Same day, two labs cut prices — the real signal is where the margin is coming from

On September 22, Anthropic cut list pricing on Claude Opus 5.5 roughly 20 percent (to $4/$20 per million input/output tokens, from $5/$25), with the company and early enterprise users citing additional effective savings from faster output and fewer tokens needed per task. The same day, OpenAI released two cost-optimized GPT-6 variants, Sol and Luna, priced at roughly half their GPT-5.6-era equivalents and explicitly pitched at high-volume, routine workloads rather than frontier reasoning tasks. Two labs making serving-cost moves on the same day isn't necessarily coordinated, but it is a clean marker of where competitive pressure has moved: toward how cheaply a model can be run, not just how capable it is.

The more revealing number sits underneath the price cuts rather than in them. SemiAnalysis has reported that Anthropic's inference gross margins rose from roughly 38 percent to over 70 percent even as the company cut prices by a factor of three across recent model generations — a combination that only works if serving efficiency is improving faster than prices are falling. That pattern, software and systems efficiency outpacing sticker-price cuts, is the more durable story than any single discount announcement, because it describes a capability (cheaper inference per unit of model quality) rather than a one-time promotional move.

Why the industry's attention is shifting

Part of the backdrop is a live debate about whether pretraining scale still buys what it used to. Ilya Sutskever has argued publicly, including in a widely discussed November 2025 conversation, that pretraining's returns are flattening as readily available training data is largely exhausted — a claim that doesn't by itself prove labs are reallocating R&D budgets, but does explain why serving-time efficiency (squeezing more useful output per dollar from the models already built) is attracting investment and attention that might previously have gone toward simply training a bigger model. Google DeepMind's Demis Hassabis has made a related point from the demand side: falling inference costs aren't making inference free, because reasoning and agentic workloads are absorbing the savings about as fast as they appear, which keeps serving efficiency commercially relevant rather than a solved problem.

None of this amounts to a quantified industry-wide reallocation of R&D spend, and claims along those lines should be treated skeptically absent real disclosure from the labs themselves. What's verifiable is narrower and still notable: on at least one day in September, two of the largest frontier labs both chose to compete on serving cost rather than capability alone, and the lab with public margin data available is reportedly getting cheaper to run faster than it's cutting prices.