If you only track one thing in AI as a founder, track the gap between two numbers that both landed in the last two weeks.

The first: $1 per million tokens. That's the input price of GPT-5.6 Luna, the cheapest of the three models OpenAI shipped on July 9. The second: 89%. That's the share of all tracked AI-startup revenue now captured by just OpenAI and Anthropic, per The Information.

Read together, they describe the actual shape of the market you're building in: the intelligence is getting cheap, and the money is getting concentrated. Those aren't contradictory. They're the two halves of the same story, and the space between them is where a startup either finds a moat or quietly becomes a thin wrapper on someone else's API.

Here's the week, read for what you should do about it.

The 30-second version#

1. The model layer is in a price war#

A year ago, "frontier-quality" output meant paying frontier prices. That link is breaking.

GPT-5.6's middle tier, Terra, is explicitly positioned as GPT-5.5-class quality at roughly half the cost — and Luna takes cost-sensitive workloads down to $1 input / $6 output. On the same two-day window, SpaceXAI's Grok 4.5 landed at $2/$6 with Elon Musk calling it "Opus-class, but faster and lower cost," and the independent benchmarking firm Artificial Analysis clocked it near the top of its task-completion leaderboard at about $0.49 per completed task — roughly 90% cheaper than the models ranked above it.

And then there's the floor under all of it: GLM-5.2, an open-weight model from China's Z.ai, MIT-licensed, running at $1.40/$4.40 and — by independent benchmarks — beating GPT-5.5 on parts of the coding suite. Open weights mean you can rent it or run it on your own hardware, which caps how much any hosted vendor can charge for comparable quality.

When "good enough for production" costs a fifth of what it did last year, model choice stops being an identity and becomes a line item.

What to do: Stop hard-coding a single model. Keep a small eval set of your real tasks, route calls through an OpenAI-compatible gateway or router, and re-benchmark whenever a plausibly-cheaper model ships. The goal isn't to always run the cheapest option — switching has real costs in re-tested prompts and eval drift — it's to make switching a config change, not a rewrite. (We've written the founder's buyer's guide to picking an LLM API without lock-in separately; if you want the developer-level breakdown of the new OpenAI tiers, see GPT-5.6 Sol vs Terra vs Luna.)

2. The money went the other way#

You'd think commoditizing intelligence would spread the revenue around. The opposite is happening.

The Information now tracks 34 leading AI startups generating roughly $80B in annualized revenue — and 89% of it flows to just OpenAI and Anthropic, a share that rose 4.5 points in six months. Anthropic reportedly passed OpenAI on the strength of its coding tools. Everyone else — Perplexity, ElevenLabs, Cognition, all reportedly past $500M ARR — is fighting over the remaining sliver.

The concentration shows up in the infrastructure too. On July 6, Anthropic signed a 20-year lease with the bitcoin-miner-turned-data-center operator TeraWulf worth about $19B in contracted revenue for up to ~401 MW of AI compute in Kentucky. That is not a company hedging its bets. That is a company that intends to own the substrate for two decades.

What to do: Assume the foundation-model layer is a two-horse oligopoly for the medium term, and build so you don't compete with it. If your product is a general assistant or a thin model wrapper, the labs will ship your feature and undercut your price. If your product uses their model as one input among several, their price war works for you.

3. Where the defensible startups are actually betting#

The most useful signal isn't the model launches — it's what got funded around them. Follow the money and a pattern falls out. The big rounds this summer bought three things a foundation model can't hand you:

None of the three is trying to out-model OpenAI. Each owns something upstream or downstream of the model. That's the template.

The takeaway for founders#

The week's headline isn't "GPT-5.6 is out" or "Grok got cheaper." It's the structural fact underneath: the model is becoming a cheap, swappable input, and the durable value is moving to whatever the model can't provide — your data, your distribution, your ownership of a specific workflow.

So ask the uncomfortable version of the question about your own startup: if a competitor could call the same model you do, for a dollar a million tokens, tomorrow — what's left that's yours? If the answer is "our prompt" or "our nicer UI," this week was a warning. If the answer is a dataset, a channel, or a wedge no lab will bother to fight for, this week was a tailwind.

Cheap intelligence is the best thing that ever happened to a founder who isn't selling intelligence.