The narrative that dominated last week’s CNBC coverage was that AI is getting cheap. Tomasz Tunguz, partner at Theory Ventures, published a post on May 20 that tells a more complicated story. Prices dropped when labs were buying market share, and prices are rising now that margins matter.
Tunguz’s data covers the three major vendors at their current flagship tiers. Google’s Gemini 3.1 Pro sits at $2.00 per million input tokens and $12.00 per million output tokens. Anthropic’s Claude Opus 4.7 runs $5.00 input and $25.00 output. OpenAI’s GPT-5.5 lands at $5.00 input and $30.00 output. Google is still the low-cost player, but Tunguz notes its models have been tripling in price roughly each year. OpenAI’s flagship was subsidized for a stretch before prices rose again. Anthropic held a premium position for most of the past two years before recently cutting rates on its most powerful tier.
The pattern Tunguz identifies is clean: cuts come when cash is plentiful and share matters, increases come when cash is tight and margins matter. He says the latter now describes all three vendors. That assessment lands on the same day Anthropic is tracking toward $10.9 billion in Q2 revenue, a figure that sounds like abundance until you price in what labs are spending on compute. Capex at the frontier has not stopped breaking records, and no lab has demonstrated a path to margin that does not eventually flow through higher API pricing.
The “cheap AI” framing from CNBC and similar outlets is not wrong on the tactical level. There are genuinely low-cost models available, budget-tier pricing has expanded, and tokenizer efficiency gains mean a dollar buys more output than it did two years ago. But Tunguz is pointing at the flagship tiers, which is where the enterprise contracts and the serious workloads sit. If the pattern holds, flagship prices will keep rising even as open-weight and budget models hold the bottom of the market down.
My read is that Tunguz is closer to correct than the “cheap AI” camp. The cheap layer is real, but it is not the layer enterprises negotiate contracts against when deploying at scale. The vendors all face the same math: frontier training costs are not declining, inference hardware cycles are expensive, and the competitive pressure to release the next model is relentless. Cutting prices on a model that is already two generations old costs nothing. Cutting prices on the flagship you are betting the company on is a different calculation.
The broader framing matters because procurement decisions are being made against narrative as much as against pricing sheets. An enterprise that budgets for 2026 AI spend assuming the cheap-AI thesis holds may find itself renegotiating contracts mid-year as flagship rates continue to drift upward.
Operators locking in annual AI commitments right now should price their flagship-tier usage against a 15 to 20 percent upward drift over the next twelve months, not a flat or declining curve.
Posted by Tomasz Tunguz on 2026-05-20.
