Benchmark Update — September 27, 2026: The Four KPIs That Just Got Repriced

Benchmarks don’t move because someone publishes a new survey. They move because the underlying unit economics shift and the survey catches up nine months later. Today is one of those days.

Three things landed on the same morning: OpenAI shipped GPT-6 Astra with explicit AGI framing while its companion tier launched at “half the price” of prior frontier access. Anthropic lost a $200M Pentagon contract over model-control clauses and pushed a $2T IPO to November. And Fresha, a beauty and wellness booking marketplace with zero AI-native positioning, raised $80M from KKR at a $1B valuation.

Each one hits a different line of your KPI stack. Here’s the read.

Data Integrity First

The KPI Hub’s search layer returned zero indexed results across all 15 queries today. Everything below is built on primary-source headlines plus previously published benchmark data that we cite by source and vintage. Where we extrapolate, we say so. If a number isn’t attributed, treat it as directional — not decision-grade. That distinction is the entire product.

CAC and CAC Payback: The Denominator Is Moving Faster Than the Numerator

The standard reference points most SaaS operators still quote come from the 2024–2025 benchmark cycle: median CAC payback for private B2B SaaS clustering in the 18–24 month range, with best-in-class sub-12 and anything past 30 months flagged as structurally unhealthy (SaaS Capital and ICONIQ Growth survey cohorts, last published vintage). Those numbers were computed in a world where inference cost was a rounding error or a fixed line item.

GPT-6 Sol & Luna launching at half the price of prior frontier access breaks that assumption in both directions.

The good news

If AI features are a material share of your gross margin, a 50% cut to frontier access is a direct gross-margin expansion event. CAC payback is calculated on gross-margin-adjusted revenue — so a two-point gross margin improvement shortens payback without you acquiring a single additional customer. If your AI COGS ran 8–12% of revenue and half of that is frontier-model spend, you just recovered roughly 200–300 basis points of gross margin. On a 22-month payback, that’s roughly a one-to-two month improvement. Inference, not a surveyed figure — run it against your own COGS split.

The bad news

OpenAI’s internal forecast puts total burn through 2030 at $278 billion (via 20VC x SaaStr). A vendor burning that much capital does not hold prices flat forever. The correct planning assumption is a 12-month AI COGS renegotiation cycle, not 36 months. If you’ve modelled three-year gross margin expansion off today’s inference pricing, your LTV is overstated.

Practical revision: run two CAC payback scenarios — one at current inference pricing, one at a 40% cost increase on the same volume. If the second scenario pushes payback past 30 months, your AI feature set is a marketing expense wearing a product costume.

Churn and NRR: Control Clauses Are the New Concentration Risk

The Anthropic-Pentagon break is the most underrated KPI story of the quarter. A $200M contract died — not over price, performance, or procurement politics, but over model control terms: autonomous weapons and domestic surveillance use. DoD moved to alternate vendors.

Net revenue retention benchmarks have been grinding down for three years. The last widely cited private-company medians sat around 102% for B2B SaaS overall, with enterprise-focused cohorts holding 105–110% and SMB cohorts frequently below 100% (SaaS Capital 2025 survey vintage). Public comparables tell the same story — Freshworks has reported net dollar retention in the 104–107% band across recent quarters, a long way from the 115%+ that defined 2021-era enterprise SaaS.

What Anthropic adds is a new churn vector that doesn’t appear in any existing benchmark: values-based termination. Federal and large-enterprise ARR now carries a control-clause