SWE-bench cost analysis: tiny Inkling 2 Mini costs 3-5x more per task than DeepSeek V4.1 Flash
hyudryu · reddit · 2026-11-22
A Redditor cost-audited the hyped Inkling 2 Mini. Locally it's excellent: 4500 tok/s decode with 32 concurrent coding sessions (140 tok/s each) on 2x RTX Pro 6000. But on a 200-problem SWE-bench subset across three harnesses (Kimi Code / Deepseek / OMP), the cost pattern is stark: pass rates 72.5-73.5% vs DeepSeek V4.1 Flash's 77.0-80.5%; 2x the model calls (16.7-19.3k vs 8.7-9.4k) and far more input tokens (e.g. 1.45B vs 529M under Kimi Code); and per-task API cost of $23.11-29.88 vs DeepSeek's $6.11-7.68 — the smaller model ends up 3-5x more expensive. Takeaway: small models' higher call counts and token burn can wipe out their price advantage.
More from coding & agent
- Walmart and Muse to launch first agent-to-agent agentic commerce integration over MCP — AccBalanced · 2026-11-22
- scientific-toolkit: one MCP server fronting nine repos that refuses to fabricate results — Emergency-Tutor-7 · 2026-11-22
- Zero-degradation Strata tuning: fixing layer_split lifts decode ~10-15% and halves CPU expert misses — PathfinderTactician · 2026-11-22
- Cross-provider adversarial review: hedge shared model blind spots — craigbalding · 2026-11-22
- Inference Stack Launches One-Command Self-Hosted LLM Serving with vLLM, LiteLLM and llm-d — AccBalanced · 2026-11-22
- Users complain GPT-5.5 writes bloatware and duplicates functions — yacineMTB · 2026-11-22