Meta walks away from open weights as the vendor race accelerates

Meta has retired Llama for a proprietary cloud-only model just as xAI undercuts the field on price and OpenAI says GPT-5.5 is its strongest launch yet — the open era is closing fast.

Three vendor moves this week point to a market in flux: Meta has retired Llama in favor of a proprietary model, xAI shipped Grok 4.3 at a price aggressive enough to reshape comparison shopping, and Musk admitted under oath to training on competitors' outputs. The open-weights branding that defined the prior cycle is fraying just as enterprise spending on agentic tools accelerates.

The clearest signal comes from Meta. According to The New Stack, the company has pivoted from Llama's downloadable open weights to Muse Spark — cloud-only, no self-hosting, and only in private API preview as of April 2026. For teams that built infrastructure assuming Llama would keep getting better and stay free to run, that assumption is now broken. As AI Advances put it, Muse Spark sits behind a paywall "right next to the OpenAIs and Anthropics" Meta once positioned itself against.

Pricing pressure is arriving from a different direction. VentureBeat reports xAI shipped Grok 4.3 with a 1,000,000-token context window (per OpenRouter and NanoGPT listings, May 2026) at pricing aimed squarely at the cost-per-intelligence chart. The credibility cost arrived the same week: in a California federal court on April 30, 2026, TechCrunch reports Musk testified that xAI used distillation on OpenAI models to train Grok, calling it standard industry practice.

Meanwhile, the incumbents are consolidating. OpenAI says GPT-5.5, launched in late April 2026, is its strongest launch yet, with API revenue growing more than 2x faster than any prior release and Codex doubling revenue in under 7 days. Anthropic took a different tack, publishing analysis of 1 million conversations to identify where Claude slips into sycophancy — Opus 4.7 (April 2026) cuts the rate by half versus Opus 4.6 on relationship guidance, with Mythos Preview halving that again. Self-critique as marketing is itself a competitive move.

For builders, the open-weights playbook just lost its anchor tenant; cost-sensitive bulk inference and on-prem requirements need a Plan B. For buyers, the price war is real but so is the stability question — Grok's distillation admission and Meta's pivot are both procurement risks. For creators, the field is wider than ever, but the next year will likely reward portability over loyalty.

The bottom line: Treat your model vendor as an infrastructure dependency, not a research partner, and keep your prompts, evals, and inference paths portable across at least two providers.

Sources