The Vibe Father
Analysis
"You can self-host an open-weight model" is true, and it gets repeated as if it settles something. It does not. "Can" and "should" are different claims, and the gap between them is a spreadsheet most people never fill in. This is that spreadsheet — the honest calculus of when self-hosting a coding model actually pays, and when it is a costly way to feel independent. We run these models daily and route across hosted and self-hosted setups, so this is the working version, not the brochure.
The three things self-hosting genuinely buys
There are exactly three reasons that hold up under scrutiny, and it is worth being blunt that most "I want to self-host" impulses do not map to any of them.
- Privacy and compliance. If code contractually cannot leave your perimeter, or you are under a regime that demands it, self-hosting is the only camp that delivers "the data never left the building." This is the strongest reason and the most common real one.
- Very high volume. Per-token cloud pricing scales linearly; fixed local cost does not. Once you are running agents flat-out all day, every day, the marginal self-hosted token approaches free, and the hardware amortizes. Below that volume line, cloud is cheaper.
- An open model that is "good enough" for your workload. Self-hosting only makes sense if an open-weight model actually covers the work you route to it. For a lot of well-scoped implementation, it does — but you have to confirm that on your tasks, not assume it.
If your situation hits one of these, keep reading. If it hits none of them, the honest answer is usually "use a hosted API" and you can stop here.
Is an open model good enough to self-host?
This is the load-bearing question, so look at the numbers. The best open-weight SWE-bench Verified scores on our board at /benchmarks are strong: GLM 5.2 at 78.7, DeepSeek V4 Pro at 77.6, Qwen3.7 Max at 77.3, Kimi K2.6 at 76.7. For a lot of routine, well-scoped work behind a reviewer, that is genuinely enough.
| Model | SWE-bench | Weights | Hosted API price (in/out per M) |
|---|---|---|---|
 GLM 5.2 | 78.7 | open | — |
 DeepSeek V4 Pro | 77.6 | open | $0.435 / $0.87 |
 Qwen3.7 Max | 77.3 | open | $2.50 / $7.50 |
 Claude Fable 5 | 95.0 | closed | $10 / $50 |
Now the counterweight, which is the reason this is a real decision. The closed flagships still sit well above the open field: Fable 5 at 95.0 is sixteen points ahead of the best open score. If the work you would self-host lives at the hard, ambiguous tail, self-hosting means choosing a materially weaker model to save money — and on tasks where a wrong turn costs a day, that trade usually loses. Self-hosting makes sense for the volume seats, not the frontier ones.
When it does not pay
Three situations where "just use a hosted API" is the right answer, and self-hosting is a trap dressed as independence.
- You would need big GPUs you do not have. Frontier-scale open models are large; serving one at production quality means serious hardware you buy, power, and depreciate. If that capital is not already justified, per-token cloud is cheaper for a long time.
- You would carry the ops burden. An inference stack is a service you now operate — updates, monitoring, capacity, the 2 a.m. page when it falls over. That engineering time is a real cost people leave off the spreadsheet, and it is often the line that flips the decision.
- You need frontier capability. No self-hosted model currently matches Fable 5 or Opus 4.8 on SWE-bench. If depth decides your outcomes, self-hosting is optimizing the wrong variable.
And note the sneaky one: hosted APIs for open models are often startlingly cheap precisely because the weights are open and providers compete to serve them. DeepSeek V4 Pro at $0.435/$0.87 may well undercut your fully-loaded self-hosted cost per token until your volume is very high. You can capture most of the open-weight benefit — price pressure, no rug-pulls, portability — through a hosted API without provisioning a single GPU.
The insurance you get for free
Here is the underrated part: even if you never self-host, the fact that you could disciplines the market. An open model's hosted API price cannot drift far from the cost of serving it yourself, and the model cannot be deprecated out from under a workflow you depend on. That optionality is worth something on its own — some of the value of open weights accrues to you without ever downloading them. We treat this as a core part of the cost story in the economics of AI coding.
Bottom line
Self-hosting pays when a hard requirement forces it — privacy, compliance, very high volume — and when an open model is genuinely good enough for the work you would route to it. It does not pay when you would be buying GPUs and an ops burden to run a weaker model than the cloud offers, which describes most teams most of the time. The right default is hosted APIs for open models, with self-hosting as the option you exercise deliberately. This is why we build model-agnostic: run open and closed, local and cloud, and route each job to the seat it deserves. The full open-versus-closed picture is in our open versus closed analysis, and live scores stay at /benchmarks.
