The LLM API for long-running agents.
GLM-5.3, DeepSeek V4.1 1017 and Nemotron in Marathon Mode, in the OpenAI and Anthropic formats. Long runs finish 2x faster at 50–80% lower cost, with 5M+ tokens of context.
AI SDK
LangGraph
Mastra
Pydantic AI
CrewAI
Agno
TemporalSwitch in two lines.
Every model is served in both the OpenAI and the Anthropic format. Keep the SDK your agent already uses and point it at Subconscious.
API format
Language
Endpoint
https://api.subconscious.dev/v1/chat/completions
Model
subconscious/glm-5.3-marathon
Works with
OpenAI SDK
Intact
Tool calls, streaming, structured output
from openai import OpenAI
client = OpenAI(
base_url="https://api.subconscious.dev/v1",
api_key="YOUR_API_KEY",
)
response = client.chat.completions.create(
model="subconscious/glm-5.3-marathon",
messages=[
{
"role": "user",
"content": "Refactor the billing module and update every affected test.",
}
],
)
print(response.choices[0].message.content)More finished sessions per dollar.
An agent resends its history on every step, so a long session gets expensive fast. Marathon Mode compresses stale context before it is billed and caches the rest, so each step costs less.
$1,000 of long agent sessions buys
Assumes one session is 750 steps. Subconscious allows more workloads with faster speed and higher accuracy due to context compression and highly efficient caching.
Speed and accuracy that hold deep into a run.
Compressing stale context keeps generation fast and gives the model a cleaner view of the task. On the DeepSWE long-horizon benchmark, compression and ASpec raised GLM-5.3 from 60.6% to 67.6% of tasks resolved.
DeepSWE (Agentic Coding Benchmark)
Tasks resolved
Pay per token. No commitment.
Prices per million tokens. Running agents all day? A Team Token Plan buys more usage for the same budget.
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| GLM-5.3 MarathonFrontier coding and agentic work | $1.40 | $0.26 | $4.40 |
| DeepSeek V4.1 1017 MarathonMultimodal & high-throughput | $0.30 | $0.007 | $1.20 |
Questions
- One whose cost and speed hold as the session grows. Subconscious serves GLM-5.3, DeepSeek V4.1 1017 and Nemotron in Marathon Mode. Against standard inference providers, long runs finish 2x faster at 50–80% lower cost, with 5M+ tokens of effective context.
- Yes. Every model is served at an OpenAI chat completions endpoint and an Anthropic messages endpoint. Point the SDK you already use at Subconscious and change the model name. Tool calls, streaming and structured output work as before.
- You pay per token, with no monthly commitment. GLM-5.3 Marathon costs $1.40 per million input tokens, $0.26 cached and $4.40 output. Teams running agents all day can switch to a Team Token Plan for 5x the usage.
- Most providers serve each request on its own. Subconscious is built for the agent loop: the runtime compresses stale context and caches the rest, so cost and speed hold past 200k tokens. Our GLM-5.3 rate matches Z.ai’s list price, so every dollar saved comes from the runtime.
- Runs keep their history past the model’s own window, up to 5M+ tokens of effective context, without being compacted into a summary that drops what mattered.
- We do not store your prompts or completions. The API keeps only token counts for billing. Teams with stricter rules can also deploy on their own infrastructure.
- Yes. A Dedicated Deployment runs any open model, or your own, with no rate limits, as a Single-Tenant API in our cloud or an On-Prem Deployment on your GPUs. It is licensed per GPU.
Put your agents in Marathon Mode.
Mint a key, change the base URL, and run your next session on Subconscious.