Today we're shipping one of our most requested features: multi-model AI support. Pro and Max subscribers can now choose which AI model powers each task in Concept Sandbox, selecting from Claude Opus, Claude Sonnet, OpenAI's Codex, and Google's Gemini Pro.
Why Multiple Models Matter
Every AI model has strengths and weaknesses. Through months of testing and user feedback, we've learned that no single model is best at everything. Claude Opus produces the most thorough and well-structured PRDs, but it's slower and more expensive. Sonnet is fast and capable for quick critiques and iteration. Codex excels at generating technical specifications and code scaffolds. Gemini Pro offers a strong balance of speed and quality for general tasks at a lower cost.
Until now, Concept Sandbox used a single model for all AI tasks. That meant users were either paying for more power than they needed for simple tasks, or getting less capability than they wanted for complex ones. Multi-model support solves this by putting the choice in the user's hands.
How It Works
Every AI-powered action in Concept Sandbox now includes a model selector. When you generate a PRD, run an AI critique, or create a prototype scaffold, you'll see a dropdown next to the generate button. The dropdown shows available models with a brief description of each model's strengths.
Your model selection is remembered per task type. If you prefer Opus for PRD generation and Sonnet for critiques, Concept Sandbox will remember those preferences and default to them next time. You can always override the default for any individual task.
Model Availability by Tier
- Free tier: Sonnet (limited to 10 AI actions per month)
- Pro tier: Sonnet, Gemini Pro, and Codex (100 AI actions per month)
- Max tier: All models including Opus (unlimited AI actions)
We've structured access this way because Opus, while powerful, is significantly more expensive to run. Max tier subscribers get unlimited access to all models, making it the clear choice for power users who rely on AI throughout their workflow.
- Gemini Flash
- Sonnet
- Gemini Pro
- Codex
- All models incl. Opus
Our Recommendations
After extensive testing, here are our suggested model assignments:
PRD Generation — Claude Opus. PRDs require deep understanding of requirements, edge cases, and technical constraints. Opus consistently produces the most comprehensive and well-organized PRDs, often catching requirements that other models miss. If you're on the Max tier, this is where Opus shines brightest.
AI Critique — Claude Sonnet. Critiques need to be fast and actionable. Sonnet delivers sharp, focused feedback in seconds. The speed matters here because you'll often run multiple critique cycles as you refine a concept, and waiting 30+ seconds per cycle breaks your flow.
Auto-Fill from Brain Dump — Gemini Pro. The auto-fill feature extracts structured data (title, description, stage, tags) from freeform text. This is a straightforward extraction task where Gemini Pro performs comparably to more expensive models at a fraction of the cost.
Technical Scaffolding — Codex. When generating code-oriented content like API specifications, database schemas, or component hierarchies within a PRD, Codex's training on code gives it an edge. It produces more idiomatic technical specifications.
The Provider Abstraction
On the engineering side, multi-model support is powered by a provider abstraction layer in our codebase. Each AI provider (Anthropic, OpenAI, Google) implements a common interface with generate() for batch operations and stream() for real-time streaming. This means the rest of our application code doesn't need to know which model is running — it just calls the abstraction layer with a model identifier.
Streaming was particularly important to get right. All models now stream their output via Server-Sent Events through our /api/ai/stream endpoint. Whether you're using Opus or Gemini, you see the AI's output appear word by word in real time. This consistent streaming experience was one of our key engineering goals, and we're proud of how smooth it feels across all providers.
What's Next
We're monitoring usage patterns closely to understand which models users prefer for different tasks. This data will help us refine our recommendations and potentially introduce smart defaults that automatically select the best model based on the task and concept complexity. We're also watching for new model releases — when providers ship improvements, we'll add them to the selector as quickly as we can validate their output quality.
Multi-model AI is available now for all Pro and Max subscribers. Free tier users continue to have access to Sonnet. Log in and try different models on your next concept — we think you'll appreciate having the choice.