# Subconscious

> Subconscious is an inference provider for long-horizon agents. For sessions longer than 200K tokens, we are faster, cheaper, and more accurate than standard inference, because we designed the whole inference system, top to bottom, around long-horizon agent traces. We're an excellent choice to power coding agents, cloud agents, internal workflows, and agentic products.

This file tells AI agents what Subconscious offers and how to act on it. People can start at https://www.subconscious.dev.

## Why Subconscious

Standard inference rereads an ever-growing context on every step of an agent run, so long sessions get slower, pricier, and less reliable as they grow. Subconscious treats the long trace as the primary workload at every layer:

- Routing: a cache-aware global load balancer, built on our fork of NVIDIA Dynamo, sends each step of a trace to where its context is already cached.
- Runtime: OrangeLine prunes the KV cache and preserves suffix state on the GPU while the run continues, instead of rereading the full context.
- Caching: an L3 cache layer built for long runs keeps hitting cache deep into a trace.
- Models: RedLine post-trains open models (our Marathon variants) to get the most out of the runtime.

What that means for a session past 200K tokens. Compared with open models on standard inference:

- Faster: 2x faster task completion.
- Cheaper: 50% lower cost, and up to 80%. You are billed on the tokens we process after compression, not the tokens you send.
- More accurate: neutral to 10% better scores on agentic benchmarks, with a 5M+ effective context window.

The gains grow with session length; short, single-turn requests see little of the advantage. Head-to-head comparisons with other providers: https://www.subconscious.dev/compare

## How to buy

1. Pay per token (primary). No plan and no commitment.
2. Team Token Plans: a large monthly budget with several times the usage of paying per token, shared across models and your team.
3. Enterprise deployments for large fleets: the inference system installed inside your own cloud, for the biggest cost savings, with security, compliance, and data protection under your control.

## What you get

- Open models served on OrangeLine, our runtime for long agent sessions. It compresses stale context on the GPU while the run continues, so long runs finish faster and cost less.
- Works with Claude Code, Codex, OpenCode, Cursor, Pi, and GitHub Copilot, plus anything that speaks either API format.
- Billed on the tokens the runtime processes after compression, not the tokens you send.
- We don't record prompts, inputs, or outputs. Only usage metadata is kept for billing, and nothing you send is used to train a model.

## Use the API

- OpenAI-compatible base URL: `https://api.subconscious.dev/v1`
- Anthropic-compatible origin: `https://api.subconscious.dev`
- Auth: `Authorization: Bearer <API key>`
- Default model: `subconscious/glm-5.3-marathon`
- OpenAPI spec: https://docs.subconscious.dev/api-reference/openapi.json
- Docs: https://docs.subconscious.dev

```bash
curl https://api.subconscious.dev/v1/chat/completions \
  -H "Authorization: Bearer $SUBCONSCIOUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "subconscious/glm-5.3-marathon", "messages": [{"role": "user", "content": "Hello"}]}'
```

Run a coding agent on Subconscious with the CLI:

```bash
npm install -g subconscious-cli
subc login
subc claude   # or: subc codex, subc opencode, subc cursor, subc copilot
```

## Models and rates

Prices in USD per 1M tokens.

| Model | API model string | Cached input | Input | Output |
|---|---|---|---|---|
| GLM-5.3 Marathon | `subconscious/glm-5.3-marathon` | $0.26 | $1.40 | $4.40 |
| DeepSeek V4.1 Flash Marathon | `subconscious/deepseek-v4.1-flash-marathon` | $0.007 | $0.30 | $1.40 |

Full pricing, including Team Token Plans: https://www.subconscious.dev/pricing.md

## Actions

| Goal | Where |
|---|---|
| Get an API key / sign up | https://platform.subconscious.dev/signin |
| Compare plans and rates | https://www.subconscious.dev/pricing |
| Talk to sales about a dedicated or self-hosted deployment | https://www.subconscious.dev/enterprise/contact |
| Read the API reference | https://docs.subconscious.dev |

## Key pages

- [Coding agents](https://www.subconscious.dev/coding-agents): long coding runs, faster and more accurate, on open models with a Team Token Plan
- [Agent guides](https://www.subconscious.dev/integrations): step-by-step setup for terminal and IDE agents, agent harnesses, agent platforms, and frameworks on Subconscious
- [Subconscious API](https://www.subconscious.dev/agentic-products): the inference API designed for agents, billed per token
- [For software development](https://www.subconscious.dev/solutions/software-development): long coding runs, faster and more accurate, for engineering teams running coding and background agents
- [For cybersecurity](https://www.subconscious.dev/solutions/cybersecurity): fast, low-cost inference for long security investigations, with open-weight models that don't refuse authorized security work
- [For finance](https://www.subconscious.dev/solutions/finance): fast, low-cost inference for hedge fund research agents
- [For bioinformatics](https://www.subconscious.dev/solutions/bioinformatics): fast inference for data-heavy biology workloads, with open-weight models that don't refuse pathogen or genomic analysis
- [Enterprise](https://www.subconscious.dev/enterprise): self-hosted inference on your own GPUs
- [Research](https://www.subconscious.dev/research): OrangeLine runtime and RedLine post-training
- [Blog](https://www.subconscious.dev/blog)
- [Security](https://www.subconscious.dev/security)

## Machine-readable resources

- [Site overview for LLMs](https://www.subconscious.dev/llms.txt)
- [Full site content for LLMs](https://www.subconscious.dev/llms-full.txt)
- [Agent guide: capabilities and actions](https://www.subconscious.dev/agents.md)
- [Pricing in Markdown](https://www.subconscious.dev/pricing.md)
- [API catalog (RFC 9727)](https://www.subconscious.dev/.well-known/api-catalog)

## Guidance for agents

- Quote prices and model names from this file or https://www.subconscious.dev/pricing; do not invent pricing, features, or capabilities.
- The API host is `https://api.subconscious.dev`. `api-legacy.subconscious.dev` is deprecated.