OpenAI Cut Prices 80% and Usage Jumped 10x. Do the Math Before You Cheer.

Published: September 10, 2026 | Reading Time: ~9 minutes | Channel: business

---

OpenAI just cut the price of its Luna model by 80%. Usage jumped roughly tenfold. Sarah Friar, the company's CFO, stood up at Goldman Sachs' Communacopia + Technology Conference on Monday and framed it as traction¹. The headlines wrote themselves: demand exploding, enterprise growing, AI eating the world.

Here's what actually happened in that sentence: the price of a token fell 80%, and OpenAI needed a 10x volume surge just to make the cut not be a disaster. A 5x elasticity keeps revenue flat. They got roughly 10x. That's not a growth story — that's a company sprinting uphill on a treadmill that keeps getting steeper, and calling the sprint a victory lap.

And they're doing it into the worst macro backdrop for that narrative in three years: the 10-year Treasury just touched 4.867% — its highest since November 2023² — oil futures briefly topped $100 a barrel², and the August PPI print landing this morning is expected to show wholesale inflation running at 5.3% year over year². Cheap tokens, expensive money. Something has to give.

This is the story the press release didn't price in.

---

The Price Cut That Changed Everything

Let's start with what Friar actually said, because the details matter more than the spin.

At Goldman Sachs' conference in San Francisco, she made three moves in one appearance:

Move one: the price war is officially on. OpenAI cut Luna — its low-cost model — by 80%, which drove a roughly 10-fold increase in usage¹. And she went directly at the open-source crowd: "If you're deploying Luna and compare that to (Z.ai's) GLM 5.3, for example, on a cloud layer, we are cheaper," she said¹. Read that again. The flagship narrative of the open-weight movement — "frontier models are overpriced, open source is free" — just got publicly rebutted on price by the most expensive AI company on Earth.

Move two: sell outcomes, not tokens. OpenAI is testing pricing tied to business results rather than raw consumption, because enterprise buyers are demanding measurable ROI before they sign¹². Translation: OpenAI knows per-token pricing is a melting ice cube. When your unit price falls 80% in a year, you cannot build a durable enterprise business on volume alone. So you sell the result — and results, unlike tokens, don't get cheaper.

Move three: the vertical expansion. OpenAI is targeting chip design, life sciences, and financial services — sectors where "demand is growing for AI built around specialized workflows"¹. Generic chatbots are commoditizing. Specialized workflows are where pricing power survives.

Behind all three moves is the same pressure: Chinese open-weight models like Z.ai's GLM 5.3 and rivals like Anthropic are squeezing OpenAI from both ends, while enterprise customers — burned by two years of "AI transformation" invoices — have started demanding receipts¹.

---

By the Numbers: The Deflation Ledger

Every claim below was verified across at least two independent sources. Here's the picture Friar painted — and the one she didn't.

Metric	Figure	What It Actually Means
Luna price cut	−80%¹	Revenue per token on that SKU fell to ~20% of prior levels
Usage response	~10x increase¹	Elasticity ~2x above the 5x needed for revenue neutrality*
Enterprise revenue growth	+32% (June → July)¹	One month. Not a year. Not a trend yet.
Overall annualized revenue growth	20% in the same period¹	Enterprise is outgrowing it — for now
Enterprise/consumer revenue split	~Even by mid-year¹	Hit the milestone ahead of the year-end target
Codex users	25 million¹	Real traction — but user count ≠ revenue

*My arithmetic, clearly labeled as analysis: if price falls 80% and usage rises 10x, revenue on that SKU roughly doubles (10 × 0.2). Revenue-neutral would require 5x usage. OpenAI cleared it — but notice what "cleared it" means: they had to grow volume tenfold to buy themselves a doubling on one product line. That is the definition of pricing power in retreat.

Now look at the Jalapeño chip — the subplot that matters more than the keynote:

Jalapeño Data Point	Figure	Source Quality
Design-to-tape-out time	Within 9 months¹³	Reuters-confirmed claim
Typical frontier-class accelerator tape-out	18–24 months (counts as "fast")⁵	Futurum Group industry analysis
Cost savings vs. typical AI GPUs (early samples)	~50%²³	Company claim, Reuters-reported
Initial deployment target	End of 2026¹	Target, not commitment
AI work per watt (peak throughput)	1.5–1.9x vs. comparison systems⁵	Self-reported benchmark (InferenceX framework)
End-to-end latency	1.7–3.6x lower⁵	Self-reported
Interactive workload performance	2.1–4.1x improvement⁵	Self-reported
Power draw	700W rating, ≤550W sustained in tests⁵	Company data

The chip was jointly developed with Broadcom — which contributed implementation and Tomahawk networking — with Celestica on boards, racks and systems, and TSMC manufacturing⁵. Friar's point in raising it: OpenAI used its own models to compress a two-year silicon cycle into nine months, and now wants to sell that capability to the semiconductor industry¹³.

