The Vibe Father
Analysis
The open-versus-closed debate for coding models has finally stopped being tribal and started being a spreadsheet. In mid-2026 the honest picture is boring in the best way: closed flagships still own the top of the capability curve, open-weight models own the benchmark-per-dollar curve, and which one you should use depends almost entirely on the workload. This post lays out the trade-offs with numbers rather than allegiances, using our board at /benchmarks.
The state of play, on one axis
Start with SWE-bench Verified, the benchmark closest to real repository work and the one we weight highest. Here is the frontier of each camp.
| Model | SWE-bench | Camp | Price (in/out per M) |
|---|---|---|---|
 Claude Fable 5 | 95.0 | closed | $10 / $50 |
| Claude Opus 4.8 | 88.6 | closed | $5 / $25 |
 Gemini 3.5 Flash | 79.3 | closed | $1.50 / $9 |
 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 |
 Kimi K2.6 | 76.7 | open | — |
Two facts fall straight out of the table. First, the top is closed: Fable 5's 95.0 and Opus 4.8's 88.6 sit above every open model, and the gap from Fable to the best open score (GLM 5.2's 78.7) is a ten-to-sixteen-point canyon depending on which open model you compare. On hard multi-file tasks you feel that gap. Second, the open cluster is tight and cheap: GLM 5.2, DeepSeek V4 Pro, Qwen3.7 Max, and Kimi K2.6 land within two points of each other, and DeepSeek does it at $0.435 in and $0.87 out — roughly a rounding error against the flagships. The open models have caught the closed mid-tier and are a benchmark-per-dollar bargain; they have not caught the closed frontier.
What "open-weight" actually buys you
Beyond price, open weights buy three things that do not show up on a benchmark. Control: you can run the model inside your own perimeter, no code leaving the building — the whole reason self-hosting exists. Durability: weights on your own disk cannot be deprecated, repriced, or silently changed under a workflow you depend on, whereas a closed model can be any of those on the vendor's schedule. And market discipline: even if you never self-host, the fact that you could keeps an open model's hosted API price tethered to the cost of serving it. Some of the value of open weights accrues to you without downloading a thing.
What closed models buy you is the top of the capability curve, plus first-party infrastructure you do not operate, plus — usually — the newest frontier features first. For work at the genuinely hard tail, that is worth paying for, and pretending otherwise is how teams ship subtly broken refactors to save a few dollars a million tokens.
Trade-offs by workload
The decision is per-job, not per-team. A few patterns we actually run:
- Hard, ambiguous, multi-file work — architecture, gnarly cross-cutting refactors, the ticket where a wrong turn costs a day — goes to a closed flagship. The 10-16 SWE-bench points buy you first-pass correctness, and correctness is cheaper than retries.
- High-volume, well-scoped implementation — the bulk of a normal sprint — goes to an open builder like DeepSeek V4 Pro or GLM 5.2 behind a reviewer. You get ~90% of a closed builder at 10-30% of the price.
- Reviewing, drafting, scouting, cheap parallel passes — go open every time; the price lets you run more of them.
- Anything privacy- or compliance-bound — goes open and self-hosted, because that is the only camp that can, and the quality is now good enough that self-hosting no longer requires an apology.
Why picking one camp is the mistake
The framing error is treating this as a team-wide religious choice. It is a routing problem. The right answer for most shops is "both" — a closed flagship in the seats where depth decides the outcome, open builders carrying volume behind gates, and the freedom to move a job between them as the models and prices change week to week. That is precisely the case for a model-agnostic setup, and it is why The Vibe Father runs open and closed side by side and routes each job to the right one on your own keys at a flat price rather than marking up tokens. We make the full argument in why a model-agnostic harness wins.
Bottom line
Open versus closed is not a winner-take-all fight and treating it as one costs you money either way — you overpay by putting a flagship on drafting, and you ship bugs by putting an open mid-tier on the hard tail. Closed still owns the frontier; open owns benchmark-per-dollar and everything privacy touches. Route accordingly. If your constraint is spend, start with our cheapest models roundup; if you want the open field in depth, read the open-weight roundup. Live scores at /benchmarks.
