Models
All models route through model providers. The platform supports 167 models across 21 provider groups. The tables below highlight models available in the catalog. Use these IDs in your assemblies.
Hobby can use the free-tier catalog. Pro and Pro Max use the same complete paid catalog. Pro Max adds a larger monthly compute allowance and a higher round limit. It does not unlock extra models.
Paid catalog, frontier models
These give the best reasoning. Use them for high-stakes decisions, complex synthesis, and code review, where answer quality matters more than cost.
| Model ID | Notes |
|---|---|
openai/gpt-6.1-sol | 1.05M context; coding, document work, and reasoning. Thinking is always enabled. |
anthropic/claude-opus-5.5 | 1M context; complex reasoning and code review. Thinking is always enabled. |
x-ai/grok-4.7 | 500K context; coding, reasoning, and tool workflows. |
openai/gpt-6-astra | OpenAI's flagship computer-operator model. 1.05M context; agentic workflows and reasoning. |
x-ai/grok-4.6 | SpaceXAI frontier model. 500K context; text, image, and file input. |
deepseek/deepseek-v4-pro-0813 | DeepSeek V4 Pro GA release. 1M context and low list pricing. |
qwen/qwen3.8-max | Alibaba's flagship. 1M context, multimodal; strong on agentic and long-horizon work. |
anthropic/claude-fable-5 | Anthropic's autonomous knowledge-work model. 1M context and reasoning support. |
openai/gpt-5.5 | OpenAI's strongest. Strong general reasoning. |
openai/gpt-5.5-pro | Longer-context variant. |
anthropic/claude-opus-4.8 | Anthropic's strongest. Best at instruction-following and long-form synthesis. |
google/gemini-3.1-pro-preview | Google's strongest. Strong on multi-modal and search-augmented tasks. |
x-ai/grok-4.5 | Previous xAI frontier model. Good for coding and adversarial roles. |
x-ai/grok-4.3 | Previous xAI frontier model. Good for adversarial / red-team roles. |
deepseek/deepseek-v4-pro | Earlier DeepSeek V4 Pro release. Used as the default chairman. |
meta-llama/llama-4-maverick | Open-weights alternative. |
Paid catalog, efficient models
These models are strong and cost less. Most production assemblies use them.
| Model ID | Notes |
|---|---|
anthropic/claude-sonnet-5.5 | 1M context; everyday coding, writing, and document work. Thinking is always enabled. |
xiaomi/mimo-v2.6-pro | Xiaomi multimodal model with a 1,048,576-token provider context limit. |
meta/muse-spark-1.3 | 1M context; multimodal reasoning and long-running workflows. |
cohere/command-a-plus | 192K context; text and image input, tool calling, and structured outputs. |
moonshotai/kimi-k3 | Moonshot's 1M-context multimodal model for agentic work. |
meta/muse-spark-1.2 | Meta's previous 1M-context multimodal reasoning model. |
meta/muse-spark-1.1 | Previous Muse Spark. Still strong for multimodal panels. |
openai/gpt-5.4 | Solid general reasoning. |
openai/gpt-5.4-pro | Longer-context variant. |
anthropic/claude-sonnet-5 | Excellent instruction-following. Good all-rounder. |
google/gemini-3.1-pro-preview | Previous-generation Google frontier model. |
Hobby (Free) tier
These models are capable and cost the least. Use them for drafts, ideation, and everyday Q&A.
