This week the biggest model labs competed on price rather than benchmarks. Anthropic's Opus 5.5 landed 40% cheaper than Opus 5; a day later OpenAI's GPT-6 Sol and Luna arrived at half their previous API rates. Both framed the releases around efficiency — fewer tokens, better caching — not a capability leap. The move is notable partly for who's making it: these are the same two companies that recently pushed the industry to slow the race down. If you buy inference by the million tokens, the direction of travel matters more than any single launch — frontier-tier quality keeps getting cheaper, and the floor keeps dropping. Worth factoring into next quarter's budget if you run agents at scale.