Analysis
I've watched three frontier labs cut prices in the same five-day window, and my first reaction wasn't "the AI race is heating up" -- it was that nobody in this specific fight is proving the economics work yet. OpenAI's GPT-6 Sol dropped to $2/$10 per million tokens. Anthropic's Claude Opus 5.5 followed at $4/$20 list, with cache reads down 60%. xAI's Grok 4.7 undercut both on cached pricing. Three labs, five days, and every one of them is telling the market the same thing: we can't win on capability alone right now, so we're competing on cost.
That's good news if you're building a startup on top of any of these APIs -- your cost of goods sold just fell again, for the third or fourth time this year, and it'll probably fall again before Q4 ends. It's a much harder story if you're an LP in one of these labs' cap tables, because price wars in infrastructure businesses compress margins for everyone simultaneously, and none of these three companies has said publicly what its actual gross margin looks like at these new price points.
“Anthropic cut cache reads 60%, deeper than its headline list-price cut, and cache reads are disproportionately where agentic and coding workloads spend their tokens.”
The tell I'd watch isn't the sticker price, it's the cache pricing specifically. Anthropic cut cache reads 60%, deeper than its headline list-price cut, and cache reads are disproportionately where agentic and coding workloads spend their tokens. That's a company optimizing to retain developers building agents on Claude specifically, not a broad capability play. Grok 4.7's tiered pricing does the same thing with its below- and above-200K-token structure. Sol's flat halving is the least targeted of the three, which either means OpenAI has more margin to give up, or means Sol's economics were the least defensible to begin with.
Room for disagreement: the optimistic read is that this is exactly what a maturing, competitive market is supposed to look like -- falling prices for equivalent or better capability is the textbook outcome of real competition, not a sign of distress, and the developers building on these APIs are the ones who actually win. Compute costs have also fallen industry-wide on the hardware side, so some of this price movement may simply be labs passing through real cost reductions rather than margin compression. I don't think that fully explains three coordinated-looking cuts in five days, but it's the more benign explanation and it deserves to be stated plainly rather than dismissed.
What I'd actually diligence if I were underwriting any of these three companies right now: ask for gross margin by product tier, not blended company-wide margin, because a $2/$10 flagship price backed by a subsidized loss-leader strategy looks identical from the outside to a genuinely efficient $2/$10 price -- and only one of those is a business model you can compound.
