Rogue Constellation

Today, we are introducing Rogue Constellation, a new multi-agent system for complex software work. It orchestrates the best frontier models, delivering the world's most powerful agentic coding experience.

An implementation can look convincing and still miss a requirement, fail an edge case, or solve the wrong problem. Exploring alternatives and deciding which result to trust takes substantial effort.

Constellation starts five agents from the same state of your project. Each writes code and runs tests independently. It then compares their approaches and the evidence produced along the way to construct a single solution.

We evaluated it on two demanding benchmarks. Here’s how it performed.

Play three games built by Constellation, with graphics and sound generated entirely in code.

Full screenOpen in a new tab ↗

We recommend playing these games on a computer for the best experience.Tap to play anyway

Resonance. A rhythm shooter. Your shots, your kills and every enemy move to a score synthesized as you play.

Efficiency

Cost was one of the main concerns with Rogue DeepThink. We took that feedback seriously and made efficiency a central design priority for Constellation.

Usage reflects the combined work of all agents, including attempts that are discarded. The chart below compares accuracy and spending at different effort levels.

0405060705102050Cost per task (USD, log scale)Score (%)LowHighMax+19.1 pts54% lower costDeepThink Max → Constellation High

Sources: Claude Fable 5.1 · GPT-6 Astra

Availability and pricing

Constellation is available today on every paid Rogue plan.

These rates determine how much of your plan’s allowance a task consumes. Usage counts the tokens of all agents.

About the evaluation

Constellation was evaluated on all 66 tasks of Terminal-Bench 4.0 with five agents per task, and on all 70 tasks of Terminal-Bench Science 0.1 with three agents per task. The product runs five.

Results for GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 are reported figures. Settings and compute budgets differ, so these comparisons do not establish a ranking at equal cost. At these sample sizes, each Constellation score has a standard error of about 5.5 percentage points.

Six Science runs include steps where the model provider fell back to an earlier model version; we kept them. One Terminal-Bench 4.0 run returned no grade and was counted as a failure.

Constellation grew out of DeepThink, our earlier research on the same idea. It is an orchestration system, not a standalone AI model.