Tuesday, 6 October 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.
Large context windows. Accurate retrieval. Per-query cost adds up.
Best value
glm-5.3-flash
Best quality
gemini-4-argon-high
Filter: Quality ≥ 1350 AND context length ≥ 128k. Sorted by value because you're answering thousands of queries. 155 models survived.
| Model | Quality | Ctx | In /1M | Out /1M | Value ↓ | |
|---|---|---|---|---|---|---|
| 01 | glm-5.3-flashvalueZ.ai | 1469.6 | 1.0M | $0.06 | $0.20 | 297230.2 |
| 02 | mimo-v2.6-flashXiaomi | 1456.4 | 1.0M | $0.14 | $0.28 | 191781.2 |
| 03 | mimo-v2.5Xiaomi | 1427.5 | 1.1M | $0.14 | $0.28 | 179624.7 |
| 04 | mimo-v2-flash (non-thinking)Xiaomi | 1411.0 | 262k | $0.10 | $0.30 | 171260.4 |
| 05 | step-3.5-flashStepFun | 1403.0 | 262k | $0.10 | $0.30 | 167904.1 |
| 06 | mimo-v2-flash (thinking)Xiaomi | 1395.4 | 262k | $0.10 | $0.30 | 164767.9 |
| 07 | qwen3-30b-a3b-instruct-2507Alibaba | 1384.1 | 262k | $0.10 | $0.30 | 160034.5 |
| 08 | solar-pro4Upstage | 1386.2 | 524k | $0.09 | $0.36 | 138433.1 |
| 09 | gemini-2.5-flash-lite-preview-09-2025-no-thinkingGoogle | 1379.2 | 1.0M | $0.10 | $0.40 | 122326.6 |
| 10 | gemini-2.5-flash-lite-preview-06-17-thinkingGoogle | 1369.3 | 1.0M | $0.10 | $0.40 | 119143.6 |
| 11 | glm-4.7-flashZ.ai | 1350.9 | 200k | $0.06 | $0.40 | 117687.6 |
| 12 | deepseek-v3.2-exp-thinkingDeepSeek | 1424.7 | 164k | $0.27 | $0.41 | 115418.6 |
| 13 | gemini-4-argon-highqualityGoogle | 1533.5 | 1.0M | $2.00 | $10.00 | 7019.2 |
| 14 | claude-opus-5.5-highAnthropic | 1511.7 | 1.0M | $4.00 | $20.00 | 3366.3 |
| 15 | claude-fable-5.1-maxAnthropic | 1510.6 | 1.0M | $10.00 | $50.00 | 1343.7 |
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.