Benchmark Update — September 25, 2026: AI Just Rewrote Your Gross Margin, CAC Payback, and NRR Targets

Benchmarks are supposed to be boring. They move 50 basis points a quarter, you update the board deck, everyone nods. That era ended today. Three data points landed in the same 24-hour window — GitLab’s $1.13B revenue print, Stripe’s AI cohort growth figures, and OpenAI’s GPT-6 “Astra” launch — and together they invalidate a meaningful chunk of the 2025 SaaS benchmark set that most operating plans are still built on.

Here’s what actually moved, what number to replace it with, and how to reforecast before your next board meeting.

Benchmark 1: Gross Margin — The New Floor Is 400 Basis Points Lower

The single most consequential number in today’s data: GitLab is paying 400 basis points of gross margin for AI. Not as a one-time experiment. As a structural line item.

Context matters here. GitLab was, four months ago, the consensus “clearest AI casualty” in public B2B — code generation was supposed to eat the DevOps layer whole. Instead, the company posted $1.13 billion in revenue with 24% billings growth, having cut 14% of headcount, exited 22 countries, and shrunk its geographic footprint roughly 37%. That’s not a casualty. That’s a company that traded margin for relevance and won the trade.

But the 400bps is the benchmark that travels. The old playbook said inference costs were transitional — that model prices would collapse fast enough to absorb them within a year. Microsoft’s MAI-Transcribe-2 undercutting OpenAI, Google, and ElevenLabs on both price and speed proves prices are collapsing. And yet GitLab’s margin hit persists, because usage expands faster than unit costs fall. Every price cut gets consumed by deeper feature adoption.

What to change in your model

• Old benchmark: 78–82% gross margin for application-layer B2B SaaS at scale.
• New benchmark: 74–78% if you ship LLM-powered features as a default part of the product experience.
• Action: Reforecast your 2027 gross margin target down 300–500bps. If your plan still shows margin expansion alongside AI feature velocity, one of those two assumptions is fiction.

The sub-benchmark nobody is tracking yet: AI cost of goods sold as a percentage of revenue, reported separately from hosting. Start instrumenting it now. Within two quarters, investors will ask for it by name.

Benchmark 2: ARR Growth Rate — AI-Native Cohorts Are at 175%

Stripe’s payment data shows the AI-native cohort growing 175% year over year. Read that against the standard reference set and the distortion becomes obvious.

Traditional benchmark bands for 2025–26 held that a top-quartile Series B SaaS company grows 80–100% YoY, and anything above 150% was outlier territory reserved for a handful of names per vintage. The AI-native median is now sitting where the old outlier band was.

This creates a benchmarking trap. If you’re a traditional vertical SaaS company growing 60% and your investors are anchored to a peer set contaminated with AI-native growth rates, you will look broken while being perfectly healthy. Conversely, if you’re AI-native and growing 90%, you are underperforming your cohort even though that number would have been celebrated 18 months ago.

The correction

Segment your benchmark comparisons by cohort type, not just ARR band:

• AI-native (product impossible pre-2023): 150–200% YoY at $1–10M ARR is the new median expectation.
• AI-augmented incumbent: 20–30% growth is strong — GitLab’s 24% billings growth at $1.13B is the reference point.
• Vertical marketplace/booking: Growth equity is still funding these. Fresha just hit a $1B valuation on an $80M raise from KKR for beauty and wellness bookings — no AI narrative required. Category durability still gets priced.

Benchmark 3: CAC Payback and the Retention-First Diligence Flip

The quieter signal in today’s capital data: VCs are publicly narrowing the AI SaaS thesis. Thin wrappers, undifferentiated copilots, and “AI for X” without proprietary data are being screened out. The mechanical consequence is that proof-of-retention now precedes proof-of-growth in diligence.

That reorders the benchmark hierarchy. For three years, CAC payback was the gating metric — get under 18 months and raise. Now the gate is net revenue retention measured on cohorts old enough to matter.

• NRR benchmark, AI-native: 110% is no longer acceptable as a headline. Diligence is asking for 120%+ with logo retention above 85%, because AI product churn in the 2024–25 vintages ran brutally high once novelty faded.
• NRR benchmark, established B2B: 105–115% remains the healthy band. GitLab’s $20M of flex commitments in six weeks is the mechanism that gets you there — consumption pricing converting existing accounts into expansion revenue without new-logo CAC.
• CAC payback: The 18-month rule holds, but the numerator is now dirtier. If inference cost is embedded in your delivery, your true contribution margin per customer is lower, which lengthens real payback by 2–4 months versus what your dashboard shows.

That last point is the one that catches operators. Most CAC payback calculations use gross margin from the last audited period. If you shipped AI features six months ago, your payback math is running on stale margin assumptions and is understating recovery time.

Benchmark 4: LTV Is Now a Capability-Risk Variable

OpenAI’s GPT-6 “Astra” launch — marketed with explicit AGI framing — does something no benchmark table captures: it resets the frontier-capability baseline for every product built on top of a model.

If your moat was prompt engineering, your LTV assumption just broke. The mechanism is straightforward: features that sat in your product layer get absorbed into the base model