Your AI vendor is pricing by the hour now. Here's what to renegotiate.

The problem this solves

Anthropic shipped Managed Agents on Friday. The product headline is the sandboxing, the orchestration, the 10-point improvement in task success over standard prompting. The real news is the pricing line: $0.08 per session-hour on top of standard Claude API tokens. Early adopters already in production include Notion, Rakuten, Sentry, Asana, and Atlassian.

This is the moment when “consumption-based AI pricing” stopped being a deck slide and became the bill of materials on the next renewal quote.

For Series B founders, the timing is awkward. Contracts signed in 2024 and 2025 are coming up for renewal right as every serious agent vendor rewrites how it charges. Founders are being asked to sign deals denominated in a currency their finance teams have never had to forecast.

The approach

The instinct is to haggle the unit price. It is the wrong move. The per-session-hour and per-token rates are the most transparent, most benchmarked numbers in the contract. What moves the P&L sits in the four clauses next to that price.

1. Demand the consumption forecasting tool before signing

   A bill that cannot be modeled cannot be run. If the vendor cannot supply a forecaster with the customer's own 90-day telemetry plugged in, the buyer is flying blind. Ask for it in writing before any price discussion starts.

2. Cap the monthly downside, not the unit price

   A monthly ceiling converts a consumption bill into a predictable number the CFO can plan around. Founders pay the lower of actual consumption or the cap. Overages become vendor credit risk, not company cash risk.

3. Get a written portability clause covering agent state

   Fine-tuned model weights were the 2023 version of lock-in. The 2026 version is agent state: conversation memory, tool configurations, permission graphs. Without explicit export rights, there is no exit. Negotiate format, timing, and assistance hours.

4. Convert uptime SLAs into task-success SLAs

   The 10-point task-success bump in Anthropic's launch is a marketing benchmark, not a contractual guarantee. An SLA should measure whether the agent finished the job, not whether the API returned 200. Uptime proves the lights are on. Task-success proves someone is home.

Why most teams get this wrong

Most Series B founders treat consumption pricing as per-seat pricing with different units. It is not. Per-seat had natural ceilings (headcount) and natural floors (headcount minus six). Consumption pricing has neither. It scales with how hard the agents are working, which is correlated with nothing any finance team has previously budgeted for.

The second mistake is trusting the vendor’s calculator. It is accurate for the vendor’s reference customers, whose workflows and data shapes look nothing like a fast-growing Series B. Token-per-task ratios depend on the actual prompts, the context windows, and how chatty the internal tooling is. Run a two-week pilot on the new model. Build a calculator from the resulting telemetry before the contract closes.

The third mistake is skipping the exit clause because the relationship feels warm. Three years in, the sales rep who cared about the signature will have turned over twice. The export path needs to be written in now, while the vendor still wants the deal.

The numbers

The sticker looks small. At Series B scale, the math does not.

One autonomous agent running 40 hours per week across 20 workflows costs roughly $64 a week in session-hours alone. Add tokens. Add supervisor agents as soon as core agents start coordinating. A meaningful Series B deployment lands around $8,000 to $14,000 per month in runtime before a single model token gets counted. Tokens typically run 1.5 to 2.5x the runtime spend for decision-heavy agents.

Variance is the real surprise. I have watched spend swing 40% month over month on identical configurations because of seasonality in inbound volume. Finance teams need to accept that fact, or the contract needs a cap.

Ship it

The 20-minute version before the next vendor call.

Pull current contract expiration dates. List which vendors have already announced consumption pricing or are likely to this quarter. Send each one a single email: “We want to evaluate the consumption model. Please share usage telemetry from our last 90 days, the pricing calculator, a monthly ceiling option, and the portability clause.”

Ninety percent of vendors will send the calculator. Half will send the telemetry. One in three will offer a ceiling. Zero will volunteer the portability clause unless asked for it by name.

That is the whole negotiation. Not the unit price. The four clauses that sit next to it.