Forbes reported in July 2026 that Telnyx was paying roughly $100,000 a day running one AI agent product on Anthropic's models, and closer to $100 a day running the same product on a cheaper alternative. Same feature, same customer-facing price, a thousand-fold swing in what it cost to deliver depending on which model answered the request. That gap is not a rounding error in someone's infrastructure bill. For any company that sells an AI product at a fixed price, per seat, per conversation, or per resolved case, it is the entire difference between a healthy margin and a business that loses money on its best month.
Pricing has mostly caught up with this. Most AI-native vendors don't charge per token anymore, because customers hate a bill that moves when their usage does. They charge per seat, per task, or per outcome, a flat, predictable number the buyer can put in a budget line. Intercom's Fin, one of the earliest and clearest examples, charges $0.99 per resolved support ticket and nothing when the agent escalates to a human. That's the right instinct for the buyer. It just quietly moves 100% of the model-cost volatility onto the seller, and most sellers are still pricing the model side like the token was still $60 per million.
What the four pricing models actually hide
| Model | What the customer sees | Where the margin risk lives |
|---|---|---|
| Per-seat | Flat monthly fee per user | Power users run 10x the queries of light users at the same price; margin depends on usage mix, not headcount |
| Per-token / credits | A bill that tracks API usage | Technically safe for the seller, but it's the bill customers churn over, and it exposes every architectural inefficiency (retries, replayed context) as a line item |
| Per-task | A fixed rate per unit of work (a document processed, a claim reviewed) | Predictable only if every task costs roughly the same to complete, which stops being true the moment task difficulty varies |
| Per-outcome | A fixed rate per completed result (a resolved ticket, a funded loan) | Maximum margin upside when it works, because cost per resolution can run 60-90% below the price charged, but the whole model assumes a stable cost-to-serve that most teams never actually measure |
Per-outcome pricing is winning the market for a good reason: it's the model buyers say they prefer, and industry pricing research this year put buyer preference for usage- or outcome-based models at roughly 86%. It's also the model with the best margin ceiling, since outcome-priced AI products can run gross margins in the 90%+ range compared to razor-thin or negative margins under naive per-token resale. But "best margin ceiling" and "reliable margin" are different claims. A ceiling only holds if the cost of producing one outcome stays roughly where it was when the price was set. Nothing about a model routing decision guarantees that.
Why the cost side moves even when the price doesn't
Three things are happening at once in 2026, and none of them are under the pricing team's control.
Token prices keep falling, unevenly. List prices for frontier tokens have dropped roughly tenfold a year since 2021, which sounds like it should only help margins. But it's a moving floor, not a fixed discount: a routing or model-selection decision made in Q1 against that quarter's price sheet is stale by Q3, and a product still hard-coded to one model isn't capturing the drop at all.
Usage volume is compounding, not linear. Google reported monthly token consumption growing roughly 330-fold between 2023 and 2026, and that growth is concentrated in exactly the agentic, multi-step workloads that outcome-based products are built on. A support agent that used to answer in one pass now plans, calls tools, and re-reads its own context before it resolves anything, and each of those extra steps bills like a full request.
Agent architecture, not model choice, is often the bigger lever. A mid-2026 cost study found that agents which keep growing their working context replay up to 3.6x the input tokens of a single compiled pass to reach the same verdict, purely from re-sending accumulated context on every step. That's a cost multiplier that has nothing to do with which model answered and everything to do with how the agent was built, and it stacks with whatever the model itself charges.
Put those three together and the cost of producing one "resolution" isn't a number you set once at launch. It's a number that moves weekly, and a pricing model that assumes it's flat is running on borrowed margin.
The tempting fix, and why it's the wrong one
The obvious response to a margin squeeze is to swap the whole product to the cheapest model that still mostly works, the Telnyx move. It's real money: Anthropic's own reported inference margins climbed from roughly 38% to over 70% during 2026 as serving got more efficient, which means the model providers themselves are proof that the same workload can be served at wildly different cost depending on execution, not just sticker price.
But a wholesale swap to one cheap model is a blunt instrument. It fixes the easy 80% of requests and silently degrades the hard 20%, the multi-step, ambiguous, or high-stakes cases that were the reason the customer was willing to pay $0.99 a resolution instead of routing the ticket to a human in the first place. Outcome-based pricing only works if the outcome is actually achieved; a wrong or incomplete resolution that gets reopened costs you the resolution fee and the redo. Model choice made once, product-wide, at the account level, can't tell the difference between a request that needs the expensive model and one that doesn't.
The fix that actually holds margin steady is making that choice per request instead of per product: send each ticket, task, or query to the cheapest model that can actually resolve it, verify the answer before it counts as a resolution, and escalate only the fraction that genuinely needs the stronger model. That's a cascade with a quality gate, not a model swap, and it's the difference between capturing the token-price drop everywhere it's safe to and eating a quality regression everywhere it isn't.
What this means if you're pricing an AI product this quarter
- Measure cost per outcome before you set the price, and re-measure it monthly. If the last time anyone checked cost-per-resolution was at launch, the number pricing was built on is already wrong.
- Don't pass token costs through 1:1, and don't ignore them either. The workable range most 2026 AI-native pricing lands in is 50-70% gross margin absorbed at the product layer, not the old 80% SaaS assumption and not a straight token resale.
- Separate "which model answered" from "did the product work." If your only lever for cost is picking one model for the whole product, you're trading quality for margin every time the market shifts instead of only on the requests where it's safe to.
- Ask what a 3x volume spike does to your cost per outcome, not just your revenue. Per-outcome pricing looks great at steady state and can break exactly when a customer's usage compounds, the moment you most need the margin to hold.
- Route, then verify, don't just swap. A verifier that checks the cheap model's answer before it ships is what makes it safe to route the easy majority of requests to a cheaper model without quietly lowering the resolution rate you're being paid on.
Nadir sits in exactly that gap: a router that sends each request to the cheapest model that clears a quality floor you set, with an optional verifier cascade for complete non-streaming responses. When a benchmark is configured and priced, the response includes the nadir_metadata.benchmark_comparison.savings_usd field. Start free and compare that routing signal with your own outcome data.
Sources: ["As Token Costs Plunge, Enterprise AI Providers Face A New Margin Squeeze," Forbes, July 28, 2026](https://www.forbes.com/sites/petercohan/2026/07/28/as-token-costs-plunge-enterprise-ai-providers-face-a-new-margin-squeeze/). [Nevermined, "Cost-Based Pricing in AI Agents Statistics"](https://nevermined.ai/blog/based-pricing-ai-agents-statistics). [MightyBot, "AI Agent Pricing Models Compared: Per-Seat, Per-Token, Per-Task, Per-Outcome"](https://mightybot.ai/blog/ai-agent-pricing-models-compared/). [MightyBot, "Token Economics of AI Agents in 2026"](https://mightybot.ai/blog/token-economics-of-ai-agents-2026/). [Causo Hub, "How to price an AI product in 2026: the margin playbook"](https://hub.causo.ai/guides/how-to-price-ai-product-token-costs-margins-2026). [Solvimon, "What is AI Agent Pricing?"](https://www.solvimon.com/glossary/ai-agent-pricing). Figures current as of August 2026 and drawn from third-party reporting on vendor pricing and margins; individual vendor economics vary.