loading
loading
The Bottom Line: Every B2B company is spending real money on AI in 2026: seats for ChatGPT and Claude, token bills from coding agents, and a growing line item for agents that run without anyone watching. Very few can say what that spend is producing. Larridin is built to answer that question. It connects AI usage and spend to the work of people and agents across a business, covering adoption and fluency, engineering performance, and opportunities to automate repeated work. It’s a16z-backed, run by a founder on his sixth company, and it just published the most useful data I’ve seen this year on what AI coding costs per engineer and where the returns stop. $213 a Week Per Engineer, and a 10x Spread Start with the data, because it’s the best argument for the product. Larridin published its first benchmark in August, built from production billing and engineering telemetry. The sample is engineers who, in the four complete weeks ending August 2, 2026, both merged code and drew billed AI-coding spend. The findings: - The median engineer draws $213 a week in billed AI-coding spend, roughly $920 a month. The 90th percentile hits $911 a week, a spread of more than 10x. - Those figures are a floor. Engineers on flat-fee plans consume tokens that never show up on a metered bill, so true consumption runs higher. Annualize the 90th percentile and you’re near $47,000 a year per engineer in tokens alone. For a 100-person engineering org, the spread between your p25 and p90 engineers is a seven-figure budget question that most CFOs can’t see today, because it’s scattered across provider invoices, corporate cards, and personal subscriptions. What managing the bill looks like in practice: Coinbase’s Kyle Cesmat on cutting AI inference costs by more than half while token usage kept climbing: Same Company, Same Tools, Same Prices: 2x the Output The second finding is the one founders should sit with. Larridin split engineers into cohorts by how much of their shipped output was AI-attributed. Both AI-native cohorts came from the same company, with the same tools, the same prices, and the same starting spend of about $170 a week, so the gap between their curves reflects skill. - Deeply AI-native engineers (79% AI-attributed): reached 11.8x output at around $1,300 a week and hadn’t hit a ceiling. At equal spend, this group shipped roughly 2x what partial adopters shipped. - Partially AI-native (37% AI-attributed): got real returns, but marginal payoff dropped by about half once spend passed roughly $600 a week. - Low-AI (15% or less AI-attributed): topped out around 1.9x and stayed flat across a 20x spend range, from $21 to $421 a week. Extra dollars bought activity, and no additional output. The practical read for anyone setting a 2027 AI budget: more budget converts to more output only where fluency already exists, so there’s no universal ideal budget. Larridin’s advice is to track your own team’s ROI curve and set review triggers where it levels off, rather than copying another company’s cap. Output here isn’t lines of code. Each merged PR is scored by model-assessed complexity on five levels, discounted for low-quality output and missing tests, and scaled by code churn. And Larridin flags its own limits, which I appreciate: the relationships are associational, since high-output engineers may spend more because they ship more. What the Platform Does: Four Products Around One Question Larridin started as a discovery tool. Early on it ran automated inventories of the AI tools in use, including shadow AI employees access outside IT’s view, then compared productivity across people with different levels of AI adoption. It has since grown into four products: Spend Intelligence. Token usage, seat licenses and cloud model costs in one view, tracing every dollar to a team, a tool or an agent before the next budget review. Built for CFOs and finance teams. This is the one I’d start with. Agent spend is the fastest-growing and least-understood line in most AI budgets, and it doesn’t map to a seat or a person. AI Impact. Connects what each team spends on AI to the hours it returns, and compares adoption, fluency and cost per AI hour to decide where to invest, where to train, and what to scale. Larridin is careful about the math here: AI capacity estimates the human-equivalent work AI contributes, and it is not automatically time saved, cash returned or a reduction in headcount. That distinction matters. Plenty of vendors will hand you an “hours saved” number and let you present it to your board as savings. Developer Intelligence. Connects engineering output, code quality and delivery to AI spend, and shows where coding agents help and where teams need support. The newest piece is Larridin Router. It scores each coding request and serves a lower-cost model when the task