Inference for coding agents.
Run GLM-5.3 and DeepSeek V4.1 1017 in OpenCode, Command Code, Claude Code or your own harness. Long sessions finish 2x faster at 50–80% lower cost.

More accurate on DeepSWE.
On the DeepSWE benchmark for long-horizon agentic coding, runtime compression and ASpec raised GLM-5.3 from 60.6% to 67.6% and DeepSeek V4.1 Flash from 67.8% to 71.8%.
DeepSWE (Agentic Coding Benchmark)
Tasks resolved
Set up your harness in one command.
Point the harness you already run at Subconscious. Every model is served in both the OpenAI and the Anthropic format.
OpenCode
# Install the Subconscious CLI
$ npm install -g subconscious-cli# Log in
$ subc login# Launch OpenCode
$ subc opencodeQuestions
- GLM-5.3 is the default for most harnesses. On Subconscious it runs as GLM-5.3 Marathon, served for long sessions. DeepSeek V4.1 1017 Marathon is a cheaper, high-throughput option that also takes image input.
- Yes. Any harness that speaks the OpenAI or Anthropic API works by changing the base URL, key and model. Setup guides cover OpenCode, Command Code, Claude Code, Codex, Factory Droid, Cline, Kilo Code and more.
- Yes. Call the API from your backend, or let your users bring their own key. As usage grows, move to a Dedicated Deployment with its own GPUs and no rate limits.
- The runtime compresses context that stopped mattering before it is billed, so a long session stops paying to reread its own history. The same open model costs 50–80% less than on standard inference.
- Yes. Team Token Plans give your team 5x the usage of paying per token, for any coding agent or harness, at a fixed monthly price.
Put your coding agent in Marathon Mode.
Start on the API. Move to a Dedicated Deployment in our cloud or on your own GPUs when you need it.