Per-Seat Vertical AI Bundles vs Per-Token Pricing: A 10-Person Worked Comparison

Two pricing shapes now compete for the same client budget, and they are not interchangeable. On September 10, 2026 OpenAI introduced ChatGPT for Financial Services, which bundles licensed financial data (Daloopa, PitchBook, LSEG News, Crunchbase) into a per-seat ChatGPT Work experience on GPT-6 Astra. That is a different unit of sale from the tokens-and-markup model most AI agencies quote with. This page works the comparison for a 10-person client team, shows the crossover point, and states the one caveat that changes agency margin maths: premium-data seats are not resellable through an agency.

Per-seat vs per-token: the two shapes side by side

The practical consequence: the per-token calculator answers “what does it cost to run this workload?” It cannot answer “what does it cost to buy this capability?” That second question is now the per-seat one.

Does OpenAI publish a price for ChatGPT for Financial Services?

No. The September 10, 2026 announcement carries no list price, no seat minimum, no contract length and no stated eligibility test. Availability is described as being for “eligible financial institutions,” with a “Contact financial services sales” call to action. American Banker reported the same day that OpenAI did not immediately respond when asked what determines eligibility. Treat every Finserv figure you see in the wild as modelled, not published — including the ones on this page.

What is published, per vendor sources, is the surrounding price structure:

Two clarifications worth keeping straight, because aggregators blur them: the enterprise ChatGPT for Financial Services product is not the consumer personal-finance experience in ChatGPT (launched May 15, 2026 for Pro users in the US); and the “Financial Services Terms” regime covers OpenAI's financial plugins and Finserv together, so partner-data restrictions should be cited as the Financial Services regime rather than as Finserv-only.

For context on the vertical race, dated: Anthropic has been in this vertical since July 15, 2025 with Claude for Financial Services and a partner-connector model, and a $1.5 billion joint venture with Blackstone, Hellman & Friedman and Goldman Sachs was reported on May 4, 2026 (a figure Anthropic has not confirmed), followed by a May 5, 2026 briefing with financial-services agent templates and Microsoft 365 add-ins. OpenAI's September 2026 entry is the first time licensed premium data is indexed and hosted by the model vendor itself.

Step 1: the per-token estimate our calculator produces today

Our calculator prices a workload from three inputs: the model's list rates, the monthly input-token volume, and the monthly output-token volume. The formula is:

metered $/month = (input tokens ÷ 1,000,000 × $10) + (output tokens ÷ 1,000,000 × $50)

To keep the comparison concrete I need a usage unit. Modelled input (ours, not a vendor figure): one model-assisted research task consumes 40,000 input tokens and 4,000 output tokens — a 10:1 input-to-output ratio, which is what document-heavy research and drafting looks like. At Astra list rates, that is:

$0.40 input + $0.20 output = $0.60 per task (40,000 ÷ 1,000,000 × $10 = $0.40; 4,000 ÷ 1,000,000 × $50 = $0.20)

Four usage intensities, each applied to 10 users:

Check the arithmetic on any row: tasks × $0.60 × 10 users × 12 months. Medium = 150 × $0.60 = $90 per user per month, × 10 users = $900 per month, × 12 = $10,800. Inputs: 40k/4k tokens per task (modelled), $10/$50 per 1M tokens (published), 10 users (assumed).

Step 2: the per-seat vertical bundle for the same team

The seat shape is simpler: annual bundle = seats × seat rate × 12. The only input we do not have is the Finserv seat rate, so we hold it as a variable instead of a number: seat_finserv = k × the published standard ChatGPT Business seat ($20/user/month annual), with k shown at 1, 2, 3 and the premium-seat precedent of 5.

Read the $20 and $100/$125 rows as published prices for other OpenAI seats, and the $40 and $60 rows as our interpolation. None of them is a ChatGPT for Financial Services price — OpenAI has not published one. What the $100 precedent buys you is a scale check: OpenAI's own premium tier sits 5× above its standard tier, so a vertical bundle with licensed data included should be expected to price well above a standard seat, not at it.

