Cost × Quality

Each point is one model's full six-task sweep. The routing thesis in one picture: the quality band is flat — what varies is what you pay to be in it. Local points carry measured electricity cost, not a mythical $0.

Scoreboard

Score 0–10 per task. Deterministic checks where possible (JSON fields, numeric answer, word-count, executed unit tests); blind LLM judge for prose. Hover a cell for the grader's note.

Consumption — the thinking tax

Completion tokens each model burned to do the same work. Local reasoning models think out loud (free at home, expensive if this were metered); efficient cloud models answer in a tenth of the tokens. Consumption, not list price, is most of what a cloud bill is made of.

The $0 myth — power economics, measured

GPU draw sampled live from the board sensor during an actual 35B generation. At residential California rates your own silicon is not free — and the cheapest cloud models undercut it on pure cash. Local buys privacy, control, and zero marginal spend; know which argument you are making.

Money — live gateway meters

Per-key spend against per-key monthly budgets, read from the gateway at publish time. The real wall is provider-side: prepaid credit, no auto-top-up.

Routing map

Two planes, by design: the work plane (BAYi + chief-of-staff seats) never touches the gateway; the lab plane never touches Anthropic. Every consumer holds its own budgeted key.

flowchart LR
  subgraph LAB["LAB PLANE — gateway :4000, tailnet-only"]
    J[John] -->|key $30| GW{{LiteLLM gateway}}
    D[Doug lab calls] -->|key $50| GW
    R[Ram · ai-box Claude] -->|key $10| GW
    B[Bench harness] -->|key $20| GW
    W[Family WebUI · opt-in via admin] -.->|key TBD| GW
    GW -->|$0| OLL[Ollama · R9700 32GB
qwen35b · gemma12b · qwen4b]
    GW -->|prepaid wall| OR[OpenRouter]
    OR --> K[Kimi K3] & G[Grok 4.3] & DS[DeepSeek v4] & Q[Qwen3-235B] & GL[GLM-5.2] & LS[Llama-4-Scout]
  end
  subgraph WORK["WORK PLANE — never crosses"]
    DG[Doug + BAYi seats] --> AN[Anthropic · Max sub]
  end
  

Exhibit A — the coffee question

Same prompt, same hardware, ten minutes apart: “How do I make a great cup of coffee?” Small models fail confidently — the tone is identical, only the facts change.

Method & honest caveats

• Fair-budget lesson: run 1 capped completions at 1,200 tokens and every local reasoning model scored 0 on math and code — they ran out of budget mid-thought. Run 2 (this page) gives all models 4,000 tokens. The “local models can’t reason” headline was a harness artifact; we kept the receipt because that mistake is the most common one in public benchmarks.
• n=1 per cell, temperature 0, model IDs version-pinned, judge (Llama-4-Scout) never sees model identity.
• Six tasks are too easy to separate frontier models — by design for v1 plumbing. A harder tier is the v2 fix.
• Latency includes network + queue + model load where applicable; local walls include thinking time.
• Power method: amdgpu hwmon sampled at 1Hz during live generation (13W idle / 218W sustained); wall-marginal ~230W with PSU losses. Rate assumption $0.35/kWh — adjust to your tariff. Framing is marginal cost (the box runs anyway); amortizing the $1,250 card over 3 years adds ~$1.15/day of standing cost regardless of use.
• This is one machine's result (John's ai-box), not enterprise deployment economics.