Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
anthropic vs. openai
TL;DR
Anthropic quietly overtook OpenAI in new business adoption — Ara Kharazian from Ramp said that in January 2026, first-time AI-buying companies flipped from choosing OpenAI at roughly a 60/40 rate to choosing Anthropic instead, driven less by a single model and more by enterprise product traction like Claude Code and Claude Co-work.
Ramp sees enterprise AI through real receipts, not vibes — because Ramp covers 50,000 businesses and about $100 billion in annual spend, it can inspect invoices showing whether companies are paying for Opus vs. Sonnet, Claude Code vs. API, giving it unusually granular visibility into model-level enterprise usage.
AI spend is exploding faster than ROI is being proven — Kharazian said API-heavy AI spend per company is up roughly 13x-14x year over year, but still under 1% of non-payroll spend, which means budgets haven’t broken yet even as companies like Uber reportedly burned through an annual AI budget in one quarter.
Coding is the center of gravity, but not the whole story — about 80% of business AI spend in Ramp’s data is API usage powering products and backend AI experiences, while most subscription dollars cluster around higher-end plans tied to coding agents, especially in software, finance, and professional services.
The expensive frontier models still capture most revenue, for now — even though cheaper models like Gemini Flash and Haiku are widely used, Kharazian said they contribute a small minority of enterprise dollars because the highest-performing models command the spend; he expects that mix to shift as CFOs push harder on budgets and routing.
Google may be more present than the paid data suggests — Gemini looks weak on Ramp’s paid-subscription chart, but Kharazian argues that Google’s bundling into Workspace likely hides substantial adoption among non-technical users, while its lack of a true Claude Code/Codex-style developer product has limited its breakout with engineers.
The Breakdown
A livestream that turns into a data-rich enterprise AI check-in
Matthew Berman opens with his usual global-roll-call energy, then brings on Ramp economist Ara Kharazian to talk about something more concrete than model leaderboards: actual business spend. The premise is simple but useful — if AI is really changing work, you should be able to see it in invoices, contracts, and budget lines before it shows up in GDP.
Why Ramp has one of the weirdest and most valuable AI datasets
Kharazian explains why an economist even exists at Ramp: the company sees spend across 50,000 businesses and roughly $100 billion annually, including uploaded receipts and invoices. That means Ramp can often tell not just that a company pays Anthropic or OpenAI, but whether it’s buying Claude Code, API access, Opus, Sonnet, or Gemini Flash — a much sharper lens than public chatter.
The big caveat: AI hype is real, but CFO math is coming
Berman pushes on the obvious tension: infrastructure investment and revenue are exploding, but productivity gains still feel fuzzy. Kharazian agrees the pressure is real — AI spend per company is up about 13x-14x over the last year, still small as a share of total spend, but growing fast enough that companies will eventually stop treating it like an experiment and start asking harder ROI questions.
Coding dominates the narrative because that’s where the money is working
When chat asks whether coding really is AI’s main use case, Kharazian adds nuance: around 80% of business spend in Ramp’s data is API usage, often for AI-native product features or backend workflows, not just internal copilots. But the heaviest subscription spend is absolutely tied to coding agents and technical workflows, especially in software, finance, and professional services — the earliest adopters sitting around 75-80% AI adoption.
Anthropic’s surprise surge past OpenAI
The Wall Street Journal hook for the episode is a Ramp chart showing Anthropic overtaking OpenAI in paid enterprise adoption. Kharazian says the real turning point came in January 2026, when brand-new AI-buying businesses abruptly switched from favoring OpenAI to favoring Anthropic; he frames it as product execution, not just benchmark wins: Anthropic nailed technical users first with Claude Code, then expanded outward with tools for less technical teams.
Frontier models make the revenue, but cheaper routing is the next fight
Even though everyone talks about cheap workhorse models, Kharazian says enterprise dollars still cluster around the expensive frontier systems because that’s where the performance is. But both he and Berman keep circling the same coming shift: CFOs will force more intelligent routing, more OpenRouter-style usage, and more task-by-task matching instead of defaulting every job to the priciest model.
Google, open source, and the gap between usage and monetization
Berman asks why Gemini looks so weak on the chart, and Kharazian points out a measurement blind spot: if Gemini is bundled “for free” inside Workspace, paid-spend datasets undercount it. On open source, the two mostly agree that there’s real demand but a brutal business-model problem — building a frontier model is wildly expensive, and once it’s open, lower-margin inference players can eat the value.
Competition is real, lock-in is partial, and SaaS isn’t dead
One of the more grounded takeaways is that these models aren’t pure commodities yet: Claude can be more expensive and more annoying operationally, and developers still stick with it. Kharazian closes by arguing that the endgame is not one model to rule them all — big AI buyers already use multiple providers, switching costs are low, and AI-native SaaS companies like Figma are proving they can absorb the new capabilities rather than just get steamrolled by them.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
