Routing Matrix
| Task | Primary | Open-weight alt | Avoid | Why |
|---|---|---|---|---|
| Long-horizon multi-tool agent | Opus 4.8 Marathon 26.0 | Qwen3.7 Max 35h proven | GLM-5.2 Tool-Decathlon ✗ DS V4 Pro DeepSWE 8% | Sustained tool chaining is the hardest agent task |
| Terminal-agent loops | Sonnet 5 ★ 80.4 Fable 5 84.3 | GLM-5.2 81.0 Gemini 3.5 Flash 76.2 MiniMax M3 66.0 (v) | Gemini 3.1 Pro 54.2 Mistral Large 3 12 | Sonnet 5 = new default. Fable 5 ceiling but classifier risk |
| Single-pass code review | Opus 4.8 SWE-Pro 69.2 | DS V4 Pro 76.2 audit GLM-5.2 62.1 Step 3.7 56.3 ($1.15/M) Step 3.5 74.4 V ($0.30/M) | — | Step 3.7 = #2 on Pro at launch, budget pick. Step 3.5 = cheapest coding-per-$ |
| Monorepo analysis | Gemini 3.1 Pro 2M ctx | Any 1M-context model | <1M ctx | Context window is the gate |
| Frontend / web dev | Gemini 3.1 Pro WebDev #1 | Qwen3.7 Max Hy3 blind eval win | — | Hy3 won blind eval on frontend tasks |
| High-volume batch / cheap | DS V4 Flash $0.28/M | Agnes 2.0 $0.20/M Step 3.5 $0.30/M Step 3.7 $1.15/M | Opus/GPT cost | Agnes = new price floor, Step = free tier |
| Security review | Opus 4.8 | GLM-5.2 IDOR 39% vs CC 32% | Gemini 3.1 Pro | GLM-5.2 independent winner on one test |
| Planning / spec mode | Sonnet 5 GDPval 1618 Fable 5 if available | Opus 4.8 fallback | — | Sonnet 5 = new default plan model. Fable 5 ceiling w/ classifier risk |
| Long-horizon autonomous | Opus 4.8 Marathon 26.0 | Qwen3.7 Max 35h / 1158 calls | DS V4 Pro 8% DeepSWE | Only Qwen3.7 has public 1000+ call evidence |