$

llm-rate

Tuesday, 23 June 2026

In the last 24 hours we dispatched 1,556 tasks across 4 models. Here's what we picked, and why.

What we ran

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:

ModelDispatchesShareWhy 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.

If you don't have a router, here are the picks per common task

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.

Reads code, finds bugs, explains tradeoffs. Mid-priced sweet spot.

Best value

qwen3-235b-a22b-thinking-2507

Alibaba · quality 1423.6 · $0.10/M blended

Best quality

claude-opus-4-6

Anthropic · quality 1535.3 · $19.00/M blended

Filter: Coding leaderboard, quality ≥ 1350. Sorted by value — for review you want diligence, not just absolute top. 177 models survived.

Model Quality Ctx In /1M Out /1M Value ↓
01 qwen3-235b-a22b-thinking-2507valueAlibaba 1423.6 262k $0.10 $0.10 423560.0
02 deepseek-v4-flashDeepSeek 1451.6 1.0M $0.09 $0.18 295169.9
03 qwen3-30b-a3b-instruct-2507Alibaba 1417.8 131k $0.05 $0.19 279322.1
04 gpt-oss-120bOpenAI 1380.5 131k $0.04 $0.18 276310.8
05 nvidia-nemotron-3-nano-30b-a3b-bf16Nvidia 1379.3 262k $0.06 $0.24 203903.2
06 mimo-v2.5Xiaomi 1467.4 1.0M $0.14 $0.28 196386.6
07 step-3.5-flashStepFun 1435.3 262k $0.09 $0.30 183658.2
08 mimo-v2-flash (non-thinking)Xiaomi 1440.6 262k $0.10 $0.30 183583.3
09 mimo-v2-flash (thinking)Xiaomi 1417.2 262k $0.10 $0.30 173833.3
10 qwen3-32bAlibaba 1358.2 131k $0.08 $0.28 162836.4
11 mistral-small-2506Mistral 1363.2 32k $0.10 $0.30 151341.7
12 deepseek-v3.2-thinkingDeepSeek 1453.0 131k $0.23 $0.34 146658.9
13 claude-opus-4-6qualityAnthropic 1535.3 1.0M $5.00 $25.00 2817.6
14 claude-opus-4-6-thinkingAnthropic 1534.2 1.0M $5.00 $25.00 2811.7
15 claude-fable-5Anthropic 1530.3 1.0M $10.00 $50.00 1395.6

What this is, and isn't

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.