Subconscious raises $5.1M to build an inference platform for long-horizon agents

CShorten30 · x · 2026-09-22

Subconscious, a MIT-born startup, announced a $5.1M raise to build an inference platform purpose-built for long-horizon agents.

Thesis: standard inference engines are inefficient for multi-step agent workloads that burn hundreds of thousands of tokens; Subconscious designed a runtime for tasks running past 200k tokens, via API or self-hosting.
Market data cited: inference usage growing 2,500% YoY, agents consuming most tokens, open models nearly closing the gap with closed ones.
Claimed results on the same chips and models: 3.5x faster sustained throughput, 80% lower cost on long tasks, up to 10% accuracy gains.
Billing example: 750K effective context, billed tokens compressed from 281.6M to 96.2M — 6.3x cheaper than Opus 5 and 2.9x cheaper than standard GLM-5.3 per agent run.
Serves open models with a "Marathon" suffix, including Qwen3.8 27B, Nemotron 3 Ultra, GLM 5.3, and DeepSeek V4 Vision.

More from Venture