One more input the seat quote normally omits: the client's existing data subscriptions do not go away. OpenAI says the built-in datasets arrive with “no separate contracts to negotiate or connectors to set up,” and it is working with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's on shared sign-in and entitlement integrations. So total cost of ownership = Finserv seats + the subscriptions the firm still pays for.

The crossover: where metered stops being cheaper

Set the two shapes equal. Working in per-user monthly terms, with metered cost per user = $15 per million input tokens equivalent (10:1 input-to-output at $10/$50 per million), the crossover is:

crossover usage per user = seat rate ÷ $15 per million input-equivalent tokens
or, in tasks at $0.60 each: crossover tasks per user = seat rate ÷ 0.60

Three things fall out of that table:

• Team size cancels out. Both shapes scale linearly per user, so the crossover is a per-user usage number, not a head-count number. A 3-person team and a 30-person team cross at the same intensity — this is the opposite of what most “when does per-seat win” intuition assumes.
• Team size re-enters only through a floor. A seat minimum or a fixed platform fee is the only thing that makes head-count matter. Example: a 3-person team forced onto a 10-seat minimum at $100/user/month pays $12,000/year for three active users — an effective $4,000 per active user per year, versus $3,240 for the same three users at medium intensity (150 tasks each) on metered inference. Minimums, not seats, are what turn the bundle into a bad deal for small teams.
• The metered number is the honest ceiling to compare against. $12,000/year is what 10 premium-precedent seats cost — and also what 240 million output tokens (20M/month team-wide, 2M per user) or 1.2 billion input tokens of Astra inference cost. If the client's real usage sits under that, metered plus a services fee is the better-argued quote.

Model the crossover for your own team

The caveat that changes agency margin maths: these seats are not resellable

If you have been pricing AI by buying tokens and reselling them at a 3–10× markup, the per-seat vertical bundle breaks that mechanism. Premium-data seats are not resellable through an agency. The restrictions are in OpenAI's Financial Services Terms, not in a blog post:

• Partner data is internal-use. PitchBook data is licensed “for your internal business operations, including internal research and analysis,” with prohibitions on model training, grounding, reconstitution, bulk export, CRM loading and competing products.
• Output excludes Partner Data. The terms say ownership of Output does not include Partner Data, and that “an export or sharing feature does not expand the rights granted under those terms.”
• Reuters News is person-restricted. It is available only to professional users employed by eligible financial firms domiciled or incorporated in the United States — an agency's staff are not those users.

So the client has to hold the entitlement in its own name, and the agency cannot stand between the vendor and the client as a reseller. A deliverable built on Partner Data cannot be redistributed by the agency either. What that does to the revenue mix:

Modelled illustration (ours, not a published price): on a $12,000/year seat contract, a reseller model would have booked roughly $2,400 at a 20% margin — and on a non-resellable seat that revenue is $0. The same engagement has to be repriced as a services engagement against the licence value. If your standard services attach is 20–30% of licence value, that is a $2,400–$3,600/year services line you have to scope, sell and defend on its own merits — implementation, template and style-guide configuration, eval and evidence-review workflows, governance and enablement. The honest framing for the client is that you are not the vendor's channel here; you are the integrator, and the metered API remains the only layer you can buy, mark up and resell.

Frequently Asked Questions

Bottom line for 2026

Quote the shape, then the number. Per-token metered inference stays the resellable layer and prices cleanly at $10/$50 per million tokens on Astra; per-seat vertical bundles with licensed premium data are a capability purchase the client makes directly, at a rate OpenAI has not published. For a 10-person team the two shapes cross at roughly 167 tasks per user per month at the $100 premium-seat precedent — below that, run metered inference and sell services on top; above it, the bundle is the cheaper capability and your revenue is the implementation, workflow, evidence-review and governance work around it. The mistake to avoid is quoting a markup on seats you cannot buy.