News · 2026-07-17

Mozilla Report: Open-Weight Models Now Route the Majority of AI Tokens

Open-weight AI models now route the majority of production tokens on OpenRouter, a major model-routing service, according to Mozilla's first 'State of Open Source AI' report, published July 2026. The share climbed from negligible two years ago to about a third in late 2025 to a majority by mid-2026, and the five highest-volume models on the platform are all open weights. The finding is the clearest evidence yet that open models are no longer a hobbyist alternative but the workhorse of real production traffic.

Key facts

• The five highest-volume OpenRouter models are all open weights, led by DeepSeek V4 Flash at 18.4 trillion tokens; the first closed model, Claude Opus 4.7, appears sixth.
• Inference cost for a GPT-4-equivalent model fell about 50x in 36 months, from $20 to $0.40 per million tokens.
• The report was published by Mozilla, drawing on data from OpenRouter, Stanford HAI, Epoch AI and a SlashData developer survey.
• Only 51% of teams using open models reach production, versus 63% for closed models -- a tooling gap, not a capability gap.

The most important nuance is that this is a token-volume lead, not a request-count lead. By number of requests, closed US providers still dominate; the open lead is concentrated in coding and agentic workloads that burn huge numbers of tokens. Mozilla is careful about this: 'By request count, closed US providers still lead -- the open lead is a token-volume lead.' The picture is a redistribution of where the heaviest work runs, not a wholesale collapse of the closed model business.

On quality, the report describes a 'jagged frontier.' Averaged across Chatbot Arena, the open-vs-closed gap shrank from about 8% in early 2024 to roughly 3.3% by March 2026, with DeepSeek-R1 briefly matching the top US model in early 2025. But that average hides real structure: open models are at or near parity on coding, instruction-following and general knowledge, while a meaningful gap persists on reasoning, long-context retrieval and agentic tasks. The report's blunt summary: 'The question is no longer whether open models are good enough. It's what you need for your workload.'

Underneath both trends is a price collapse. The cost to serve a GPT-4-class model fell roughly 50x in three years -- 'faster than dotcom-era bandwidth or PC-compute price curves,' the report says, citing Stanford's AI Index and Epoch AI. When the same capability gets 50 times cheaper, the economics of paying a premium for a closed API get harder to defend, especially for high-volume tasks. The report cites a concrete case: a company that cancelled most of its Claude Code licenses by June 30, 2026 after token billing consumed its annual AI budget in months, while testing a self-hosted open model for its heaviest workload.

The report's sharpest self-criticism is where open still loses: getting to production. Roughly 79% of developers adding AI features use open models and 71% use closed ones (many use both), but only 51% of open-model teams actually ship to production versus 63% for closed. Mozilla attributes the gap to 'operational tooling and trust, not model capability.' That leads to the report's central thesis, which it calls 'the harness is the new frontier': above the model now sits the agentic harness -- the orchestration loop, tools, memory, sandboxes and permission model -- and that is where production difficulty, and the next open-vs-closed fight, has moved. Closed labs 'just proved it by pulling the harness in-house,' a reference to products like Claude Code and Codex.

Why it matters: this report is the receipts behind the whole 'open source is winning' narrative that the Kimi K3 market shock dramatized on the same day. It also arrives with real revenue numbers -- DeepSeek at roughly $220 million in annual recurring revenue and a $50-billion-plus valuation, Mistral scaling to about $400 million in a year -- so this is not just enthusiasm. The honest caveat is attribution: these are Mozilla's figures, assembled from OpenRouter, Stanford HAI, Epoch AI and company filings, not numbers independently verified by a neutral auditor, and Mozilla is itself an advocate for open source. Mozilla's CTO Raffi Krikorian leans into that: 'Mozilla exists because one company tried to own the front door to the web, and an open community rose up to make sure it never could... We bet on open the first time. Open won. Together, we can do it again.'

Primary source, verified: read the paper →

Ground Truth. (2026, July 17). Mozilla Report: Open-Weight Models Now Route the Majority of AI Tokens. Ground Truth. https://groundtruth.day/news/mozilla-open-source-ai-report-open-weights-majority-tokens.html

@misc{groundtruth:mozilla-open-source-ai-report-open-weights-majority-tokens,
  title  = {Mozilla Report: Open-Weight Models Now Route the Majority of AI Tokens},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/mozilla-open-source-ai-report-open-weights-majority-tokens.html}
}

Topics: open-weight · report · mozilla · industry · economics