Megan Skiendel: Long week — I've been staring at two press releases that fundamentally contradict each other and I genuinely cannot tell which one is the lie.
David Sterling: OpenAI and Forbes, I assume.
Megan Skiendel: Exactly those two. Listen — July 9, GPT-5.6 Luna hits general availability. July 30, Sam Altman announces an 80% price cut. One dollar per million input tokens becomes twenty cents. Six dollars per million output becomes a dollar twenty. And OpenAI's framing is: this is the efficiency dividend, we built something that costs less to run, we're passing it on. Forbes, August 1, says this might be the start of unsustainable pricing dynamics industry-wide.
David Sterling: Both cannot be fully correct. That's actually the most precise thing Forbes said.
Megan Skiendel: Three weeks between launch and an 80% cut — I mean, honestly, when has that looked like confidence?
David Sterling: It hasn't. The question is what specifically broke. Was it adoption numbers on Luna? Was it a call from a hyperscaler? That's actually what I want to get at — the arithmetic on an 80% cut is brutal unless volume is up by a factor of four or cost is genuinely down 80%.
Megan Skiendel: And OpenAI says it's the cost. I want to know who actually made that call inside the room.
David Sterling: That's where I need you.
Megan Skiendel: Sam said it himself — GPT-5.6 Sol rewrote its own production GPU kernels inside Codex. Autonomously. That's the efficiency story. And look, I believe the engineering happened. What I don't believe is that it happened in the three weeks between Luna's launch and the price cut.
David Sterling: That's the tell. If the kernel rewrite was real before July 9, why didn't Luna launch at twenty cents?
Megan Skiendel: Exactly. That's the gap nobody's closing.
David Sterling: Think of it this way — imagine a grocery store prices milk at a dollar, then three weeks later drops it to twenty cents and says they found a more efficient cow. Maybe they did. But you'd also notice a new grocery just opened across the street. That's Kimi K3. That's DeepSeek V4. The cow story might even be true — and it still doesn't explain the timing.
Megan Skiendel: And the cow story is conveniently one-time. A kernel rewrite isn't a compounding curve — it's a step change you take once.
David Sterling: Which means — wait, actually this is the important part — the arithmetic still doesn't close. An 80% revenue cut on Luna requires either an 80% cost reduction or volume up four-hundred-plus percent. OpenAI has shown neither number publicly. And then look at Terra — only cut 20%. Sol, no cut at all. If this were a uniform efficiency story, why does it apply almost exclusively to Luna?
Megan Skiendel: Because Luna is the layer Kimi K3 and DeepSeek V4 directly threaten. The high-volume, commodity-inference layer. Sol's premium market isn't gone yet — so you hold Sol, you defend Luna.
David Sterling: That's the signal. This isn't an efficiency dividend. It's a subsidized land-grab — price Luna below what makes comfortable sense now, because the alternative is losing that segment entirely to open-weight in six months. The efficiency story is real but partial. The competitive pressure is the actual load-bearing reason.
Megan Skiendel: But that's where the circulating take goes completely wrong — everyone's saying OpenAI fired a price war across the whole market. They didn't. They started a price war in one lane while quietly raising prices in the other.
Megan Skiendel: Sol. No cut. Instead it gets a new Fast mode — up to two-and-a-half times faster, at double the standard rate. That's not a concession. That's a premium upsell inside the same announcement everyone's reading as a retreat.
David Sterling: Double the rate. At the same moment Luna drops 80%. That's — I mean, that's not one strategy. That's two simultaneous strategies wearing the same press release.
Megan Skiendel: And Anthropic didn't flinch at all. Claude Opus 5 priced at exact parity with Opus 4.8 — not a dollar lower. If you're scared of what OpenAI is doing, you match. Anthropic didn't move.
David Sterling: So Anthropic holding Opus 5 flat is actually the signal about where premium margin is defensible.
Megan Skiendel: That's exactly the tell. And Google goes the opposite direction — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, both targeting low inference costs. So you've got three labs and none of them are doing the same thing. The 'price war' framing flattens a structure that is — honestly, it's pretty deliberate segmentation.
David Sterling: Put a concrete number on it. Developer running ten million document-parsing calls a day on Luna — her monthly bill just dropped four-fifths. The enterprise team one floor up using Sol for real-time contract analysis is paying the same or more with Fast mode. That asymmetry is structural, not accidental.
Megan Skiendel: And what that asymmetry doesn't solve — we'll get to this — is whether Luna's price can hold once enterprises realize Kimi K3 doesn't need to recoup OpenAI's R&D at all.
David Sterling: And that's actually the structural problem — Kimi K3 and DeepSeek V4 don't have a centralized R&D bill to recover through API margins. OpenAI does. Every token Luna serves has to absorb some fraction of what it cost to build the thing. Kimi K3 doesn't carry that. The cost floor is just different.
Megan Skiendel: Which means the price cut buys time, not parity.
David Sterling: Correct. And here's the point — picture a Fortune 500 procurement lead. She's got Kimi K3 running on her company's own infra, she's got Luna at twenty cents per million. She doesn't actually have to migrate. She just has to show Sam's team the Kimi K3 benchmark printout and say 'match it or we move.' The threat does the negotiating.
Megan Skiendel: The leverage is real whether or not she ever runs a single Kimi K3 token in production.
David Sterling: That's the watch item. Not 'did enterprises migrate?' — that's the wrong question. Watch whether any major enterprise actually signs a self-hosted inference contract in the next two quarters. That's the signal. Because if nobody migrates, OpenAI's margin problem is manageable. If even one hyperscaler moves meaningful Luna volume to self-hosted DeepSeek V4, the precedent is — I mean, that changes the negotiating table for every renewal.
Megan Skiendel: And Sarah Friar drops the one billion active user number July 31 — one day after the cut. That's not coincidence.
David Sterling: Wait — one day after?
Megan Skiendel: July 31 blog post. One billion users, two million businesses. The sequencing is deliberate — you lead with the price cut as the headline, then twenty-four hours later you reframe it as a scale story. 'We're not bleeding margin, we're feeding a billion people.' It's narrative management before the Forbes 'race to the bottom' story had time to harden. And it worked — a lot of coverage picked up the user number and buried the margin question.
David Sterling: Frankly, the one billion figure doesn't touch Luna economics directly — those are consumer users, not API inference buyers. But it resets the emotional frame. And that's probably the point.
Megan Skiendel: And that's the part I can't get out of my head. If API inference on Luna is now trending toward utility pricing — twenty cents per million, and falling — where does OpenAI find the margin to fund whatever comes after GPT-5.6? Sol can't carry that alone. Sol is one model for one segment.
David Sterling: The math doesn't close publicly. Volume at twenty cents per million input tokens on Luna would need to be enormous — I mean genuinely enormous — to fund the next training run. And nobody has shown that arithmetic. Not OpenAI, not Sarah Friar's blog post, nobody.
Megan Skiendel: Which is the question nobody's actually sitting with. Anthropic held Claude Opus 5 flat. Google is racing to the bottom on Flash-Lite. OpenAI just told the market that managed inference might be a utility. If that's the frame that sticks — if that's what enterprises believe — the margins that built GPT-5.6 don't exist to build what's next.
David Sterling: And nobody's answered it.
Megan Skiendel: Nobody's answered it. I keep waiting for someone at OpenAI to show the actual unit economics and — it just hasn't come.
David Sterling: Frankly, that silence is its own data point. Good conversation.
