Same Output, 300x the Bill — How to Split Work Between Open-Weight and Closed Models

The same tasks went to two different AI models — summarizing documents, fixing code, drafting emails. The outputs were not noticeably different. The bills were: $50 on one side, 14 cents on the other. Three hundred times.

The usual reaction is skepticism, and it is reasonable. So here is the conclusion up front: saving 300x is not always a win, but for certain work it clearly is. Where that line sits is the whole point.

How Far Does "Free" Actually Go

Open-weight means the model's weights are published. Weights are the numbers a model learned during training. In cooking terms, you are not receiving the recipe — you are receiving the finished stock someone else simmered.

Which matters, because free stock does not make the meal free. You still need a stove and a kitchen. The compute to run the model still costs money. That 14 cents is the stove. The model is free; the operation is not.

Meta's Llama opened this lane, with DeepSeek and Alibaba's Qwen following.

The Numbers

Based on published pricing aggregates:

• Median open-weight price: $0.53 per 1M tokens
• Median closed model price: $2.63 per 1M tokens

Individual models widen the gap further:

• Llama 4 Scout: $0.18 per 1M input tokens
• Claude Sonnet 4.6: $3 per 1M input tokens

That is roughly 16x on input and 25x on output.

So Why Still Pay

Three things genuinely differ, and pretending otherwise is dishonest.

Hard problems separate them. Easy work looks the same. Summarizing, translating, reformatting — you will not feel the difference. Multi-step problems requiring judgment mid-chain are where the gap opens.

Setup effort. A paid API works the moment you open it. Open-weight needs installing, configuring, tuning. Price your own hours and that setup time can consume the entire 300x saving.

Someone to call. When something breaks, is there an accountable vendor? For work touching company data, that often outweighs price.

Open-weight is not equivalent across all workloads. Difficult reasoning, chains involving multiple tools, and top-tier coding still lag. Industry analysis says the performance gap has narrowed, not closed — roughly 12 to 16 months.

But Few Tasks Need All Three

Write down what your company actually does in a day and the picture clarifies. Most of it is repetition:

• Drafting emails
• Cleaning up meeting notes
• Polishing product descriptions
• Extracting numbers from documents
• First-pass triage of customer inquiries

Does that work need frontier performance? No. And that work is the highest volume. Hard tasks number a handful per day; easy tasks number in the hundreds. Cost multiplies by volume. The 300x is not the scary part — 300x multiplied by hundreds of daily jobs is.

So the answer is not choosing one. It is splitting: high-volume easy work to the cheap side, rare hard work to the expensive side.

Run the Math

Say you have a repetitive task running 300 times a day. Even at modest tokens per job, that is 9,000 jobs a month. Apply a 16x difference and a bill in the hundreds of dollars becomes tens.

Meanwhile, five hard judgment calls a day routed entirely to the expensive model still costs very little monthly. The failure pattern is economizing on the expensive thing while missing where the money actually goes.

This is not an either-or decision. It is attaching different tools to different jobs — the same way you draft in a word processor and move the numbers to a spreadsheet.

Data Stays Inside

There is something that may matter more than price. Open-weight models can run on your own machines, which means company documents never leave. Contracts, HR files, unreleased results. If uploading those to someone else's servers has always felt uncomfortable, this removes the concern structurally — nothing gets uploaded.

There is a catch. Running it yourself means owning the security yourself. If someone else's servers felt unsafe, ask honestly whether yours are safer. Without anyone to maintain them, they are more dangerous, not less.

Blocking access is easy. The problem is that without an alternative, employees end up using personal accounts — data leaves through a channel the company cannot see, with no log at all.

That is why self-hosting comes up. It is not free either: hardware, someone to maintain it, and rework every time models change. Below a certain scale the overhead exceeds the benefit.

One test decides it: among the documents currently leaving for someone else's servers, is there anything that genuinely should not? If yes, bring that workload in first. If no, there is no urgency.

What to Do This Week

Skip the grand plan and follow this order.

1. Open last month's bill and find where the money went. Usually one or two workloads consume most of it.
2. Ask whether those one or two are genuinely hard. If they are simple but high volume, that is your migration candidate.
3. Move that single workload to the cheap side for one week. One, not everything. Roll back if quality drops.
4. After a week, compare two things: how much the bill fell, and how much more human handling it required. Skip the second and your math is wrong.

Do you need to stand up your own server? Not at this stage. Several services run open-weight models and bill by usage, with integration close enough to what you already use that starting can mean changing an endpoint.

Hardware comes later — after the migration proves useful, volume justifies it, and a real data-residency reason emerges. Reverse that order and you buy equipment you never use.

This Is Not an Argument for Cheap

The point is not cutting costs. It is operating with knowledge of what things cost. Until recently there was one option, so there was nothing to compare. Now there are options separated by 300x. If everything still routes to the expensive side, that is likely not a performance judgment — it is that no one has ever split the workload.

Splitting it takes a week.

One more thing: the gap will not hold. Closed models are cutting prices and open-weight models are gaining capability. Today's conclusion is not guaranteed to hold in six months, so re-measure quarterly rather than deciding once.

Related Reading

AI costs only make sense when you can see the contract structure behind them. The analysis of how much of Meta's real AI cost sits off the books in unreported commitments shows how platform pricing gets built. And any migration to cheaper models needs a verification step attached — making "done" provable supplies that framework.

Bottom Line

How far to shift toward the cheap side is your call. But making the call requires looking at the numbers first — starting with last month's bill.