OpenAI cut its flagship GPT-6 Sol to $2 per million tokens, while Anthropic slashed Opus 5.5's cached read fees by 60% on the same day — two leading vendors used a price war to declare: the capability race is taking a pause. All three new flagships are fully aligned on a 1M token context window (tokens are the minimum billing unit for LLM services).

What this is

OpenAI went with a "lead plus support" pairing: GPT-6 Sol targets daily production development — 1M context, 128K token output, at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna focuses on preprocessing and large-scale retrieval at $0.10 input and $0.50 output per million tokens, designed first to filter logs and distill structure.

Anthropic's Claude Opus 5.5 targets long-horizon coding and complex architecture work. Same 1M context, same 128K token output ceiling. The core upgrade is "adaptive thinking" — no more manually setting a thinking budget; the model decides how long to reason on its own. Prompt caching (storing frequently used prompts on the model side to reduce repeated-request costs) read fees dropped by 60%, making a single complex task roughly 40% cheaper than the previous-generation Opus in real spending.

The three models are heavily converged on specs: same 1M context, same 128K token output ceiling. Differences sit mostly in pricing, reasoning style, and toolchain fit.

Industry view

We note two things. First, the signal of "capability convergence" across flagships is clear — all ship 1M context and adaptive reasoning; the capability gap is no longer the main battlefield. Second, "AI gateways" (middleware that unifies multi-model access and local scheduling, analogous to a traditional IT API gateway) are gaining traction: developers no longer wire up keys and protocols per tool, but route through a local gateway, pushing the "which model to use" decision down to the config layer.

But we also want to push back. Model companies were publicly calling for "slowing down capability improvements" just two weeks ago — and within half a month they are shipping new versions aggressively and going straight into a price war. Commercial pressure has overridden the safety consensus. The fiercer the price war, the more R&D tilts toward "cost reduction" — good for consumers in the short term, potentially bad news for the industry's ceiling over the long term.

Another underestimated risk: multi-model scheduling looks like a cost saver, but key management, protocol adaptation, and bill consolidation are costs of their own. For small and mid-sized teams, "all of the above" may not be cheaper than "betting on one vendor."

Impact on regular people

  • For enterprise IT: unit cost of AI services keeps falling, but the budget structure shifts from "buying model calls" toward "buying integration and governance" — multi-model routing becomes a new problem to tackle.
  • For individual professionals: heavy AI users see per-task cost drop to one-third to one-half of before, but toolchain management gets more complex, with daily decisions about "which model to use."
  • For the consumer market: upstream price cuts will likely pass downstream — over the coming months, consumer-facing AI products may see a wave of price cuts or feature unlocks.