Subconscious

Teams building coding and cybersecurity agents need models to work across hundreds of steps and hundreds of thousands of tokens, but standard inference engines are built for chatbots and one-shot tasks. As agent context windows fill with tool calls and results that stop mattering a few turns later, standard providers hit a context compression ceiling, process redundant tokens, and models get noisier and less accurate the longer a session runs. This costs developers in both accuracy and spend, since standard providers bill for tokens sent rather than tokens that actually matter, and periodic compaction doesn't resolve the underlying inefficiency.

Subconscious is an inference provider that treats long-running agent sessions, not chatbot queries, as its primary workload. Its runtime scores context messages for relevance as an agent runs, compresses low-relevance spans in place on the GPU, and caches tokens on both sides of each compressed span so the thread keeps hitting the cache, extending effective context past 5M tokens. The company reports this approach completes tasks 2x faster and processes 50-80% fewer tokens than general-purpose runtimes, with compression also improving accuracy by giving the model a cleaner view of the task (shown on the DeepSWE agentic coding benchmark with GLM 5.2). The API is OpenAI and Anthropic compatible, works natively with coding agents like GitHub Copilot, Claude Code, Codex, Pi, and OpenCode, and offers models including GLM-5.3 Marathon and DeepSeek V4.1 1017 Marathon, plus Dedicated Deployments (single-tenant or on-prem) for custom workloads. Billing is based on tokens actually processed after compression, not tokens sent, and the company states it does not store or train on customer prompts or outputs.