Anthropic shipped Claude Opus 4.7 with a 1M-token context window, which is the polite way of saying they caught up to Gemini on the one spec that fits in a tweet. Triples the previous Opus ceiling. The genuinely interesting move, the one operators will notice before the marketing team gets done celebrating, is that long-context pricing stays in the same band as 200K. The marginal cost curve for agents that kept hitting the ceiling just got a haircut.
The model card flexes on long-horizon tool-use evals and code-repo retrieval at the deep end of context, which is the polite way of saying “we got better at reading your entire monorepo without losing the plot.” Sub-100K performance is reported as flat versus 4.6, which is actually the useful detail. The practical question is whether your shop swaps the default model or just the long-document pipeline. Anthropic is quietly telling you: the second one.
Worth tracking, because nobody else will: API price-per-effective-token versus chunking-plus-retrieval. Long context isn’t free, it’s just billed differently. The right answer depends on whether your workload looks like a needle in a haystack or a haystack with a plot.