allows, leaving everything else on the model the developer asked for. The part that separates it from a plain model router: it ties routed sessions to output, quality, defect rate and cost per task, so you can see whether the discount held up in code review. A developer who needs a specific model for one job can pin it, and that session gets reported as pinned rather than counted against the router. Workflow Intelligence. Maps repeated work from observed activity and takes the best candidates from identified to automated, measured against a captured baseline. Gainsight Used It Before Buying Its First Enterprise LLM Gainsight used Larridin to understand AI tool adoption and inform its first enterprise LLM purchase. Larry Hill’s point from that story is simple: without knowing what your people use today, you don’t know what to buy next. That’s the most common AI procurement mistake I see. Companies sign a big enterprise license based on a pilot, then find half the org is still on personal accounts of a different tool. Other customers on the site include Vertiv, Gainsight, Klaviyo, SurveyMonkey, EcoVadis, TigerConnect and Sundt. It’s SOC 2 Type II, GDPR and HIPAA compliant. ClickUp CFO Dan Zhang on the AI investment framework that actually scales: Russ Fradin’s Sixth Company, With $17M From a16z Russ Fradin has been doing this a long time. He’s been founding and exiting companies for 30 years, from Flycast Communications in 1996 through Larridin. He previously ran Dynamic Signal and Adify and was at comScore. President Jim Larrison came from Dynamic Signal, Firstup and comScore, and CTO Ameya Kanitkar from LinkedIn, Coinbase and Groupon. The Dynamic Signal story is the relevant one. After 18 months and $5-6M in ARR, Russ concluded customers were happy and paying well but the business wasn’t sticky, and walked away from it. The rebuilt company became a $50M ARR employee communications business. A founder who’s already killed millions in ARR because it wasn’t recurring is exactly who you want building a measurement product, where the whole value is being the system people check every week. Larridin raised $17M in seed funding led by a16z, with Alex Rampell on the board, alongside Bloomberg Beta, Gradient, Haystack, Homebrew, and Refract. They started building in early 2024 and began selling in August 2024. Where It Gets Harder: Pricing, Privacy, and Causation A few things to know going in. It’s enterprise-priced. Pricing isn’t public. A competitor’s comparison piece puts enterprise pricing starting around $50K a year. Treat that number with caution given the source, but plan for an enterprise sales motion. Adoption tracking touches employee monitoring. Usage tracking runs through browser plugins and desktop agents. Larridin supports role-based access and enterprise authentication, and lets you choose the scope of measurement. Decide what you’re measuring, and tell your team, before you roll it out. Engineers who feel watched will route around the tooling. Correlation isn’t proof. Larridin says this about its own benchmark. Use it to find where spend has stopped converting, then test the fix. Who Should Look at Larridin - CFOs with an AI line item that doubled this year and no way to attribute it by team, tool, or agent. - CTOs and VPs of Engineering deciding whether to raise, cap, or reallocate coding-agent budgets for 2027. The benchmark alone is worth reading before that meeting. - Heads of AI who need to show the board which departments turned licenses into working capability and which are still paying for access nobody uses. - Companies running agents in production. At SaaStr we run 20+ agents with 3 humans. Agent spend doesn’t sit on anyone’s seat, and it’s the category most in need of this kind of attribution. If you’re a 20-person startup where every engineer is deeply AI-native and the founder reads the Anthropic invoice personally, you don’t need this yet. Once you’re past a few hundred people, or once agent spend passes seat spend, you probably do. Where to Start Read the Larridin benchmark on AI coding spend first. Then pick one decision you have to make this quarter, a renewal or a coding-agent rollout, and measure that. Larridin is also offering $5,000 in Router credits to conference attendees. Visit larridin.com. Meet Larridin at SaaStr AI 2027 Larridin is a Super Gold sponsor of SaaStr AI 2027, May 11-12 in the SF Bay Area. Bring your AI budget questions and your token bills. Russ and the team will be there to walk through what your spend is buying. SaaStr AI App of the Week is a weekly series highlighting the AI tools B2B companies are running in production.
document fetched 2026-09-28 02:55 · canonical https://saastr.com/saastr-ai-app-of-the-week-larridin-the-median-engineer-now-bills-213-a-week-in-ai-coding-tokens-larridin-tells-you-what-it-bought