Benchmarks

Which hardware runs local LLMs fastest per euro in 2026? Comparison table, tokens/s by model and hardware, our own DGX Spark benchmark and cost per token vs. cloud.

GPT-6 Astra costs exactly the same as Fable 5.1 on the API. But 62.7% vs. 99.9% on ARC-AGI-3 shows that the headline number belongs to the harness. With a reading strategy and decision matrix for agent builders.

Qwen3.8-27B runs on a surprising range of hardware – but how fast, really? Baseline vs. community-tuned benchmarks (as of Aug 19, 2026): an RTX 3090 pushing ~1,000 tok/s across 64 parallel streams, an RTX 5090 at 148 tok/s, a DGX Spark at 210 tok/s aggregate, DFlash 2 at 3.4x on the H200 – plus RTX 5080 laptop, Mac mini, MacBook Pro, Mac Studio and AMD Strix Halo with prefill, decode and concurrency.

Two API settings moved GPT-5.6 Sol from 13.3% to 38.3% on ARC-AGI-3 with no weight change. Why the harness, not the model, decided that benchmark score.

GPT-5.6 Sol and Grok 4.5 both landed cheaper than Fable 5 in 48 hours. What the benchmark flip means for your coding-agent stack—and why cost-per-completed-task, not token price, decides the switch.

Claude Mythos Preview scored 92.1% on Terminal-Bench 2.1 with a 4-hour timeout, up from 82%. Here's why evaluation conditions matter more than the score — and what it means for enterprise AI teams.