Spojit uses two kinds of models:
- Chat / agent models power agent mode in connector nodes and the AI chat sidebar.
- Embedding models turn documents into vectors so the Knowledge feature can do semantic search.
You can choose which model to use per connector node and per knowledge collection, or leave the defaults.
Chat / agent models
Anthropic Claude
| Model | Context | Strengths | Notes |
|---|
| Claude Fable 5 | 1M tokens | Most capable Claude: deepest reasoning, long autonomous agent runs | Flagship tier above Opus. Premium pricing. |
| Claude Opus 4.8 | 1M tokens | Top-tier reasoning, tool use, multi-step problem solving | Latest Opus release. Recommended for complex workflows. |
| Claude Opus 4.7 | 1M tokens | Same Opus-class capability as 4.8 | Previous-generation Opus. |
| Claude Opus 4.6 | 1M tokens | Opus-class capability, one generation older | Current default for new workflows. |
| Claude Opus 4.5 | 200K tokens | Earlier Opus generation | Kept available for comparison; you generally want 4.6 or later. |
| Claude Sonnet 5 | 1M tokens | Most capable Sonnet yet: near-Opus reasoning built for coding and agents, at Sonnet speed and cost | Latest Sonnet. A strong balance of capability and cost for high-throughput agent steps. |
| Claude Sonnet 4.6 | 1M tokens | Balanced quality and speed at lower cost than Opus | Previous-generation Sonnet; good middle-ground for high-throughput agent steps. |
| Claude Haiku 4.5 | 200K tokens | Fastest and lowest cost in the Claude family | Best for simple steps, lookups, and high-volume runs. |
Google Gemini
| Model | Context | Strengths | Notes |
|---|
| Gemini 3.1 Pro | 1M tokens | Google’s most powerful agentic and coding model; state-of-the-art reasoning | Currently in preview, so capabilities may change. |
| Gemini 2.5 Pro | 1M tokens | Strong general-purpose reasoning | Stable alternative to Claude Opus when you prefer Google. |
| Gemini 2.5 Flash | 1M tokens | Very fast and very low cost | Best for high-volume simple steps. |
Embedding models
Embedding models convert text into high-dimensional vectors that capture semantic meaning. Spojit uses them whenever you upload a document to a Knowledge collection, and again at query time so the same vector space is used for search.
| Model | Dimensions | Strengths | Notes |
|---|
| Gemini Embedding 001 | 3072 | Higher-quality semantic search; best recall on nuanced queries | Default for new collections. Higher cost per token. |
| Text Embedding 004 | 768 | Faster and much cheaper per token | Good choice for high-volume document ingestion when peak quality isn’t required. |
Once a collection is created with a given embedding model, every document and query in that collection must use the same model, because vectors from different models aren’t comparable. Choose your embedding model when you create the collection.
Where model selection appears
Connector nodes (agent mode)
When configuring a connector node in agent mode, the Model dropdown in the properties panel lets you pick which model runs that step. If you don’t select one, the default model is used.
Each connector node can use a different model, so you’re not locked into one choice for the entire workflow.
The chat sidebar has a model picker in the header bar. Select a model from the dropdown to change which model powers your chat session. Your selection persists across page reloads.
Choosing a model
- Start with the default. Claude Opus 4.6 is a strong all-rounder and a good starting point for most workflows.
- Reach for Opus 4.8 for the hardest reasoning steps: multi-tool chains, ambiguous instructions, or long-context tasks.
- Reach for Claude Fable 5 when wrong answers are expensive. It offers the deepest reasoning available and excels at long autonomous agent runs; the premium pricing is worth it for the hardest problems.
- Use Sonnet 5 for a strong balance of capability and cost: near-Opus reasoning at Sonnet speed and price, well suited to coding and agent workflows.
- Use Sonnet 4.6 or Gemini 2.5 Flash for high-throughput steps where cost and latency matter more than peak reasoning quality.
- Use Haiku 4.5 for the simplest steps: short lookups, classification, simple formatting.
- Mix models in a single workflow. Different steps can use different models, so pick the right tool for each step rather than committing to one model for the whole flow.
