Inference for agents that just keep running.

Standard inference engines are a one-size-fits-all solution for chatbots, one-shot tasks, and agents. Agents are the hard case. They run for hundreds of steps and fill the context window with tool calls and results that stop mattering a few turns later. Our runtime scores those messages as the run goes, compresses the 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. The result is a context window that extends past 5M tokens, sustained token generation about 3.5x faster than a general-purpose runtime on the same hardware, and about 40% fewer tokens processed per run.

It helps. Long agent threads fill up with noise late in the run, and models get worse as that noise piles up. Clearing it out gives the model a cleaner view of the task. We see this on key benchmarks like DeepSWE, where GLM 5.2 scored higher on Subconscious than on a standard provider at a lower cost per task. We hear it from customers running production agents on our system every day. The models work great.

GLM 5.2, GLM 5.3, and our post-trained Qwen 3.8 27B on the managed API. All open-weight, so you can see exactly what's answering your requests. For on-prem deployments, we can support virtually any open model.

Yes. Every model is available through both the OpenAI chat completions format and the Anthropic messages format. If your code already uses either SDK, change the base URL and the API key and you're running on Subconscious. Three lines.

Almost certainly. We work natively with GitHub Copilot, Claude Code, Cursor, Codex, Pi, and OpenCode. Anything else that speaks OpenAI chat completions or Anthropic messages works too. Our docs have setup pointers for each one.

Two ways to pay. Pay as you go per token, or pick a token plan with a fixed monthly price and a daily token allowance. Current rates are on the pricing page. Dedicated endpoints and self-hosted deployments are priced separately.

Only the tokens our system actually processes. You might send us a message list with 1M tokens in it. Under the hood, after compression, we're processing and charging for 200k. That gap between what you send and what gets recorded is another way the savings end up on your invoice and not just in our benchmarks.

No. We don't record prompts, inputs, or outputs. We keep usage metadata such as token and request counts for billing. Nothing you send is used to train a model.

Yes. Dedicated GPUs, no rate limits, any supported model, custom SLAs, and optional post-training on your tools and data. Sign up and open the dedicated endpoint tab to get started.