Grok 4.7 keeps Grok 4.6 pricing and speed; demos raise the task-cost question

Why it matters

Official claim — SpaceXAI says Grok 4.7 is a notable improvement over Grok 4.6 at the same price and speed; its official launch page lists $2 per million input tokens and $6 per million output tokens below 200k prompt tokens, and Elon Musk describes it as third behind Anthropic and OpenAI in agentic coding but significantly faster and lower-cost. Community evidence — collected X demos show positive one-off results for an open-world city, Blender, and an interactive SVG, while another author reports a poor Three.js rocket result at $1.59 versus about $6 for Kimi and unusually high token use in 3D/frontend work. These are attributed anecdotes, not comparable benchmarks: prompts, harnesses, reasoning settings, retries, environments, and selection are not controlled. Independent conclusion — low token price does not imply low task cost. Reproduce with a frozen public task set and identical harness, permissions, runtime, model settings, and acceptance tests; run at least five randomized trials per task and model; report first-pass success, attempts and retries, input/output/reasoning/cache tokens, wall time, tool and infrastructure spend, human rework minutes, final success, and total cost per accepted task. Publish prompts, container or lockfiles, seeds, and raw traces, and compare the community examples only after rerunning them under this protocol.