Open Models Now Run 63% of AI's Token Traffic
Two years ago, open-weight models processed less than 5% of the tokens flowing through OpenRouter, the largest model-routing platform in the AI ecosystem. Today that number is north of 63%, and rising. Mozilla's inaugural State of Open Source AI report, published July 14, dropped the receipts: the five highest-volume models on OpenRouter by token count are now all open-weight. DeepSeek V4 Flash, Xiaomi's MiMo-V2.5, Tencent's Hy3 Preview, MiniMax M3, and a stealth entry called Owl Alpha collectively process more tokens than any closed model on the leaderboard — including Anthropic's Claude family.
Read the full version with charts and embedded sources on ComputeLeap
This is not a popularity contest. It is a purchasing decision repeated billions of times per day, and the market is choosing open.
The Numbers That Matter
Mozilla's report, authored by CTO Raffi Krikorian and built on a survey of 950+ developers alongside platform telemetry, lays out the structural shift in hard data.
The cost collapse is the headline. Running a model with GPT-4-equivalent performance cost roughly $20 per million tokens in mid-2023. Today, the same capability costs about $0.40 — a 50x decline in 36 months. DeepSeek V4's pricing sits at $0.55/$2.19 per million input/output tokens, versus OpenAI's o1 at $15/$60. At 90% capability parity, closed models cost approximately 6x more per API call.
The adoption numbers confirm the economics. 79% of developers adding AI functionality now use open models, versus 71% for closed models. Half of developers use both. But here is where the story gets interesting: only 51% of open-model teams reach production, compared to 63% for closed-model teams. The gap is not about quality. It is about tooling, support, and operational maturity — what Mozilla calls "the harness."
The geographic split is stark. Greater China and East Asia lead open-source adoption at 89%. South America and Western Europe are the only two regions where closed adoption exceeds open. Twelve new national AI strategies launched in the past year, and 47 countries now restrict foreign processing for critical workloads. Open models are becoming a sovereignty play, not just a cost play.
The 50x cost cliff in context: Uber exhausted its entire annual AI budget in four months. Microsoft canceled Claude Code licenses after token billing consumed department budgets. These are not anecdotes about mismanagement — they are symptoms of a pricing structure that makes closed-model inference unsustainable at scale. The cost math has flipped: you now need to justify paying more, not less.
View original discussion on Hacker News
What "The Harness" Tells You About the Real Moat
The most consequential concept in Mozilla's report is not a model or a benchmark — it is "the harness." If the model is the engine, the harness is everything that turns that engine into a working vehicle: memory, tool connections, safety controls, and the software that decides what the AI is allowed to do on someone's behalf.
Mozilla's data shows that swapping the harness around a model can produce larger performance gains than swapping the model itself. On Terminal-Bench 2.0, a third-party scaffold running Anthropic's own weights scored 79.8% — versus Claude Code's 58.0% using the same weights with Anthropic's own harness. That is a 21.8-point spread from tooling alone. (Terminal-Bench 2.1, which let labs optimize their harnesses, closed the gap to about 3 points — proving the point rather than refuting it.)
This has direct implications for anyone choosing an inference stack. The model is increasingly the cheapest, most replaceable component. The harness — your evaluation framework, your prompt engineering, your tool orchestration — is where differentiated value lives. We wrote about this dynamic earlier this year in our analysis of why harness engineering matters more than your AI model, and Mozilla's data validates the thesis with large-scale benchmarks.
The Revenue Paradox: Winning Volume, Losing Money
Here is the contrarian read that the "open is winning" narrative obscures: open models power roughly one-third of real-world AI applications but capture only 4% of global AI revenue. That ratio — 8x more usage than revenue — is a sustainability crisis hiding inside a growth story.
The venture capital flowing into open-model infrastructure is real. Together AI just closed an $800M Series C at an $8.3B valuation, with annual bookings exceeding $1.15B. Mistral reports approximately $400M ARR with 20x year-over-year growth. DeepSeek hit approximately $220M ARR and raised $7.4B at a $50B+ valuation. But these numbers still pale against the closed-model ecosystem's revenue concentration.
Contrarian corner: Open models dominate token volume because they are cheaper, not necessarily because they are better. On reasoning, long-context retrieval, and complex agentic tasks, closed models still lead. Gemini 3's multi-needle retrieval hits 89% at 1M tokens versus DeepSeek V4-Pro's 41%. The production deployment gap (51% vs 63%) suggests that enterprises still trust closed providers with mission-critical workloads. The volume victory may be hollow if it concentrates on commodity tasks while closed models keep the high-margin, high-stakes work.
The counterargument writes itself: capability gaps have been closing at an accelerating rate. The closed-to-open gap went from 8.04% in January 2024 to 3.3% by March 2026. This spring, the strongest closed model scored 60 on aggregate benchmarks and the strongest open model scored 54. A year earlier, the leading open model managed 22. If the trend holds, parity on reasoning and long-context arrives within six to twelve months — and the cost advantage remains.
View original analysis on Substack
The China Factor: Strategy, Not Charity
The geographic composition of the open-model surge matters as much as its scale. Chinese open-weight models rose from under 2% of OpenRouter tokens in late 2024 to 61% of the top-10 most-used models by June 2026. Four of the five highest-volume models are Chinese. Meta's Llama, which led open-weight rankings two years ago, has fallen below 1% market share.
This is not accidental. China's "AI Plus" Initiative and its National Five-Year Plan treat open-source AI as a cornerstone of national strategy. Labs like DeepSeek, Moonshot (Kimi), Zhipu AI (GLM), Alibaba (Qwen), and MiniMax are releasing frontier-class models under permissive licenses at prices that undercut Western closed providers by 10-30x. Qwen alone crossed one billion Hugging Face downloads in January 2026 and now accounts for over 50% of all open-model downloads globally.