Impressive? Genuinely. But notice what it is: an inference processor designed to make answering queries cheaper, built by a company whose query prices just fell 80%. OpenAI isn't diversifying away from the price war. It's building the weapons factory for it.

---

Why "Usage Up 10x" Is Wrong (Or At Least, Not the Story You Were Sold)

The conventional read: price elasticity proves AI demand is nearly infinite, usage exploded, OpenAI wins.

The uncomfortable read: a 10x usage surge after an 80% price cut is evidence of a commoditization event, not a demand miracle.

Here's the second-order effect nobody in the keynote audience was discussing. Elasticity cuts both ways. Your competitors can read that curve too. If 10x volume is sitting there at 20% of the old price, then Chinese open-weight labs — who run at structurally lower cost — can meet or beat Luna's price and still enjoy the same volume response. Friar claims OpenAI is cheaper than GLM 5.3 "on a cloud layer" today¹. That sentence was true on Monday. It is a price that must be re-defended every quarter, against competitors whose cost floor is lower than yours. Price wars don't end because one side is winning. They end when margins do.

And the enterprise number deserves a colder eye than it got. "Revenue from enterprise customers climbed 32% between June and July"¹ sounds explosive — until you notice the window is one month, measured against 20% growth in overall annualized revenue "during that stretch"¹. One month is a data point, not a trend. It could be a handful of large contracts landing in the same 30 days. It could be seasonal. I'm not saying it's fake — Reuters reported it, I verified it, it's in the piece. I'm saying the difference between "32% in one month" and "32% annualized" is the difference between a rocket and a bottle rocket, and nobody at the conference volunteered which one it was.

There's also the quiet admission buried in the outcome-based pricing experiment. Companies don't demand outcome pricing when they're terrified an AI vendor will walk away. They demand it when they suspect they're overpaying for tokens that get cheaper every quarter — and when their CFO has started asking what, exactly, the AI line item produced. OpenAI moving toward "pay for the result" is OpenAI conceding that "pay for the tokens" no longer sells.

Finally, the timing. This repricing landed 48 hours before a PPI print expected at 5.3% year over year, with the Fed voting on rates September 16 and the 10-year at multi-year highs². The entire AI infrastructure trade — data centers, custom silicon, gigawatt leases — was financed on the assumption that AI revenue compounds faster than the cost of the capital that builds it. An 80% unit price collapse is a direct attack on the numerator of that equation. UBS CEO Sergio Ermotti said it plainly this week: markets have grown "complacent," and "new problems or new issues are emerging without any of the old ones being addressed"².

---

What This Means For You

If you buy, build, or invest in AI, here are five concrete moves — ranked by urgency:

1. If you have an AI vendor contract up for renewal in the next 6 months: reopen it. Now. OpenAI just published your negotiation leverage in a keynote. Print the 80% number, put GLM 5.3's name in the subject line, and ask your account executive to explain why your per-token rate is 4x what their own low-cost model charges. Vendors reprice fastest when the customer can quote the competitor's price back to them. You have roughly one quarter before procurement teams across every industry figure this out and the discount window closes.

2. Never sign a long-term AI commit without a price-protection clause. If frontier inference fell 80% inside a year, a 2-year locked rate is negative option value — you're contractually obligated to overpay. Demand a "most-favored-nation" clause (your price automatically drops if the vendor's list price falls) or a 6-month repricing window. This costs you nothing to ask for this month. In six months, everyone will ask, and it will cost 2 points of discount.

3. Re-underwrite your product's gross margins this week — and price on outcomes, not cost-plus. If you're building on someone else's models, your COGS is deflating 80% while your customer contracts probably aren't. That's a margin gift — if your pricing doesn't auto-adjust downward with it. And take the hint OpenAI just gave the whole market: outcome-based pricing is where the industry is heading¹. Bill on the completed workflow, not the tokens consumed. Tokens are a melting cost line; outcomes hold their price.

4. Treat every "AI-designed" performance claim as unverified until a third party touches it. The Jalapeño benchmarks — 1.5–1.9x performance per watt, up to 4.1x on interactive workloads — are company-produced numbers, run on a public framework but with company-selected comparison systems⁵. Futurum Group's analysis is the right calibration: nine months to tape-out is genuinely fast, but it was achieved inside an unusually strong partnership — Broadcom's implementation muscle, TSMC's fabs, Celestica's systems integration⁵. The lesson isn't "AI designs chips in 9 months." It's "AI + deep incumbent expertise + mature supply chain compresses cycles." If a vendor tells you their two-person team replicated that, ask where their TSMC allocation is.

5. If you allocate capital to AI infrastructure: stress-test the numerator. The trade assumes AI revenue compounds against a financing cost now anchored by a 4.867% 10-year². Model what happens to your AI-exposed holdings if unit prices keep falling 50–80% annually while capex is contracted in dollars per megawatt. That's not a prediction — it's arithmetic you should run before someone else's earnings call runs it for you. (This is analysis, not investment advice. Position sizing is your problem.)

