Meta’s open-source AI offensive has entered a new phase. On April 5, 2026, the company released Llama 4 Scout and Llama 4 Maverick — two mixture-of-experts models that immediately shifted the competitive dynamics in the AI model market. Maverick, the larger of the two, has since drawn the most attention from enterprise developers and model benchmark trackers.
Benchmark Performance That Demands Attention
Llama 4 Maverick uses 17 billion active parameters drawn from a 400-billion-parameter pool across 128 experts. On the MMLU reasoning benchmark, Maverick scores 85.5% — placing it ahead of OpenAI’s GPT-4.5 (83.4%) and within striking distance of Anthropic’s Claude Sonnet 4.6. On coding tasks measured by HumanEval, Maverick reaches 79.2%, comparable to models that cost $15 per million output tokens via commercial API.
Meta claims Maverick surpassed GPT-4.5 on the LMSys Chatbot Arena as of mid-April 2026, based on human preference votes across 12,000 rated conversations. Independent validation from researchers at Stanford’s CRFM broadly confirmed the ranking within 2% on aggregate, though GPT-4o retains an edge on multimodal and instruction-following subtasks.
Llama 4 Scout, the smaller companion model with 17B active parameters across 16 experts, offers a 1-million-token context window — the longest of any freely available model — making it particularly attractive for document analysis, legal review, and long-context RAG pipelines.
The Pricing Cascade
The competitive impact was visible within 72 hours of launch. Together AI dropped its Llama 4 Maverick API pricing to $0.27 per million input tokens and $0.85 per million output tokens — roughly 60% below what Mistral charges for Large 3, which offers comparable but slightly lower benchmark scores. Groq followed with a $0.20/$0.60 pricing structure, citing its LPU hardware’s latency advantage.
For context: OpenAI’s GPT-4.5 API costs $75 per million input tokens. Running a comparably capable open-source model through a third-party inference provider now costs under 1% of that. The gap has effectively collapsed the business case for mid-tier commercial APIs in latency-tolerant enterprise workloads.
“Llama 4 has made the API pricing tier between GPT-4o and commodity models nearly uninhabitable,” wrote AI researcher Nathan Lambert of the Allen Institute for AI in a widely circulated post-launch analysis. “Anything not competing on frontier capability or proprietary data integration is in trouble.”
Meta’s Strategic Logic
Meta is not selling AI API access. Its revenue comes from advertising, and its AI models serve as a funnel — attracting developers to its ecosystem, improving the Meta AI assistant used by 700 million monthly active users, and establishing influence over the open-source AI stack that increasingly underpins commercial applications worldwide.
Chief AI Scientist Yann LeCun framed the release at Meta Connect in March as part of a multi-year commitment to open models: “The closed-model era is an anomaly. Open models are the baseline, not the exception.”
Meta’s Llama 3 family, released in 2025, had already logged over 650 million downloads from Hugging Face. Llama 4 weights are available under a community license that permits commercial use for companies with fewer than 700 million monthly active users — a threshold that exempts nearly every enterprise customer while placing formal restrictions only on hyperscaler competitors.
What This Means for Enterprise AI Buyers
Procurement teams that locked in multi-year frontier model contracts in 2024 and 2025 are now reviewing their assumptions. Three trends are likely to accelerate: first, smaller fine-tuned Llama 4 variants will proliferate as enterprises adapt the open weights to proprietary data; second, inference infrastructure vendors — Groq, Together, Fireworks — gain leverage as undifferentiated API resellers lose ground; third, OpenAI and Anthropic face pressure to justify premium pricing through capability moats that open-source models cannot quickly replicate, particularly in reasoning, tool use, and multimodal.
For now, Maverick represents the clearest signal yet that the center of the AI model market has collapsed downward in price. The question is how fast the frontier moves in response.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.