tier's margin swings with usage.
Outlook — what to watch
Logged as emerging in July 2026 on three same-day adopters at the platform layer — a directional cluster, not yet a category standard. It graduates to holds if the next cycles add app-layer or seat-based vendors decoupling AI onto its own currency (not just BYOK passthrough, which zeroes a meter rather than creating one). It stays emerging, or fades, if the platform-layer moves remain a data-warehouse / dev-platform quirk while the mass of SaaS keeps folding AI into the seat. The boundary to watch is the BYOK trend: that lever lets a buyer zero the passthrough meter; this trend is about the vendor giving AI its own currency regardless of who supplies the key.
Bottom line
Three platform vendors — Snowflake, Vercel, Anthropic — pulled AI usage onto a dedicated meter on 2026-07-06, pricing it as a thin fee over pass-through on its own axis (routing, model, marketplace) rather than inheriting the base seat or edition. The driver is unit-economics honesty as inference becomes a large, volatile cost line; the countervailing force is the still-dominant instinct to fold AI into the seat.
FAQ
What does it mean to put AI usage on a separate meter?
It means AI features bill in their own currency — a dedicated AI credit, a per-token platform fee, or a marketplace consumption unit — priced independently of the base plan's seat or edition. Snowflake's Cortex AI Credit ($2.00 global / $2.20 regional) is flat across every Snowflake edition, so identical AI usage costs the same on Business Critical as on Standard, while the warehouse compute underneath still bills at the edition rate.
Why would a vendor decouple AI from the base plan price?
Because AI inference is the largest and most volatile cost-of-goods line. If AI is folded into a seat or edition, a model-price cut can't reach the buyer without re-rating the whole plan, and a usage spike crushes the margin on a bundled tier. A dedicated AI meter lets the price float on its own axis — routing, model, or marketplace — so savings pass through automatically and AI spend becomes separately forecastable.
How is this different from bring-your-own-key (BYOK) pricing?
BYOK lets a buyer zero the model-cost passthrough meter by supplying their own API key. This trend is about the vendor giving AI its own billing currency regardless of who supplies the key — a structural pricing choice, not a discount lever. A vendor can run a separate AI meter and offer BYOK on top of it.
Does a separate AI meter make AI cheaper?
Not by itself, but it changes how price changes reach you. On a thin-fee, pass-through meter (Vercel Agent at $0.25/1M plus model cost), a model-price drop flows straight through — Vercel's v0 Max Fast fell ~3x and buyers got it automatically. A seat that bundles AI can capture that same drop as vendor margin instead. Watch the routing and discount rules too: Snowflake's data-residency routing costs ~10% more and capacity discounts don't apply to AI Credits.