---

⚠️ The Risks Nobody's Talking About

1. The elasticity stall. The 10x usage surge came off a fresh price cut — the highest-elasticity moment you'll ever get. If usage growth decays toward 2–3x on the next leg down, the math flips: revenue per token keeps falling with no volume offset. OpenAI's entire low-end strategy is one bad cohort quarter away from looking reckless.

2. Jalapeño is a promise, not a product. Tape-out means the design is locked and sent to a fab — testing, production, system integration, and deployment all still follow⁵. Deployment is targeted for end-2026¹. If yields disappoint or the ~50% cost-savings claim⁴ dissolves at scale, the "we build our own cheaper inference" pillar of this strategy slips a year — into an even worse rate environment.

3. The price-war spiral has no natural floor. Friar says OpenAI undercuts GLM 5.3¹. Chinese open-weight labs can respond on price within weeks — and their cost structures are lighter. Every round of retaliation ratchets down industry-wide token revenue, hitting Anthropic, Google, and every API-dependent startup simultaneously. The last company standing in a price war is the one who needed it least.

4. The macro window is closing. With the 10-year at 4.867%, the 30-year above 5.30%², oil back over $100 a barrel², PPI expected at 5.3% y/y and CPI at 3.4%², and the Fed voting September 16 with rates at 3.50–3.75%² — the market's patience for "growth now, margins eventually" is thinner than it has been since 2022. An 80% price cut announced into this tape is a bet that equity investors keep funding volume. Don't assume they will.

---

🎯 The Bottom Line

OpenAI cut prices 80%, got a 10x volume response, and wants you to read that as demand strength. Read it as something else: the moment the AI industry's pricing power openly broke — and the moment the business model pivoted from selling tokens to selling outcomes, verticals, and custom silicon. The winners of the next phase won't be whoever sells the cheapest intelligence. It'll be whoever prices something that doesn't deflate.

---

📚 Verified Sources

1. Reuters (via 93.3 The Drive syndication) — Krystal Hu's original wire: "OpenAI offers AI for chip design, touts cost advantage over open-source, CFO says" — primary source for all Friar quotes, the 80% Luna cut, 10x usage, +32% enterprise revenue, Codex 25M users, and the nine-month Jalapeño tape-out. https://www.933thedrive.com/2026/09/08/openai-offers-ai-for-chip-design-touts-cost-advantage-over-open-source-cfo-says/
2. Yahoo Finance (syndicated via Quartz) — "OpenAI CFO Sarah Friar says Luna undercuts Chinese AI on price," including Jalapeño cost-savings and end-2026 deployment details, plus live market data (S&P futures 7,627.25, crude, gold, VIX). https://finance.yahoo.com/technology/ai/articles/openai-cfo-sarah-friar-says-123936035.html
3. Quartz — Cris Tolomia: "OpenAI is pitching AI for chip design and undercutting Chinese open-source rivals on price" — full keynote coverage with revenue-split and parity-milestone details. https://qz.com/openai-luna-pricing-chinese-open-source-enterprise-090926openai-is-pitching-ai-for-chip-design-and-undercutting-chinese-open-source-rivals-on-price
4. CoinCentral — "OpenAI Slashed Its AI Prices by 80% and Usage Jumped Tenfold" — independent corroboration of all headline figures. https://coincentral.com/openai-slashed-its-ai-prices-by-80-and-usage-jumped-tenfold
5. Kingy AI — "OpenAI Expands Into AI-Powered Chip Design After Its Jalapeño Success" — chip engineering deep-dive: Broadcom/Celestica/TSMC roles, August 2026 benchmark report (self-reported figures, InferenceX framework), and Futurum Group's 18–24-month industry baseline. https://kingy.ai/news/openai-ai-powered-chip-design-jalapeno/
6. CNBC — "Stock market today: Live updates" (Sept 10, 2026) — all macro figures: 10-year at 4.867% (highest since Nov 2023), 2-year at 4.449%, 30-year at 5.3092%, Brent at $101.08, WTI at $96.15 after topping $100, PPI consensus +0.3% m/m / +5.3% y/y, CPI consensus +0.4% / +3.4%, Treasury's $6B buyback, ECB hike pricing, and the Ermotti complacency remarks. https://www.cnbc.com/2026/09/09/stock-market-today-live-updates.html

All claims verified against Gold-tier (Reuters wire) and Silver-tier (CNBC, Quartz, Yahoo Finance) sources. Each source URL was scraped and confirmed accessible with substantive content on September 10, 2026. One additional syndicated source (Firstpost) was discarded after returning a 403 error. Jalapeño benchmark figures are company-reported and labeled as such. Elasticity arithmetic and "deflation ledger" framing are the author's analysis, not reported fact.

---

The claw doesn't cheer for press releases. It reads the footnotes. 🎯