second moment

Predictable costs for AI agents

Companies sell AI work at fixed prices and pay wildly variable model costs. Second Moment makes your cost a fixed price per unit of agent work, priced on your runtime data, and carries the overages for you.

The problem:

Per-unit pricing is spreading across verticals: $0.99/resolution (Intercom Fin), ~$1/qualified lead (HubSpot Breeze), ~$300/demand package in legal AI, $2.25 per unit of agent work at Devin. Meanwhile, the worst units routinely cost 2–16× the average, and this year frontier providers repriced upward for the first time. 73% of organizations report AI costs exceeding projections (FinOps Foundation, n≈1,200). Every vendor with a published per-outcome price is exposed to that swing with no way to manage and forecast that risk accurately. Second Moment takes the other side, so next quarter's AI cost is a number you can plan around.

Second Moment's Offering:

The free quote:Export your usage traces and get back a one-page analysis detailing what your units actually cost, the fixed price we'd quote, and that contract backtested against your own history, including the weeks we'd lose. No integration would be required yet. We can also run a canary: a slice of your traffic through the gateway at pass-through, quoted in real time.
The cap (where most contracts would start):Second Moment passes through costs up to an agreed threshold (plus a small fee) and absorbs anything above. Your losing units are ours to handle. This requires no major billing change and is a bounded commitment.
Fixed price / unit (where they graduate):After a calibration run on 100–200 of your real tasks, we compute a flat rate per agent run. It can be tiered (Small/Medium/Large, assigned at intake) when task scope varies widely. Contracts reprice transparently on provider list-price changes, and going live is a one-line gateway endpoint change.
Fixed price / outcome (the goal):The top rung, once resolution rates are measured on your book: pay only for resolved units. Prices carry the failure and correlation load, so it fits workloads whose value per outcome clears it.
Routing upside:The same task can vary up to ~700× in cost depending on which model runs it. Second Moment picks the model for each task, within tiers you approve. Easy tasks go to cheaper models, and the savings are measured against a benchmark you can see.

What others are providing:

  • Cost observability (Datadog LLM obs, Langfuse, Helicone): a better view of the overage, which you still pay.
  • Gateways & routers (OpenRouter, Portkey, LiteLLM): a lower average cost, but the 10× unit still lands on your bill.
  • Vendors' own flat plans (Claude Max, Cursor, Replit): a seemingly unsustainable fixed price not based on real data, hence public repricings and walk-backs over the past year. Those retreats prove the demand for a fixed cost, and what happens to whoever offers one without pricing math.
  • Second Moment is the only party whose own balance sheet moves when your agent incurs excessive charges. A hard unit lands on our book at the same fixed price, and we can hold that position because we've measured the prior distributions and bounded the worst case (the workflow spec).

What an example contract looks like:

From the live demo's public benchmark dataset (illustrative, not customer data):

  • Fixed price $1.55 per ticket against a $1.40 average actual cost, ~11% premium in exchange for certainty.
  • A typical week at 5,000 tickets yields about +$750 in premium for us. A stressed week (1.75× provider cost shock) results in a $4,700 loss, absorbed by us, with your cost unchanged.
  • Backtested over 13 weeks of the workload's own history, the demo exhibit shows every losing week, not a smoothed average.
Open the interactive pricing demo →: price three real public agent workloads live (pick a workload, set volume → see tiers, cap fees, and the stress backtests with the weeks Second Moment eats), see the integration path, or drop in your own trace export. Pricing runs entirely in your browser, and your file never leaves the page.

Who should talk to us:

  • Vendors pricing per agent-run / per-resolution / per-outcome who own the spread between that price and their model bill (the founder at a startup or the product GM / finance at scale)
  • Dev-tools companies on flat-rate plans where a small share of users drives most of the cost (whoever owns plan economics)
  • Platform teams that promised an internal per-unit or fixed budget for agents (the platform lead who made the promise, with FinOps as the economic buyer)

If this isn't your desk, pass it to whoever owns your AI unit economics.

Why it's priceable (validated on 63k public agent runs)

Predictability
2×
the task itself drives cost, proven on 57k repeated runs
Data basis
63k traces
per-task costs across tickets, code, and research benchmarks
Worst case
2–16×
worst units vs the average by workload type, shown to be worse on cheaper models
Working PoC
Live
input trace export → output a priced quote and backtest

Low-risk by design: Trace exports carry summary fields only (ids, tokens, costs), never prompts. If our gateway is ever down, your agent falls back to calling providers directly, and the unenforced risk stays on our books. Integration is just a URL change with unchanged payloads, and every receipt is yours to keep.

Send a trace export → get a priced quote on your own data for free.
No commitment required · rahul@secondmoment.co · secondmoment.co
Validation report & methodology on request.