| Model ID | Notes |
|---|---|
openai/gpt-6-luna | 1.05M context; efficient drafting and lightweight agent workflows. Supports reasoning off. |
deepseek/deepseek-v4.1-flash | 1M context; low-cost reasoning with text and image input. |
google/gemini-3.8-flash | 1M context; multimodal reasoning and tool workflows. |
deepseek/deepseek-v4-flash | Earlier fast DeepSeek variant. |
openai/gpt-5.4-mini | OpenAI mini. |
openai/gpt-5.4-nano | OpenAI nano. Note: first-byte latency can exceed 15s on model providers. |
anthropic/claude-haiku-4.5 | Anthropic Haiku. |
google/gemini-3.7-flash | Google's fast multimodal model for agentic workflows, coding, and reasoning. |
google/gemini-3.6-flash | Google's efficient coding and agentic-workflow model. |
google/gemini-3.5-flash | Google Flash. |
google/gemini-3.5-flash-lite | Google's efficient model for focused agent and subagent tasks. |
qwen/qwen3.6-flash | Alibaba Qwen. |
Choosing models for a panel
A good panel has role specialization, not just three expensive models:
| Role | Good fit | Why |
|---|---|---|
| Anchor (sets the initial framing) | GPT-6 Astra, Opus 5.5, GPT-6.1 Sol, Gemini 3.1 Pro | Strong general reasoning. |
| Critic (finds holes) | Opus 5.5, Grok 4.7, Qwen3.8 Max, Gemini 3.1 Pro | Sharp, willing to disagree. |
| Synthesizer (writes the strongest final version) | GPT-6 Astra, Opus 5.5, Sonnet 5.5, Muse Spark 1.3 | Strong writing, good at synthesis. |
| Adversary (DA format) | Grok 4.7, Gemini 3.1 Pro | Willing to attack. |
| Attacker (Red Team) | Grok 4.7, Gemini 3.1 Pro | Good at finding creative misuse. |
| Defender (Red Team) | Opus 5.5, Sonnet 5.5 | Careful instruction-following. |
| Interrogator (Socratic) | GPT-6 Astra, Opus 5.5, GPT-6.1 Sol | Sharp question-asking. |
| Respondent (Socratic) | GPT-6 Astra, GPT-6.1 Sol, Sonnet 5.5 | Confident, will defend. |
For most assemblies, diverse providers > same-provider top models. A panel of GPT-6.1 Sol + Opus 5.5 + Gemini 3.8 Flash will disagree more than three copies of GPT-6 Astra.
Capabilities
Per-model controls:
webSearch: true. The model can issue web searches during its turn. It adds 30-60% per-turn cost. Use it for current research. Skip it for "rewrite this" tasks.reasoningEffort. Choose an effort supported by the selected model. The default is low. Mandatory-thinking models do not offer off; GPT-6 Luna supports off. Reasoning tokens increase usage cost.role: "...". A free-form string. It appears in the UI transcript for your own bookkeeping. Common values:INITIATOR,REVIEWER,SYNTHESIZER,PROPOSITION,OPPOSITION,ATTACKER,DEFENDER,JUDGE,INTERROGATOR,RESPONDENT.
Cost reference
The model and tier determine per-token cost. Frontier models are typically 5-10× the cost of Hobby models. A turn costs roughly 0.5-1.5× the input + output tokens of the turn.
For typical usage, a 3-model Refinement debate costs:
- 3 frontier models, 3 rounds, ~12k tokens total: ~$0.40-0.80
- 3 pro models, 2 rounds, ~8k tokens total: ~$0.15-0.30
- 3 hobby models, 1-2 rounds, ~6k tokens total: ~$0.03-0.08
See cost optimization for every way to lower cost.
Latency notes
Models have different latencies on model providers. A few cases to watch:
- OpenAI nano models can take 15-30s to first byte. That delay is fine for async work and painful for synchronous UX.
- Gemini 3.1 Pro is fast.
- Claude Opus is fast but can slow down during peak times.
- DeepSeek V4 Pro, the default chairman, is fast and reliable.
If a specific model keeps timing out, set a custom timeout in its model config or swap to a different model.
What's not here yet
- Open-source model selection. We currently surface a curated set. Contact us for access to the full model providers catalog.
- Fine-tuned / custom models. These are for high-volume enterprise customers. Contact us.
- Input support depends on the selected model. The catalog includes multimodal models; upload and processing limits still apply.
For the live list in the web app, see the model picker in the Assembly editor. It shows current availability and per-tier access.