--- title: "Open Weights Took 62% of the Tokens and Under 9% of the Spend" slug: "open-weight-models-62-percent-volume-9-percent-spend" author: "ibl.ai" date: "2026-08-28 12:00:00" category: "Premium" topics: "open weight models, vercel ai gateway, model routing, inference cost, deepseek, qwen, glm, model-agnostic, vendor lock-in, llm pricing" summary: "Vercel's AI Gateway put open-weight models at 62% of token volume in late August, up from 11% in April — while closed models still took roughly two-thirds of the spend. That split is not a contradiction, it is what a correctly routed AI estate looks like, and it is only available if switching models is a config change." banner: "" thumbnail: "" linkedin: | Open-weight models reached about 62% of token volume on Vercel's AI Gateway on 22 August, up from 28.4% in late June and 11% in April. The number circulating alongside it is usually wrong, so worth separating: the 62% is open-weight models overall, not Chinese models specifically. Chinese-built open models were around 29% of tokens in June, at roughly a tenth of the average token price. The far more interesting figure is the other half of the same dataset. While open weights took 62% of the volume, they captured under 9% of the spend — and Anthropic's closed models held roughly 61-65% of it. Most people read that as a contradiction, or as evidence open models are not really winning. It is neither. It is what a correctly routed estate looks like. The volume is classification, extraction, summarization, routing, tool-calling — high-count, low-stakes work where a cheap open model is indistinguishable in output and 10x cheaper. The spend is the narrow set of genuinely hard reasoning tasks where the frontier still earns its price. That is not a market splitting. That is engineers doing cost engineering, at scale, in production. Which makes the strategic question simple, and it is not "open or closed." It is: can you move a workload between them without a rewrite? If routing a task from a frontier API to an open model is a configuration change, this chart is an opportunity you can act on every quarter as prices move. If it is a migration, the chart is just news. Model lock-in is the new technical debt, and this dataset is what paying it down looks like. On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing. #iblai #OpenSource #LLM #AIInfrastructure #EnterpriseAI #ModelAgnostic --- ## The Short Answer **Open-weight models reached roughly 62% of token volume on Vercel's AI Gateway on 22 August 2026, up from 28.4% in late June and 11% in April — while capturing under 9% of the spend, with closed models holding around two-thirds of it. That split is a routing strategy, not a contradiction. On ibl.ai you own all the code and the data and run it model-agnostic across any LLM, which is what makes acting on it a configuration change.** Volume and spend answer different questions. Reading only one of them produces the two most common wrong takes about this dataset. ## What does the Vercel AI Gateway data actually say? The gateway data says open-weight models won the volume and closed models kept the revenue, over a remarkably short period. Vercel reports open-weight share of tokens moving from **11% in April** to **29% in June** to about **62% on 22 August**. The June figure was 28.4% on the 24th, so the rise across two months is roughly a doubling. Vercel publishes this as a recurring [AI Gateway production index](https://vercel.com/blog/ai-gateway-production-index), which is what makes it unusually useful: it is measured production routing across many customers, not a survey of intentions. **One correction, because it is being repeated.** The 62% figure is **open-weight models overall**, not Chinese models specifically. Chinese-built open models were around **29% of tokens in June 2026**, at roughly **one-tenth** the average token price of US frontier systems on the platform. Both facts are interesting; conflating them overstates one and erases the other. ## Why do open models dominate volume while closed models keep the spend? Open models dominate volume and closed models keep spend because the two categories are being pointed at different classes of work. At 62% of tokens, open-weight models captured **under 9% of spend**, while Anthropic's closed models held roughly **61–65%** of it. The explanation is unglamorous. High-volume production work is classification, extraction, summarization, routing, reformatting, and tool-calling — tasks where a competent open model produces output a reviewer cannot distinguish from a frontier model's, at a fraction of the price. That work generates enormous token counts and small invoices. The spend concentrates in the narrow band where frontier reasoning genuinely changes the outcome: long multi-step agentic runs, hard synthesis, ambiguous judgment. Fewer calls, much higher value per call. So the barbell is not a market fragmenting. It is cost engineering, executed at scale, in production — and it is roughly what a well-run AI estate should look like. ## Is this evidence that open models have "won"? No, and the framing is the problem. This data shows open models winning a specific argument — the one about routine work — rather than winning outright. The maximalist read ("open weights have taken over") ignores that two-thirds of the money still goes to closed models for the hardest tasks. The dismissive read ("they only have 9% of spend, so it's hype") ignores that 62% of production traffic now runs on models anyone can download and self-host. Both readings share a mistaken assumption: that this is a contest with one winner. The organizations in this dataset are not choosing sides. They are routing. The practical implication is that "which model should we standardize on" is the wrong question, and increasingly an expensive one. The right question is what fraction of your workload is currently over-served by an expensive model, which is a measurable number rather than a philosophical position. ## What does it take to act on this? Acting on it requires that moving a workload between models is a configuration change rather than a migration. That is the whole difference between this chart being an opportunity and being news. If your application code, prompts, tool definitions, retrieval layer, and evaluation harness are bound to one vendor's API, then re-routing a workload to a cheaper open model is a project — and by the time it ships, the prices have moved again. Three things make routing cheap: **A model-agnostic platform boundary.** The application asks for a capability; the platform decides which model serves it. Swapping the answer does not touch application code. **Your own evaluation set.** You cannot route by cost without knowing where quality actually degrades. Public benchmarks will not tell you; a few hundred examples of your real traffic will. **Somewhere to run the open models.** The 62% is only fully available to you if you can host, which is a serving-stack question. We cover the runtime layer in the [self-hosted LLM serving stack](/blog/self-hosted-llm-serving-stack-ollama-vllm-tgi), and the architectural reason the newest open models are cheaper to serve in [GLM-5.3-Flash and the KV cache](/blog/glm-53-flash-hybrid-attention-kv-cache-self-hosting). On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ## Why is model lock-in a form of technical debt? Model lock-in behaves like technical debt because it accrues silently and the interest is paid in foregone savings rather than in visible failures. Nothing breaks when you are locked in. The system works. What happens instead is that every price move, every capable open release, and every capability shift arrives as an opportunity you cannot take — and the gap compounds quarterly. This dataset makes the cost legible. An organization that can route moved a majority of its tokens onto models costing roughly a tenth as much within four months. An organization that cannot route did not, and paid the difference. That is also why the ownership question sits underneath the routing question. Being able to run any model is only half of it; the other half is owning the layer that decides which one runs, so the decision stays yours as the field moves. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. ## Frequently asked questions **Is the 62% figure about Chinese models?** No. It is open-weight models overall on Vercel's AI Gateway as of 22 August 2026. Chinese-built open models were a subset — about 29% of tokens in June — and the two numbers are frequently conflated. **Does low spend mean open models are low value?** No, it means they are cheap. The same dataset shows them carrying the majority of production tokens, which is a statement about how much work they are doing, not how little they are worth. **Should an organization move everything to open weights?** No. The pattern worth copying is routing, not switching. Move the high-volume, low-ambiguity work where quality is indistinguishable, keep frontier models for the tasks that genuinely need them, and re-measure as prices change. ## The bottom line Two numbers from the same dataset, and you need both: 62% of the tokens, under 9% of the spend. Read together they describe an industry that has stopped asking which model is best and started asking which model is sufficient for each task. That question only pays if switching is cheap — which is a property of your architecture, not of the model market. ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.