Claude Opus 4.7 and the End of the Frontier Cost War
Opus 4.7 ships at the same $5 input price as 4.6, with the same 1M context window, and a 1503 Elo that tops the Arena. That nothing about pricing moved is precisely the story.
Claude Opus 4.7 shipped today with the same $5 input pricing as 4.6 and the same 1M context window. The only thing that changed is the model — and that is precisely the story. For two years, every frontier release has landed with a pricing narrative attached: cheaper tier, new batch discount, smaller context at lower cost. 4.7 is the first flagship in recent memory where the price sheet is a copy-paste from the last version. That shifts the evaluation question back to capability, which is exactly where Anthropic wants it.
On the current LMArena board, Opus 4.7 sits at 1503 — ten points above Gemini 3.1 Pro and nineteen above GPT-5.4 High. Internal benchmarks emphasize planning depth and self-correction on long-running coding tasks, and the "Coding King" tag in helloai's directory reflects what developers have been reporting in production: fewer abandoned branches, fewer wrong-file edits, and meaningfully better recovery when a plan doesn't survive contact with the repo. None of that shows up cleanly in a static eval, which is why the Arena Elo matters more than the incremental MMLU bump.
The more interesting question is what 4.7 does inside an Advisor Strategy loop. The executor-advisor split published last month showed Haiku-plus-Opus-advisor more than doubling BrowseComp performance at 85% less cost than Sonnet running solo. Swap the 4.6 advisor for 4.7 and the math compounds — the advisor's marginal judgment lift gets amplified by every fork it touches across the trajectory. For teams that restructured around that pattern in the last thirty days, 4.7 isn't a model upgrade. It is a system-wide capability lift without a single line of application code changing.
The counterargument is version fatigue, and it is fair. 4.6 dropped in price two weeks ago. The Mythos story broke a month before that. Developers have been re-running evaluation suites more or less continuously since February. Shipping 4.7 into that cadence risks the reasonable response of deferring the evaluation until Q3. That would be a mistake on pure-coding workloads, where the delta over 4.6 appears large enough to justify a re-run, but it is probably the right call for chat, extraction, or any workload where 4.6 was already saturating your rubric.
What this release actually signals is the end of the cost-war phase of the frontier. The pricing has converged: Opus, Gemini Pro, and GPT-5.4 sit within a few dollars of each other on input. The context windows are all a million tokens or better. The batch discounts are roughly equivalent. From here the differentiation is capability and the shape of a model's failure modes under agentic workloads. 4.7 is the first release that leans entirely on that and doesn't hide behind a price cut. Expect the next year of frontier releases to look the same.