As the top-voted comment in the HN thread put it: frontier models are "an edge and a liability — astronomically expensive to train." The thread surfaced genuine uncertainty about whether open models represent a structural threat to frontier AI companies, with most participants acknowledging trade-offs between capability, cost, and accessibility rather than declaring a clear winner.
The Fable 5 Incident: Why Control Matters
Mozilla's report contains one of the sharpest case studies for the open-model argument. In June 2026, Claude Fable 5 went dark globally following a single government export order. Access was cut for everyone at 5:21 p.m. on a Friday — no advance notice, no geographic targeting, no gradual wind-down. Every team that had built their production stack on Fable 5 lost access simultaneously.
This is the single-vendor risk that token-share data abstracts away. When your model is an API call, your model is someone else's policy decision. Mozilla's framing — "who gets to decide when an AI model disappears?" — resonates because the event happened. It is not a theoretical risk. It is a documented incident with real production outages.
The NTIA's recommendation, cited in the report, is to monitor open weights rather than restrict them, noting that security concerns are addressable via harness-layer controls.
What the Community Is Saying
The Hacker News discussion (386 points, 284 comments) reflects the nuanced reality practitioners face. The top-voted comment argues that open models will undermine Anthropic and OpenAI's business models because frontier models are "astronomically expensive to train" and the real value is in the harness that makes models deterministic.
But the thread is not uniformly bullish on open. Multiple commenters note that local open weights still underperform Claude Sonnet on real-world coding tasks despite 48GB RAM setups. Others predict that frontier companies will maintain dominance through hardware access control. The counterpoint: Meta and others selling excess compute will commoditize hardware costs, and RAM availability follows exponential growth patterns.
The debate on X tells a similar story. Mozilla called out that "the AI stack is reorganizing in real time," while Jason Calacanis (1,838 likes) captured the practitioner sentiment: "Open Source models are compounding. Frontier Models are refining." This tracks with what we saw just yesterday, when Kimi K3 and Thinking Machines' Inkling shipped within 24 hours — both benchmarked against Opus 4.8 and Fable 5.
The Enterprise Calculus: Repatriation Is Real
The open-model shift is accelerating an adjacent trend: cloud repatriation. Mozilla's report cites that 80% of enterprises are repatriating AI workloads from cloud providers. The economics are straightforward: AWS S3 egress runs $90k-$120k per petabyte, and companies like 37signals documented savings from $3.2M to under $1M by moving off cloud. GEICO reported spending 2.5x its planned cloud budget before pulling back.
When you combine open-weight models (no per-token API fees) with self-hosted inference (no cloud markup), the total cost of ownership drops by an order of magnitude. Stripe achieved a 73% inference cost reduction by moving to vLLM for self-hosted open-model inference. Together AI's $1.15B in annual bookings represents enterprises making exactly this calculation at scale.
This is the structural shift that matters for infrastructure planning. The question is no longer "should we use open models?" — it is "what is the minimum capability threshold that justifies paying closed-model prices?" For most production workloads that are not frontier reasoning or 1M-token context retrieval, the answer is increasingly: nothing. We traced the broader inference inflection earlier this year, and the trendlines have only steepened.
What This Means for You
If you are building or maintaining an AI-powered product, here is the decision framework this data supports:
Default to open-weight models for production inference. The cost advantage is 6-20x, the capability gap is 3.3% on aggregate benchmarks, and the vendor risk of closed models is now a documented production hazard. Start with DeepSeek V4 Flash or Qwen 3 for general tasks, GLM-5.2 for coding workloads.
Invest in your harness, not your model loyalty. Mozilla's Terminal-Bench data shows that scaffolding quality drives a 21.8-point performance swing — far more than model selection. Build evaluation frameworks, tool orchestration, and prompt engineering that work across models. Your harness is your moat; the model is a commodity input.
Budget for the transition. Together AI, Fireworks, and the open-inference providers have crossed the $1B revenue mark collectively. The tooling is enterprise-ready. Self-hosted vLLM on commodity GPUs is a proven path. The 12-point production deployment gap (51% vs 63%) exists because of tooling immaturity, not model quality — and that gap is closing fast as the ecosystem matures.
Watch the capability frontier, not the volume charts. Open models win on volume and cost. Closed models still lead on frontier reasoning and long-context tasks. Your stack needs a fallback path to closed models for the 10-15% of workloads where the capability gap matters. Design for model portability from day one — we covered why this portability matters in our breakdown of AI's $700B subsidy clock.
The one-line takeaway: Open-weight models are no longer the alternative. They are the default. Closed models are the specialty tool you reach for when the default is not enough — and you should be measuring exactly how often that happens.
The Trajectory
Mozilla's report captures a moment, but the trajectory is the story. Two years ago, open models were an experiment. One year ago, they were competitive. Today, they process the majority of the internet's AI token traffic. The 50x cost collapse, the 3.3% capability gap, and the geopolitical push toward AI sovereignty are all compounding in the same direction.
The question for closed-model providers is no longer whether they can maintain a capability edge. It is whether the edge is wide enough to justify prices that are 6-20x higher, while the open ecosystem closes the gap at an accelerating rate and governments around the world bet on open infrastructure.
Raffi Krikorian framed it plainly: "I don't want seven AGIs, one for every single one of the big companies. I want seven billion AGIs."
The token-share data suggests the market agrees.
Originally published at ComputeLeap
Top comments (0)