Friday, 21 August 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
solar-pro4
Best quality
claude-opus-5-high
Filter: Quality ≥ 1350 AND context length ≥ 128k. Sorted by value because you're answering thousands of queries. 140 models survived.
| Model | Quality | Ctx | In /1M | Out /1M | Value ↓ | |
|---|---|---|---|---|---|---|
| 01 | solar-pro4valueUpstage | 1376.7 | 524k | $0.03 | $0.12 | 405005.6 |
| 02 | gpt-oss-120bOpenAI | 1365.5 | 131k | $0.03 | $0.17 | 285574.5 |
| 03 | qwen3-30b-a3b-instruct-2507Alibaba | 1384.3 | 262k | $0.05 | $0.19 | 256952.5 |
| 04 | mimo-v2.5Xiaomi | 1427.3 | 1.1M | $0.14 | $0.28 | 179530.8 |
| 05 | mimo-v2-flash (non-thinking)Xiaomi | 1411.1 | 262k | $0.10 | $0.30 | 171300.5 |
| 06 | step-3.5-flashStepFun | 1403.9 | 262k | $0.10 | $0.30 | 168278.3 |
| 07 | mimo-v2-flash (thinking)Xiaomi | 1394.7 | 262k | $0.10 | $0.30 | 164471.1 |
| 08 | gemma-4-31bGoogle | 1441.8 | 262k | $0.14 | $0.40 | 137212.3 |
| 09 | gemini-2.5-flash-lite-preview-09-2025-no-thinkingGoogle | 1379.3 | 1.0M | $0.10 | $0.40 | 122344.5 |
| 10 | gemini-2.5-flash-lite-preview-06-17-thinkingGoogle | 1368.9 | 1.0M | $0.10 | $0.40 | 118994.0 |
| 11 | glm-4.7-flashZ.ai | 1352.9 | 203k | $0.06 | $0.40 | 118435.7 |
| 12 | deepseek-v3.2DeepSeek | 1424.7 | 164k | $0.27 | $0.40 | 117736.5 |
| 13 | claude-opus-5-highqualityAnthropic | 1505.1 | 1.0M | $5.00 | $25.00 | 2658.4 |
| 14 | claude-opus-5-maxAnthropic | 1503.5 | 1.0M | $5.00 | $25.00 | 2649.9 |
| 15 | claude-opus-4-6-highAnthropic | 1502.8 | 1.0M | $5.00 | $25.00 | 2646.1 |
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.