A product manager at a 60-person SaaS company told her team last month that switching three internal workflows to GPT-5.4 Nano would cost roughly $40 a month. She was right. She was also building on a floor that nobody actually knows how to price at cost.
The average cost per million tokens across major providers fell from roughly $10 to $2.50 in a single year, according to Ramp's enterprise spending data. That looks like progress - and it is - but the reason it's possible tells you something worth understanding before you restructure your software stack around it.
The inference subsidy is real, and it is large
In 2025, OpenAI generated $3.7 billion in revenue and lost an estimated $5 billion - spending $1.35 for every dollar it earned - driven not by R&D but by the cost of serving inference requests. Anthropic's situation is similar: Anthropic's gross margins came in at 40% in 2025, ten percentage points below their own projections, with inference costs running 23% higher than anticipated.
OpenAI, Google, Anthropic, and Meta are all pricing inference below cost to capture market share. When frontier model providers are subsidizing your API calls, it creates a false floor in the market - one that will eventually normalize upward when capital discipline returns.
Both companies have told investors they expect margins to improve dramatically: Anthropic has projected gross margins above 70% by 2027, and OpenAI has projected gross margins of at least 70% by 2029 - closer to the margins of publicly traded software and cloud firms. Getting there means the price you are paying today is not the price they need to charge tomorrow.
There is a parallel from recent history worth keeping in mind. Silicon Valley has seen this before: VC money helped underwrite cheap Uber rides and DoorDash deliveries; Amazon built its base with low prices, free shipping, and for years no sales tax in most states. Eventually, all of these companies had to charge enough to cover costs - and make a profit.
Why this matters more now that agents are eating the per-seat model
The timing is inconvenient. Just as cheap inference is making AI agents feel economically obvious compared to SaaS subscriptions, the economics underneath those agents are propped up by losses.
Per-seat pricing is structurally broken for AI agents: the better the agent works, the fewer seats a buyer needs, so the vendor is paid to under-deliver. That logic is sound. The SaaS business model was built on a simple equation: more employees using software meant more revenue. Per-seat pricing worked beautifully when every new hire needed their own login, dashboard, and license. When AI makes one person as productive as five, the company does not need five seats.
In February 2026, approximately $285 billion in market value vanished from software stocks in a single trading session - ServiceNow dropped 7%, Salesforce fell 7%, Intuit plummeted 11%. Investors were pricing in the structural pressure on per-seat revenue. They were not wrong about the direction.
But "agents are cheaper than SaaS" is partly a comparison between the real cost of SaaS and the subsidized cost of inference. Agentic workloads consume 5-30x more tokens per task than a simple chatbot does. Run those agent workflows at the cost the labs actually need to charge, and the economics look different - not necessarily bad, but materially different from the spreadsheet you ran last quarter.
The steelman for the other side is real: NVIDIA's Rubin platform targets a 10x reduction in inference token costs compared to its Blackwell architecture. Hardware efficiency gains are genuine and compound. If compute falls fast enough, the labs could reach margin targets without raising prices. That is possible. Betting your stack architecture on it is a different thing.
What to actually do about it
None of this means do not build on AI APIs. It means build in a way that survives a price adjustment.
A few concrete moves:
- Use caching aggressively now. Both OpenAI and Anthropic now offer ~90% off cached input, making effective costs nearly equal for cache-heavy workloads. Teams that engineer for high cache hit rates are paying far less per task and will be more insulated from future price changes.
- Separate high-volume routine calls from frontier model calls. Open-weight model APIs - Llama 4, DeepSeek V4 - have pushed the price floor to $0.07-$0.14 per million input tokens. Routing classification, extraction, and formatting tasks to cheaper models insulates you from frontier price moves.
- Model your agent costs at 2-3x current token prices. Enterprises that planned budgets around 2024 token rates are finding that agentic AI workflows at 2026 adoption levels consume multiples of what the spreadsheet projected. The same gap will reappear if subsidy normalization and volume growth compound together.
- Watch how the labs price agentic workloads specifically. Tool-enabled requests already add 300-700 extra input tokens per call on both OpenAI and Anthropic, billed at full rates. That overhead is not in most cost estimates.
Beagle in action#engineering, 10:45amThe outcome-pricing shift compounds the uncertainty
There is a second variable teams tend to ignore: the move from per-seat to per-outcome pricing does not just affect SaaS vendors - it affects how you will be sold AI products going forward.
Outcome-based agent pricing typically runs $0.50 to $2.00 per resolution with no charge on escalations, and above roughly 3,000 monthly conversations it usually beats per-seat once you fold in implementation, helpdesk fees, and agent salaries. That math works - but only if the inference cost inside the vendor's product stays where it is today.
The contract risk is the word 'resolution': define it precisely or the vendor will. Add inference price normalization to that ambiguity, and outcome-based contracts signed this year could look very different in eighteen months.
Outcome-based pricing is not a trap. It is just a contract that deserves the same scrutiny as any arrangement where one side controls the cost of goods.
AI inference pricing for enterprise teams: common questions
Why are AI inference costs falling so fast?
Google builds its own Tensor Processing Units, reducing dependence on third-party GPU pricing, and its developers were processing roughly half a trillion tokens per day inside its internal platform by March 2026. That internal scale creates a data flywheel that improves model efficiency and reduces per-token serving costs over time. Hardware improvements and competition are both real - but so is below-cost pricing.
Are the major AI labs actually losing money on inference?
Yes. The labs are not profitable on inference. Anthropic's inference costs ran 23% higher than anticipated in 2025. OpenAI's gross margins fell from 40% in 2024 to 33% in 2025, missing their own 46% forecast, as inference costs grew fourfold year-over-year to roughly $8.4 billion.
Will per-seat SaaS pricing actually disappear?
Not entirely, but it is losing ground fast. Seat-based AI pricing fell from 21% to 15% of companies in a single year, and seat-only vendors now risk immediate disqualification from deals.
The hybrid 'seat plus usage' model is the 2026 equilibrium, but it is not a stable final state.
How much more do agentic workloads cost than chatbot workflows?
Significantly more. "Cheaper per token" does not mean cheaper overall when agentic workloads consume 5-30x more tokens per task than a chatbot. Most teams underestimate this because they benchmark against simple chat interactions, not multi-step agent loops with tool calls.
What is the safest way to build on AI APIs given the pricing uncertainty?
Prioritize prompt caching, route routine tasks to open-weight models, and stress-test your cost model at 2-3x current token rates before committing to an architecture. Managed model APIs remain the easiest default, but open-weight deployments are increasingly credible for stable, repetitive workloads
- and they are not subject to a single vendor's margin recovery timeline.