LLM Analysis

Model-agnostic LLM nodes.
Claude, GPT, Gemini, in one workflow.

Pick the right model for each step. Run Anthropic Claude (Sonnet 4.6, Haiku 4.5, Opus 4.7), OpenAI GPT, Google Gemini side by side in the same agent. Tool use, vision input, structured JSON output, prompt versioning, per-node budgets. Your data lands in the prompt; the model never goes hunting for context.

Get a live workflow build → Back to Agents

app.demandsphere.com - LLM Node

Provider

Claude

GPT

Gemini

Model

claude-sonnet-4.6

Capabilities

Tool use Vision JSON output 1M ctx

Temperature 0.3

Prompt

Summarize this week's citation changes from the upstream DS Citations node. Return JSON.

Structured output

{
"wins": ["AI Overview re-cited", ...],
"losses": ["lost #1 on q4 guide"],
"summary": "3 wins, 1 loss..."
}

fallback → claude-haiku-4.5 1.19 AI credits / run

LLM node features

Every model, every capability, one configuration form.

Multi-provider, one workflow

Use Claude Sonnet for analysis, GPT for generation, Gemini for grounding, Haiku for cheap classification. Switch per node, pin versions for production.

Grounded by default

Upstream DS data nodes land in the prompt as structured context. No hallucinated facts about your domain.

Tool use mid-workflow

The model can call back into the workflow - fetch more data, run a sub-agent, ask for human approval - and resume with the result.

Structured JSON output

Define a schema, get validated JSON back. Downstream nodes consume typed fields, not free-form text. Reduces brittle string parsing to zero.

Vision input

SERP screenshots, chart images, page captures - pass them through Claude vision or GPT-4o for analysis. Useful for AIO mockup classification.

Budget controls + fallback

Per-agent credit caps. Auto-fallback to a cheaper model on retry. Per-node cost shown before run. No surprise bills.

Smart routing

Route by difficulty. Pay for what each task needs.

        TaskRouted modelRelative cost
      

Classify intent

Haiku 4.5 $

Weekly summary

Sonnet 4.6 $$

Competitive deep-dive

Opus 4.7 $$$

One workflow can mix all three - the router sends each task to the cheapest model that can do the job, and falls back automatically on error.

Compose with

Visual Builder →

The drag-and-drop canvas LLM nodes live on.

Sub-Agents →

Wrap your LLM logic as a callable sub-agent.

Citation Analytics →

The data primitives LLM nodes typically ground on.

Model Tracker →

Picking the right model? Benchmarks for every frontier release.

Common questions

Frequently asked

Anthropic Claude (Sonnet, Haiku, Opus), OpenAI GPT series, and Google Gemini. You can switch models per node, mix multiple models in one workflow, and pin specific versions for production reproducibility.

Each LLM node shows its credit cost before you run. You can set per-agent budgets, fall back to cheaper models on retry, and route easy tasks to Haiku-class models while sending the hard work to Sonnet or Opus.

Yes. LLM nodes support tool use - the model can pause the workflow to call back to a data source, a transform, or a sub-agent before continuing. Outputs flow back into the LLM context automatically.

That's the point. Connect a DS data source node (rankings, citations, crawl, GSC) upstream of the LLM node. The data lands in the prompt as structured context. You can also use a URL Loader, CSV Input, or any sub-agent output.

Run frontier LLMs on your data

Book a strategy session and we'll wire a Claude / GPT / Gemini node into a workflow on your real numbers.

Get a live workflow build → See Pricing

Model-agnostic LLM nodes.Claude, GPT, Gemini, in one workflow.