Gemma 4 31B rents for $0.09 per million input tokens. And that single number explains why the open-weight LLM price war matters to anyone running a lean operation. A survey that read the pricing pages of 18 inference providers on October 18, 2026 recorded 80 offers across 12 open-weight models. And the floors keep dropping (Developers Digest). One August analysis citing BenchLM data put the median blended open-weight API price at $0.53 per million tokens, 82% below the $3.00 median for proprietary models (yage.ai). What that does to the local inference argument is mostly kill it: the reported crossover where self-hosting beats hosted APIs sits around 10-30 million tokens per day. And most small operations never push that volume (packet.ai).
The Floor Dropped Out of Open-Weight Pricing
The headline stat is the 82% gap between open-weight and proprietary medians. But the per-model numbers are where you actually make decisions. One comparison of open-weight models lists prices that now span a genuinely absurd range (Codingscape):
| Model | Input $/M tokens | Output $/M tokens |
|—|—|—|
| DeepSeek V4 Pro | $0.435 | $0.87 |
| Mistral Large 3 | $0.50 | $1.50 |
| GLM 5.3 | $1.40 | $4.40 |
| Qwen 3.8 Max | $1.65 | $4.95 |
| Kimi K3 | $3.00 | $15.00 |
Read that table twice, because it breaks the assumption most people still carry. “Open-weight” is not a price tier.
DeepSeek V4 Pro at $0.435 in and $0.87 out is commodity pricing, while Kimi K3 at $3.00 in and $15.00 out is premium pricing wearing an open license.
The same source puts DeepSeek V4 Flash at $0.14 per million input tokens against Fable 5’s $10, which is not a competitive gap so much as two different industries.
One cost analysis put the aggregate picture at open-weight inference running 5-10x cheaper than closed APIs at scale (packet.ai). My take: the price collapse is real, but it benefits people who shop. It does nothing for people who picked a provider in 2025 and stopped checking.
Same Weights, Nine Another Prices
The uncomfortable part of open-weight pricing is that the same weights quote differently depending on who serves them.
A 2026 survey put Llama 3.3 70B input pricing between $0.12 and $1.05 per million tokens across providers, roughly a ninefold spread, as of June 2026 (engineering handbook). Same model. Same weights. Nine distinct prices.
You can see the same confusion across surveys. The 18-provider October survey recorded DeepSeek V4 Pro’s lowest listed input price at $1.30 per million tokens (Developers Digest), while the Codingscape comparison lists the same model at $0.435 in and $0.87 out. Neither source is wrong; they just caught separate providers on other days. That is the actual state of this market.
The practical consequence for a small operator: provider choice now moves your bill more than model choice does. Picking between Mistral Large 3 at $0.50 and GLM 5.3 at $1.40 is a smaller decision than picking between the $0.12 and $1.05 ends of the Llama 3.3 70B spread. I treat provider pricing as a quarterly audit, not a one-time setup step, since these numbers are clearly not stable enough to set and forget.
When Local Inference Actually Pays
So when does hardware make sense?
The reported self-hosting crossover is 10-30 million tokens per day.
And it moves depending on model size, utilization, and your input-output mix (packet.ai). That is a lot of tokens.
A solo operator doing content generation, classification, and the odd coding assist is usually nowhere near it.
Here is the hardware ladder those decisions run against (Tokios docs):
| Model size | Min VRAM | Example GPUs | Hardware cost |
|—|—|—|—|
| 7B-8B | 8 GB | RTX 4060, RTX 3060 | $300-$500 |
| 14B | 16 GB | RTX 4080, RTX 4070 Ti | $600-$1,200 |
| 32B | 24 GB | RTX 4090, RTX 3090 | $1,600-$2,000 |
| 70B | 48 GB | RTX A6000, 2x RTX 3090 | $4,000-$8,000 |
The $300 entry point for a 7B-8B setup looks tempting until you compare it against Llama 4 Maverick hosted at $0.25 per million input tokens.
You would be spending hundreds of dollars of capital to avoid a bill that rounds to zero at small volumes. And before you buy anything, rent: one worked example ran a Qwen3-7B model quantized to INT4 for five hours on an RTX 4090 at $0.39 per hour, total cost $1.95 (packet.ai).
That is the correct way to test whether local inference fits your workloads, and it costs less than lunch.
My honest read: if you go local in this market, go for a reason the token market cannot quote. Cost stopped being that reason sometime in the last year.
The Decision Rule I Use Now
There is a middle path that most coverage skips, and it is where I would start.
Flat-rate subscriptions for open models exist now: z.ai’s Coding Plan Max at $80 per month and Ollama Cloud Max at $100 per month were both reported to comfortably handle heavy usage volumes (yage.ai). A flat $80 or $100 against a per-token bill is the comparison worth running before you do anything else.
The decision process I follow, in order:
– Pull last month’s actual token usage before anything else. Every number in this post is useless without yours.
– If you are below roughly 10 million tokens per day, stay hosted and shop the floor price across providers rather than defaulting to whoever you signed up with.
– Check flat-rate open-model subscriptions against your per-token bill. At small volumes, one of them usually wins outright.
– Only price hardware at sustained high volume, and rent an RTX 4090 by the hour first to validate the workload before spending $1,600 to $8,000 on GPUs.
The open-weight price war is good news. But only for operators who treat inference as a line item they actively manage. The prices moved; the discipline did not move with them.
Open your provider dashboard today and pull your real token volume, then check it against that 10-30 million per day crossover. If you are under it, spend an hour shopping the 80 offers in that October survey and cut your rate before the next billing cycle.
That hour is the highest-paying hour you will work this week.
