STRATUS X1

A World Model to unlock the next generation of Agents.

Loading...
The Agentic World Model Layer

while LLMs predict tokens..

Stratus predicts worlds

STRATUS is a world model built for agents. It sits above LLMs, not replacing them, supplying the predictive intelligence needed for reliable planning, action, and coordination.

JEPA World Model

Same SDK.
Smarter agent.

Drop in. Don't look back.

Higher accuracy on complex tasks
+24 F1 on reasoning benchmarks
Fewer tokens, lower cost
68% token reduction on average
More successful task completions
Up to +150% task success
Faster end-to-end responses
4× faster in production flows

One line. Same SDK.

Standard SDK
import OpenAI from "openai"const client = new OpenAI({  apiKey: process.env.OPENAI_API_KEY})

Change: baseURL + Stratus API key

With Stratus
import OpenAI from "openai"const client = new OpenAI({  apiKey: process.env.STRATUS_API_KEY,  baseURL: "https://api.stratus.run/v1"})

Works with OpenAI, Anthropic, Gemini, and any compatible API.

Benchmarks

Agents that see before they act

STRATUS X1 predicts state transitions in embedding space, giving agents the foresight to simulate, plan, and validate before execution. Task success doubles. Real autonomy emerges.

WebArena Task Success Rate

Real-world web automation tasks. STRATUS X1 boosts navigation, interaction, and completion rates.

Claude Sonnet 4
+105%
Baseline
20%
With X1
41%
GPT-4o
+100%
Baseline
17%
With X1
34%
Llama 4 Scout
+150%
Baseline
12%
With X1
30%
DeepSeek V3
+93%
Baseline
15%
With X1
29%

HotpotQA Performance

Multi-hop reasoning benchmark. STRATUS X1 lifts both EM and F1 by simulating outcomes before responding.

DeepSeek V3
Exact Match
+68%
37%vs62%
F1 Score
+48%
0.46vs0.68
GPT-4o
Exact Match
+55%
44%vs68%
F1 Score
+46%
0.54vs0.79
Claude Sonnet 4
Exact Match
+51%
41%vs62%
F1 Score
+57%
0.49vs0.77
Llama 4 Scout
Exact Match
+66%
35%vs58%
F1 Score
+56%
0.43vs0.67

Increases completions, lowers token usage, and provides near-zero latency.

0.0×
Task Success Rate
0%
Token Reduction
+0%
Exact Match Improvement
<0ms
Inference Latency
Compatible with:OpenClawMiniMaxClaudeGPTKimiGeminiand more
Predict worlds,
not words..
Architecture

Six ideas. One engine.

STRATUS X1 is built on a chain of interlocking concepts — each one a prerequisite for the next. Together they form the first practical world model for production agents.

01
State over tokens

World Model

Stratus X1 builds a continuous representation of the environment — not next-word probabilities, but next-state predictions in embedding space.

02
Joint Embedding Prediction

JEPA

The Joint Embedding Predictive Architecture learns structured representations by predicting the abstract embedding of a target state — not the raw pixels or tokens.

03
See before you act

Prediction

Before the agent touches the environment, X1 forecasts outcome embeddings across candidate action sequences — converting blind execution into foresighted planning.

04
Branch possible worlds

Simulation

X1 rolls out multiple futures in latent space simultaneously. Each branch is a simulated world — explored in under 10 ms, before a single real action is taken.

05
Rank trajectories

Comparison

A learned value function scores each simulated trajectory against the task objective. The agent selects the highest-ranked path — not the first plausible one.

06
Commit with confidence

Validation

After execution, X1 verifies the resulting state matches the predicted embedding. Drift triggers re-planning — turning one-shot agents into self-correcting systems.

Infrastructure

Latent Encoder
Compresses observations into compact state vectors
Trajectory Planner
Searches action sequences in embedding space
Value Network
Scores candidate paths against task objectives
State Monitor
Tracks drift between predicted and actual states
Context Fusion
Merges tool outputs, memory, and observations
Replay Buffer
Stores successful trajectories for few-shot transfer
Integration

Bring Nothing,
Build Everything

No provider signups. No API keys to juggle. Hit any OpenAI, Anthropic, or Google model instantly — Formation routes it.

0
Provider accounts required
Start immediately
All models
OpenAI, Anthropic, Google
Via Formation's pool
BYOK
Bring your own key
Zero markup, anytime
import { StratusClient } from '@formthefog/stratus-sdk-ts';

const client = new StratusClient({
  apiKey: process.env.STRATUS_API_KEY!,
});

const response = await client.chat.completions.create({
  model: 'stratus-x1ac-base-gpt-4o',
  messages: [
    { role: 'system', content: 'Current state: Homepage. Search box visible.' },
    { role: 'user', content: 'Search for wireless headphones' }
  ]
});

console.log(response.choices[0].message.content);
console.log(response.stratus?.overall_confidence);
console.log(response.stratus?.action_sequence);

The native SDK exposes response.stratus — action sequences, confidence scores, and world-model metadata not available through third-party SDKs.

Formation Pool
Any model, instantly
25% markup on LLM cost — Formation covers your provider spend
Your Own Key
Direct to provider
No markup — pay your provider at their rate, add anytime

Every API response includes key_source and formation_markup_applied — fully transparent billing.

X1