On September 12, 2026, the AI industry reached a rare moment of apparent consensus. Anthropic CEO Dario Amodei published We Must Pace the Frontier, a framework for coordinated industry-wide deceleration. Within hours, Elon Musk endorsed it on X: “Dario is right.” Sam Altman and Demis Hassabis followed. Nine days later, xAI shipped Grok 4.7 — and the pricing alone told you the consensus was already dead.
Grok 4.7 costs $2 per million input tokens and $6 per million output for prompts under 200,000 tokens. Cached input reads cost $0.50. Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra both sit at $10 and $50. The output-token discount is 80 percent. A fast variant, restricted to Cursor and Grok Build, runs at $4 and $12. The message is not subtle: xAI chose to ship now, at this price point, against the very coordination narrative its CEO publicly endorsed.
The timing makes the move legible as a competitive maneuver. The antitrust lawsuit filed September 18 in the Northern District of California — Buist et al. v. Anthropic PBC et al. — alleges that the coordinated public endorsements of an AI slowdown amount to an output-restricting cartel under Section 1 of the Sherman Act. Senators Josh Hawley and Ted Cruz had already blocked an antitrust exemption for AI companies in the NDAA days before the filing. Cohere CEO Aidan Gomez called the push for a FINRA-style safety body “a cartel by any other name.” Against that backdrop, Grok 4.7 reads as xAI betting that price competition — not coordination — defines the next phase.
For builders evaluating models in production, the pricing gap reshapes the evaluation tax calculus. The September model flood — five frontier-class releases in ten days — already compounded the benchmarking, migration, and compliance costs that labs never see. Grok 4.7 adds a sixth model with a 500,000-token context window, four configurable reasoning modes, and pricing that undercuts the field by an order of magnitude. For startups running high-frequency agentic workflows where output-token costs dominate, the switch economics are significant.
The counter-argument to the “no safety” critique has substance. Grok 4.5, launched July 9, shipped with no model card and no safety benchmarks of any kind. Grok 4.7 arrives with an entirely redesigned safeguard stack: a detailed model card, jailbreak metrics across standard, strong, and long-horizon attack suites, and an invite-only red-team program for cybersecurity partners. Standard jailbreak compliance fell from 0.73 percent in version 4.5 to 0.01 percent in 4.7. On LatchBio’s biosafety benchmark, the model scored 62.4 percent — top of the reported field. HackerBench v0.3 shows only 3.3 percent of risky dual-use cyber prompts allowed through. These are vendor-reported figures, and independent verification has not yet arrived, but the disclosure posture is categorically different from the July release.
The ~2.1 trillion parameter count circulating in builder communities has not been officially confirmed by xAI. What is confirmed: the model is available on Cursor, Grok Build, the xAI API, OpenRouter, Vercel, and Cloudflare — a distribution breadth that makes it immediately accessible to the developer ecosystem.
The Anthropic IPO and the pacing framework coverage both identified the same structural tension: the incentive to coordinate on safety runs directly against the incentive to compete on capability. Grok 4.7 makes that tension concrete. Nine days after endorsing the slowdown, Musk’s company shipped the most aggressively priced frontier model of the cycle. The question is no longer whether the labs can coordinate — it is whether any of them can afford not to undercut when the pricing signal is this clear.