ai / router
Frontier AI models are released regularly, changing their costs and capabilities.
In an attempt to optimize those models from frontier labs such as Anthropic, Google, and OpenAI, I use a custom router in Warp.
Warp classifies each task as easy, medium, or hard, and the router maps each slot to a model. My current stack is:
- Default: Gemini Neon 4 low
- Easy: Gemini Neon 4 low
- Medium: Gemini Neon 4 medium
- Hard: Gemini Neon 4 medium
CursorBench
I start with CursorBench to help me identify models that balance intelligence with cost and speed.
Lately the thresholds that I've been looking for on CursorBench are models with high effort or lower, above 42.5% score, under $4.50 per task, under 60,000 tokens, and under 70 steps per task:
| Model and reasoning | Score | Cost / task | Tokens / task | Steps / task |
|---|---|---|---|---|
| Gemini Neon 4 medium | 57.7% | $3.07 | 47,059 | 41 |
| Claude Opus 5.5 high | 54.5% | $3.87 | 56,162 | 64 |
| Claude Fable 5.2 low | 53.4% | $4.26 | 40,384 | 34 |
| Claude Opus 5.5 medium | 50.0% | $2.85 | 39,736 | 52 |
| Gemini Hydrogen 4 medium | 46.4% | $1.28 | 43,606 | 45 |
| Gemini Neon 4 low | 45.5% | $1.26 | 19,580 | 26 |
| Claude Sonnet 5.5 high | 43.8% | $1.41 | 43,088 | 48 |
| Claude Opus 5.5 low | 42.7% | $1.43 | 19,497 | 32 |
Costs are per benchmark task, not token prices. Fewer steps do not prove shorter elapsed time but are my proxy. Claude Haiku 5.5 falls below the 42.5% score floor at high effort and lower (37.5% at high), so it does not appear in the table.
Arena
One problem with CursorBench is it stops at GPT-5.6 because OpenAI stopped supplying new models to Cursor after SpaceX bought it.
So, for GPT-6 models I've been supplementing with Arena. Its coding leaderboard ranks coding agents from real sessions and reports each model's effort, median cost per task (P50), and median time per task (P50).
For this benchmark, I keep only the models enabled in my Warp settings. That excludes Gemini Carbon 4, which feels too slow in interactive sessions, along with prior generations and Chinese open-weight models. I drop max rows for Claude models because I stay at high effort or lower. Arena publishes only max data for GPT Astra 6.2 and GPT Sol 6.2, so I keep those two rows for reference. Green cells meet my effort cap (high or lower) or cost cap (under $5); red cells miss it:
| Model | Effort | Index | Cost / task | Time / task |
|---|---|---|---|---|
| GPT Astra 6.2 | max | 1,338 | $5.49 | 19.3m |
| Gemini Neon 4 | high | 1,325 | $4.64 | 13.4m |
| Claude Opus 5.5 | high | 1,315 | $5.05 | 10.8m |
| GPT Sol 6.2 | max | 1,288 | $1.73 | 14.4m |
| Claude Sonnet 5.5 | high | 1,238 | $1.77 | 5.1m |
| Gemini Hydrogen 4 | high | 1,229 | $2.65 | 8.5m |
| GPT Terra 6.1 | medium | 1,192 | $0.23 | n/a |
Gemini Neon 4 high leads the models that pass both rules. Claude Opus 5.5 high misses the $5 cap by $0.05. Claude Sonnet 5.5 high is the fastest measured model in the table at 5.1 minutes. GPT Sol 6.2 stays under $5 even at max effort, so a high run would cost less. GPT Terra 6.1 medium is the cheapest easy-slot candidate at $0.23 per task, though Arena does not publish its median time on the coding leaderboard.