The most interesting number in AI this month is not a benchmark score. It is an invoice, and the share of it that is no longer going to a frontier lab.
1. Open Weights Took the Majority of Production Traffic and Nobody Announced It #
Vercel's AI Gateway published its August numbers and the crossover is done: open-weight models handled 56% of production tokens, up from 36% in July and under 10% last December, peaking at 62% on one day. Those same models took 14% of estimated spending. Closed tokens cost about 7.8 times as much on average, which fully explains the gap between those two numbers.
The buying behavior underneath is the part to watch. The average price per token through the gateway fell 23.2% in August, and among teams spending over ten million tokens in both months the median fell 7.6%, more than double July's decline. The substitution is not only open-for-closed: within Anthropic's own lineup, the most capable model's share of spend dropped from 13.2% to 4.9% while a model at about half the price climbed to 22.5%, with nine in ten teams cutting use of the expensive one. Anthropic still takes 64% of the money. It is taking it on cheaper models.
Treat the figures with care — one gateway's anonymized traffic, spend estimated from list prices, no visibility into direct API calls or private deployments. But the direction matches what capability data says is now possible: Mozilla's 2026 open-source AI report puts the benchmark gap between leading open and closed models at about 3.3 points, with near-parity on coding and instruction-following. When the gap is three points and the price gap is 7.8x, routing stops being a matter of taste.
Why it matters:
- For ICs: the skill is no longer picking the best model, it is knowing which tier your task needs. Build the eval that tells you, or keep guessing expensively.
- For leaders: if every request hits your most capable model, you have a 7x line item and no policy. A router with a cheap default and an escalation path is a week of work.
- For founders: hard-coding one vendor's SDK is a pricing risk now, not just a portability one. Model-agnostic harnesses and IDE integrations are where this curve is bottlenecked.
2. The Model You Could Not Run Last Year Fits on the Machine Under Your Desk #
A CMU lab spent the week publishing what a couple of GPUs can now do, and the inference numbers deserve a second read. Their framework runs Qwen 3.8 Flash Next, 125 billion parameters, on a single 24 GB desktop card. On a 128 GB MacBook, an AMD Strix box or a DGX Spark, it runs DeepSeek V4.1 at 550 billion. Quantized Metal inference of a 35B mixture-of-experts model hits 450 tokens per second at 1.5 bits per weight — a tenth of the memory a half-precision copy needs, fast enough to feel like a local process rather than a remote service.
The Metal kernels were not hand-written. They pointed an agent harness at the repository, told it to optimize the kernels, and left. That is the durable claim here: the expensive part of running models locally was never only the hardware, it was the engineering expertise, and that half is now itself automatable. The same lab's context-compaction technique reports halving session cost while letting agent runs stretch past a hundred million tokens, with one partner company measuring a 45% cut in total AI spend after deploying it.
Long-running agents, batch evaluation and reinforcement-learning loops tolerate latency and do not care what silicon they land on, which makes them the natural first workloads to leave hosted APIs — for privacy, for cost, or because the box is already paid for. Frontier-class capability is no longer what is scarce. Deciding where it should run is.
Why it matters:
- For ICs: run a 100B-class model locally once this quarter. The ceiling on what fits has moved so far that most engineers' intuitions are a year stale.
- For leaders: you have latency-tolerant workloads — nightly evals, backfills, bulk classification — sitting on premium per-token pricing out of habit. That is the cheapest migration available to you.
- For founders: the gap between good local weights and a product someone can actually run is still wide. Tooling that hides quantization, serving and context management is an open lane.
3. The Substrate Everyone Is Standardizing On Has a Country of Origin #
Here is the part nobody puts on the slide. The open-weight ecosystem is overwhelmingly Chinese, and the lead is widening. Chinese models sit at 3.2 billion Hugging Face downloads, roughly twice the American total. On the Artificial Analysis intelligence index the top open models score 45, 44 and 42, against 26 and 23 for the best American open releases. OpenRouter's open-model traffic went from about 1 trillion tokens a week a year ago to roughly 80 trillion today, with the Chinese share climbing from 70% to over 80%; one open-source coding agent reports 95% or higher.
This is already in your dependency tree whether or not you put it there. Cursor, Harvey, Airbnb, DoorDash and Perplexity all build on Chinese open weights, and American startups are increasingly signing enterprise agreements directly with those labs. In research it is starker: Qwen appears in about 30% of new machine-learning papers against Llama's 21%, and some Chinese open model in over 40%. The engineers being trained right now learned on these weights.
The strategic read is procedural, not alarmist. A model is a dependency with a license, a jurisdiction and a migration cost, and most teams track none of those for the component doing the most work. Regulatory attention on open weights is real and the legal mechanism is unsettled, so the cheap insurance is boring: know which checkpoints you depend on, keep an eval suite that makes swapping one a two-day job, and do not let a single provider's endpoint become the only place your weights can run.
Why it matters:
- For ICs: read the license on the weights you pulled. "Open" covers a wide range of terms, and some of them do not permit what your employer is doing.
- For leaders: add models to the dependency inventory you already keep for libraries. Same question, higher stakes: what breaks if this becomes unavailable on Monday?
- For founders: portability across checkpoints is a feature you can sell. Buyers are starting to ask where their weights come from, and most vendors cannot answer.
- The pattern across all three: intelligence is commoditizing faster than anyone's procurement process, and the teams that win the next year are the ones treating model choice as an engineering decision rather than a brand one.
The Verdict: Real or Hype? #
Open weights as the production default → Real. Majority of tokens, a seventh of the spend, and a capability gap small enough that the price difference decides it. Self-hosted frontier-class inference → Real but unevenly distributed. The hardware math works today for latency-tolerant work and still does not for interactive, high-concurrency serving. Paying frontier prices for routine coding tokens → Hype. Teams are already stepping down inside vendor lineups, and the ones who have not are funding someone else's benchmark.