OpenAI's CFO says it's pushing into chip design, health care and finance -- and testing pricing tied to results, not tokens

OpenAI CFO Sarah Friar told investors at Goldman Sachs' Communacopia + Technology Conference on Sept. 8 that the company is pushing beyond chatbots and coding assistants into specialized industrial work -- chip design, life sciences and financial services -- while testing a pricing model tied to business results rather than how many tokens a customer consumes. It's a CFO's version of a product announcement: no new model, no launch, just a spending and pricing strategy laid out in front of the people who fund it. Two of the numbers she offered are checkable against OpenAI's own business, and they hold up. The third is a comparison that's accurate exactly as far as she took it, and no further.

Of the three verticals, only chip design has a concrete artifact behind it. Friar said OpenAI used its own frontier models to help design Jalapeno, the inference chip it has been co-developing with Broadcom since October 2025, and that the design cycle reached tape-out -- the point where a chip's design is finalized and sent to a fab to be manufactured -- in under nine months. That's among the fastest publicly reported cycles for a chip of its class, and it sits on top of, not instead of, what OpenAI's own blog already disclosed in August: Jalapeno's early benchmark results against Nvidia's Blackwell systems, run on a public third-party suite and reported here at the time. What's new in Friar's remarks isn't the chip's performance -- it's the claim that OpenAI's own models meaningfully sped up the human engineering work of designing it, a claim neither company has isolated with a before/after comparison against a design cycle done without AI assistance.

Life sciences and financial services got no equivalent specifics -- no named pilot customer, no product, no contract, no figure resembling Jalapeno's nine-month number. The closest concrete anchor either vertical has is OpenAI's own August integration of ChatGPT with Epic's electronic health record system, reported separately and not something Friar's Sept. 8 remarks connected back to this specific pitch. That gap matters for how much weight to put on the framing: 'chip design, life sciences and financial services' reads as three parallel bets, but only one of the three currently has a shipped artifact attached to it.

The clearest number in Friar's remarks was Luna's price. On July 30, OpenAI cut GPT-5.6 Luna from $1.00 per million input tokens and $6.00 per million output tokens down to $0.20 and $1.20 -- a combined rate of $1.40 versus $7.00 before, an 80% cut -- and Friar said usage rose roughly 10x afterward. The cut landed in the middle of a price war that Chinese open-weight labs mostly started, undercutting frontier US labs on a pure cost-per-token basis for much of 2026.

"If you're deploying Luna and compare that to [Z.ai's] GLM 5.3, for example, on a cloud layer, we are cheaper," Friar said -- a specific, checkable comparison, and by VentureBeat's own reporting, an accurate one: GLM 5.3's standard API pricing runs $1.40 per million input tokens and $4.40 per million output tokens, a combined $5.80, more than four times Luna's post-cut rate. It is also a narrower claim than the framing suggests, and the gap between the two is exactly the kind of thing a wire rewrite of the Reuters story wouldn't flag.

That tier mismatch is the whole reconciliation: DeepSeek's Flash model runs about $0.42 combined per million tokens and Xiaomi's MiMo-V2.5 Flash about $0.40 -- both genuinely cheaper than Luna, but budget-tier models Friar wasn't comparing Luna to. Her comparison and VentureBeat's market survey are not actually in conflict; they're answering different questions.

The revenue numbers are less contestable, because they describe OpenAI's own business rather than a competitor's pricing page. Enterprise customer revenue climbed 32% between June and July, outpacing the company's 20% overall annualized growth rate over the same stretch, and enterprise and consumer revenue reached roughly even parity by mid-year -- ahead of OpenAI's own year-end target for that balance. Codex, OpenAI's coding tool, has reached 25 million users. Put together, the pitch is coherent even where it's unproven: cut prices to win volume against cheaper open-weight rivals on the low end, then recover margin on the high end by selling outcomes -- a completed chip design, a processed diagnosis, a reviewed filing -- to enterprises who no longer want to pay for tokens they can't tie to a result.

Outcome-based pricing is itself a bet that enterprise AI buyers have changed what they'll sign. A year of usage-based API bills that scaled with adoption but not always with measurable value has left procurement teams asking a harder question before renewal: what did this actually get us? Tying price to a completed result -- rather than a volume of tokens, seats, or API calls -- shifts that risk from the buyer to the seller, which is a much easier pitch to make from a position like OpenAI's, with enough enterprise revenue growth to absorb the experiment, than it would be for a smaller lab still trying to prove usage-based pricing works at all.

A CFO's conference remarks are checkable in the parts that describe her own company's ledger, and unverifiable in the parts that describe what her company hopes to sell next year.

None of this is a launch. It's a CFO narrating a strategy at a bank's investor conference, which is exactly the kind of statement that tends to get repeated as settled fact once enough outlets run the same wire copy. The concrete parts -- the price cut, the enterprise revenue split, the taped-out chip -- are checkable and, so far, check out. The parts that aren't yet concrete -- what an outcome-based contract actually looks like, and what OpenAI has actually shipped in life sciences or financial services beyond one hospital-records integration -- are still just the sentence a CFO said out loud, not a product a customer can buy.