- MegaOne’s index tracks 164 open-weight and 212 proprietary models with published pricing — open-weight is now a third of the priced market.
- Open-weight and proprietary models meet at the same $0.03-per-million-token price floor.
- Databricks’ internal benchmark across 3,000 engineers found the open-weight GLM 5.2 handling its highest-difficulty coding tasks at lower total cost than proprietary frontier models.
- Meta’s Llama scores 9 out of 10 on MegaOne’s Engine Score — tied with Claude Code and the major API platforms.
The debate over whether open-weight AI models can compete with proprietary ones is settled on cost, and 2026 is the year the data made it undeniable. MegaOne’s index now tracks 164 open-weight models with published pricing against 212 proprietary ones — open-weight has grown from a rounding error to roughly a third of the priced market — and at the price floor, the two are indistinguishable: both bottom out at $0.03 per million output tokens.
Price parity has arrived at the floor
The cheapest open-weight models — Meta’s Llama Guard 3 8B, Mistral Nemo — are priced at $0.03 per million output tokens. The cheapest proprietary models sit at exactly the same $0.03. For high-volume, cost-sensitive work — classification, extraction, moderation, routing — there is no longer a price argument for choosing proprietary. The floor is shared, and open-weight adds the option to self-host and keep both the data and the model in-house.
The capability gap closed on real tasks, not just benchmarks
The stronger evidence is operational. Databricks ran an internal benchmark across the real coding tasks its 3,000 engineers perform and reported that open models — the open-weight GLM 5.2 in particular — handled even the highest-difficulty coding at a lower total cost than proprietary models from Anthropic and OpenAI. That is a cost-per-completed-task finding from a company managing its own multimillion-dollar AI bill, not a leaderboard score. It is why Databricks, now valued at $188 billion, has become one of the most visible enterprise champions of open-weight models.
Where proprietary still wins
Open-weight has not won everywhere. The premium reasoning tier — the models built for the hardest multi-step problems — remains proprietary and expensive, running $120 to $600 per million output tokens on MegaOne’s index. For a legal-analysis or financial-modeling workload where a single wrong answer is costly, that premium is defensible. The pattern is clear: proprietary labs are ceding the commodity tier and defending the frontier, where reasoning depth and reliability still command a price.
The MegaOne Engine Score view
On capability, MegaOne’s Engine Score — an independent 1-to-10 rating across 155 tracked tools — puts Meta’s Llama at 9 out of 10, tied with Claude Code, the OpenAI and Anthropic API platforms, and other top-tier tools. An open-weight model sitting in the same score band as the leading proprietary platforms would have been implausible eighteen months ago. Combined with the shared price floor, it means the real question for most teams is no longer “open or proprietary” but “which tier does this specific workload need” — a decision the Engine Scores and the Price Index are built to answer.
What to do about it
Audit your workloads by stakes, not by habit. Route bulk, low-stakes, high-volume traffic to the sub-dollar open-weight tier; keep the frontier proprietary reasoning models for the small share of tasks where being wrong is expensive. Teams still defaulting every call to a frontier proprietary model are, in most cases, overpaying by an order of magnitude for capability their workload never uses.
Related Reading
- The MegaOne LLM Price Index: What AI Models Actually Cost in 2026
- The MegaOne LLM Price Index: What AI Models Actually Cost in 2026
- China’s Kimi K3 Just Took #1 on the Coding Leaderboard — 2.8 Trillion Parameters, Free, and It Beats Claude in 76% of Matchups
- White House Says Moonshot Used Banned Nvidia Chips to Build Kimi K3
- Bloomberg Intelligence: China’s AI Model Gap Narrowed to 6% in June
