What happened
Agno’s October 23, 2026 knowledge-base guide, “How to price an AI feature in your SaaS product,” is attributed to Cosette Cressler, Kaustubh Shukla, and Harsh Sinha. It recommends four pricing approaches: bundle a feature into seats, charge for a customer-visible usage metric such as invoices sent or tickets resolved, sell a separate AI tier, or charge by the end users reached. The guide argues against billing customers per token or per trace because those units are hard to forecast and may misalign incentives. It says teams should still record tokens and cost internally.
The guide illustrates the choices with a support agent serving a customer with 10,000 end users. Assuming 20% contact support monthly, six turns per conversation, and 4,000 input and 300 output tokens per turn, it estimates model cost at about $0.07 per conversation on GPT-6 Sol, then about $0.15 all-in after retrieval, guardrails, and occasional retries. At a 70% resolution rate, billing $0.99 per outcome yields $1,386 monthly revenue and a stated 78% gross margin. Charging $0.25 per reached end user yields $500 and a 40% margin. The guide notes that a 20-turn conversation doubles input cost, while switching to GPT-6 Astra raises the estimated all-in conversation cost to about $0.75.
Its second example is an in-product copilot for a 500-seat customer paying $50 per seat. At 30 daily requests per active seat, each with 6,000 input and 400 output tokens, the guide estimates monthly cost per seat at $10.56 on GPT-6 Sol and $0.53 on GPT-6 Luna. A mix routing 80% of requests to Luna and 20% to Sol costs about $2.53 per seat monthly. A $20 AI tier would have a stated 87% margin on that mix and 47% on Sol alone. But if 5% of seats make 150 requests daily, the guide estimates their cost at $53 each on Sol.
For cost control and billing, Agno recommends recording each run’s tokens, cost, user, session, and tenant, then aggregating by tenant per billing cycle. It describes allowances and overages, model routing, session summarization, and alerts as ways to manage exposure. For anonymous public agents, it suggests rate limits, caching, a cheaper default model, and a hard budget cap. Agno says its AgentOS records per-run metrics and supports tenant-based queries. The guide also lists Agno platform prices as of September 22, 2026: Free at $0, Pro at $150 per month, and Enterprise at custom pricing; model costs are paid directly to providers.
What the archive adds
New Runtime’s earlier coverage examined how SaaS companies structure AI pricing and account for inference and human-in-the-loop costs, including a note dated 2026-03-20. This guide adds worked examples and specific cost arithmetic to a known topic, rather than reporting a new market change.
Trend signal
This guide is consistent with agent economics moving to completed work: it argues that buyers should pay against outcomes or other observable business units, not model calls. The examples are vendor-authored estimates, however, and do not establish that the wider market has adopted outcome-based pricing.
Why now
Agno frames the issue as a consequence of AI features’ variable cost of goods: provider price lists, model verbosity, customer usage, and routing choices all affect margin. Its examples use OpenAI list prices from September 22, 2026, and it advises budgeting for a 2× cost swing within a year.
Why it matters
A SaaS team can choose a price that looks healthy at average usage and still lose money on heavy users, longer conversations, or a more expensive model. Agno’s examples make the distribution problem tangible: in the copilot case, routing routine requests to a cheaper model reduces estimated cost from $10.56 to $2.53 per seat monthly, while a small group of high-volume users can cost far more than the average. Those figures are assumptions in a worked example, not a universal benchmark, but they identify the variables worth measuring.
For builders, the operational implication is to connect product analytics, billing, and runtime controls. Track cost by tenant and user before deciding whether to bundle, charge on outcomes, or create a tier. Then define allowances, overages, and alerts before launch. A customer-visible unit makes a bill easier to forecast; internal metering and routing give the team room to preserve margin without making customers reason about tokens.
Limits and uncertainty
The cost calculations are illustrative, and actual results depend on prompt and context size, traffic, retries, retrieval, guardrails, and vendor rates. Agno cites OpenAI prices as of September 22, 2026, and explicitly notes that prices change. Its recommendation to assume a 2× annual cost swing is guidance, not a measured forecast. The guide’s worked examples are not independent evaluations, and the stated gross margins depend on its chosen assumptions and pricing. No evidence here establishes which pricing model performs best across SaaS products.
Reader takeaway
Treat token and trace counts as internal cost signals, not automatically as the customer’s billable unit. Measure cost and usage by tenant, test the distribution rather than just the average, and choose a customer-visible unit with allowances and overages that can absorb heavy use and model-price changes.
Open question
How often do SaaS teams revise customer-facing AI pricing when usage distributions or model costs diverge from launch assumptions?