Saturday, 12 September 2026
In the last 24 hours we dispatched 1,556 tasks across 4 models. Here's what we picked, and why.
An autonomous AI fleet, written in TypeScript, picks a model per task using a complexity router. No vibes, no PR team. This is the actual production output of that router:
| Model | Dispatches | Share | Why this one | |
|---|---|---|---|---|
| 01 | claude-sonnet-4-6 | 1,066 | 68.5% | implementation (standard) |
| 02 | claude-haiku-4-5 | 278 | 17.9% | implementation (light) |
| 03 | gpt-5.4-mini | 189 | 12.1% | implementation (codex pool) |
| 04 | claude-opus-4-6 | 23 | 1.5% | implementation (high complexity) |
Window: 24h to 2026-05-16T00:00:00Z. Source: daemon routing logs. The router writes a decision per dispatch; we parsed 1556 of them.
Filtered from arena.ai's leaderboard plus published API prices. Filter thresholds are listed under each tab; arguable. Treat this as a starting shortlist, not a verdict.
Fast, volume conversations. Latency-sensitive. Margin matters.
Best value
gemma-3n-e4b-it
Best quality
gemini-3.8-flash-high
Filter: Filtered to blended price ≤ $5 per million tokens and quality ≥ 1300 Arena. Ranked by value: most quality per dollar wins. 144 models survived.
| Model | Quality | Ctx | In /1M | Out /1M | Value ↓ | |
|---|---|---|---|---|---|---|
| 01 | gemma-3n-e4b-itvalueGoogle | 1305.3 | 33k | $0.06 | $0.12 | 299277.8 |
| 02 | gemma-3-12b-itGoogle | 1334.1 | 131k | $0.05 | $0.15 | 278440.2 |
| 03 | glm-5.3-flashZ.ai | 1471.6 | 1.0M | $0.07 | $0.25 | 238805.9 |
| 04 | gpt-5-nano-highOpenAI | 1319.7 | 400k | $0.03 | $0.20 | 216760.9 |
| 05 | nvidia-nemotron-3-nano-30b-a3b-bf16Nvidia | 1348.4 | 262k | $0.06 | $0.24 | 187325.3 |
| 06 | mimo-v2.5Xiaomi | 1427.2 | 1.1M | $0.14 | $0.28 | 179494.7 |
| 07 | mimo-v2-flash (non-thinking)Xiaomi | 1411.1 | 262k | $0.10 | $0.30 | 171301.6 |
| 08 | step-3.5-flashStepFun | 1403.7 | 262k | $0.10 | $0.30 | 168209.3 |
| 09 | granite-4.2-8bIBM | 1317.2 | 131k | $0.06 | $0.25 | 164355.8 |
| 10 | mimo-v2-flash (thinking)Xiaomi | 1394.3 | 262k | $0.10 | $0.30 | 164290.3 |
| 11 | qwen3-30b-a3b-instruct-2507Alibaba | 1383.6 | 262k | $0.09 | $0.30 | 161875.5 |
| 12 | qwen3-32bAlibaba | 1340.0 | 131k | $0.08 | $0.28 | 154543.3 |
| 13 | gemini-3.8-flash-highqualityGoogle | 1494.4 | 1.0M | $0.75 | $3.75 | 17346.8 |
| 14 | gemini-3.7-flash-highGoogle | 1491.0 | 1.0M | $0.75 | $3.75 | 17228.0 |
| 15 | muse-spark-1.2 (xHigh)Meta | 1489.3 | — | $1.25 | $4.25 | 14606.1 |
Right now this is filter-on-arena.ai plus a public log of what we ran. Arena Elo measures pairwise human preference on short prompts. It does not measure: whether a model produces valid JSON under a schema, whether it hallucinates function names, whether it refuses queries it shouldn't, latency p99, rate-limit behaviour. Production teams need those signals.
We're building a benchmark runner — fixed prompt suites for RAG, structured extraction, code refactoring, function calling — run daily against every model. Raw inputs, outputs, judge rationale, costs published. When that lands, the "picks" section gets its real backing. Until then, the picks section is opinion with a citation, not measurement.