How Startica uses AI

Startica uses an AI model, with your own API key, for everything that needs judgment: understanding your product, scoring conversations, writing drafts, planning the roadmap, auditing your site and checking whether AI assistants recommend your product. It works with two providers: Claude (Anthropic) and OpenAI. You pick one in the welcome wizard or in Settings → General → AI. This page explains how the work is split across models, how calls are made and checked, and where you can see them.

Providers

Provider	Key	Where to get it
Claude (Anthropic)	Anthropic API key, starts with sk-ant-	Anthropic API key
OpenAI	OpenAI API key, starts with sk-	OpenAI API key

You need a key for the provider you use. You can save both. The AI card in Settings → General says which one is in use (“Using Claude (Anthropic).” or “Using OpenAI.”).

Each task is routed to a model, and the model decides which provider the call goes to. So with both keys saved, you can mix them: for example OpenAI for drafting and Claude for strategy. A call to a provider without a saved key fails with “No Anthropic API key: add one in Settings → General → AI (the ping task uses claude-haiku-5-5)” or the same for OpenAI.

Switching provider

In Settings → General → AI, click Change keys or budget (or Add API key), choose the other provider under Use, paste its key if it isn’t saved yet, and save. Every task still on the old provider’s models moves to the new provider’s recommended models (the defaults below). Tasks already on the new provider’s models keep their settings, so switching back and forth doesn’t lose your edits.

Tasks and models

Every AI call belongs to a task. Each task is routed to a model, an effort level and an output limit. These are the defaults for each provider:

Task	Claude default	OpenAI default	Effort	Used for
strategy	Claude Opus 5.5 (claude-opus-5-5)	GPT-6.1 Sol (gpt-6.1-sol)	high	The Intel brief (synthesis), voice profiles, the roadmap and weekly replan, audit findings, ad campaign plans
research	Claude Opus 5.5	GPT-6.1 Sol	high	Web research for the Intel brief and the GEO audit, using the provider’s web search
draft	Claude Sonnet 5.5 (claude-sonnet-5-5)	GPT-6.1 Sol	medium	Reply drafts, Content Studio pieces, topic suggestions, ad rewrites, suggested AI visibility questions, fix drafts on Opportunities, accessibility fix suggestions and summaries, and Security’s explain and fix, report import and triage of reported problems
answer	Claude Sonnet 5.5	GPT-6.1 Sol	low	AI visibility: answering your questions with web search, as an assistant would
vision	Claude Sonnet 5.5	GPT-6.1 Sol	medium	Reading images: the accessibility alt text review, and Ask AI on Needs review items, which reads the page screenshot
score	Claude Haiku 5.5 (claude-haiku-5-5)	GPT-6 Luna (gpt-6-luna)	low	Scoring each conversation for relevance and intent; reading each AI visibility answer
ping	Claude Haiku 5.5	GPT-6 Luna	low	The Run test job health check on the System page
image	GPT Image 2.5 Flare (gpt-image-2.5-flare)	GPT Image 2.5 Flare	medium (the image quality)	Drawing the pictures for ad creatives. OpenAI only.

Max output tokens default to 16,000 for strategy and research, 4,000 for draft, vision and answer, 1,024 for score and 256 for ping on Claude. OpenAI’s defaults are higher (32,000, 8,000, 2,048 and 1,024), because OpenAI counts reasoning tokens inside the output limit. The image task’s row shows 4,096, but the image call doesn’t use it: the picture’s size and quality set how many tokens it takes.

Why this split:

• The strongest default model for strategy and research. These calls are rare (a brief per product, a weekly replan, an occasional audit) and benefit most from careful reasoning over a lot of context: a whole site crawl, months of metrics. Effort is set to high explicitly.
• A mid-priced model for writing. Drafts and content are frequent and need good writing in a specific voice. Sonnet and GPT-6.1 Sol write well at half Opus’s price.
• The cheapest model for scoring. Scoring runs on every new conversation found, up to 60 per product from each source per hourly scan. It needs to be cheap and fast, and it returns just two scores and a short reason. AI visibility uses the same model to read each answer.
• A mid-priced model, at low effort, for answers. AI visibility’s answer task stands in for an AI assistant answering a customer’s question. A mid-priced model with web search is close to what assistants use, and the answer only needs to be a good everyday answer, not deep reasoning. This model also decides which provider’s assistant you’re measuring; see Models.

The image task

The image task draws the pictures for ad creatives, and only OpenAI’s image models can do it. It’s the same whichever provider you use for everything else:

• It needs an OpenAI API key, even when Claude is your provider. Without one, the Draw it with AI option on a creative is disabled (“needs an OpenAI key in Settings”); you can still upload your own pictures. See OpenAI API key. If you only want pictures from OpenAI, save both keys and keep Use on Claude.
• Switching provider doesn’t move it. It stays on its OpenAI image model.
• Effort is the image quality. The dialog calls it Quality: low, medium (the default), high, xhigh or max. Higher quality costs more; see Costs.
• Two models: GPT Image 2.5 Flare (gpt-image-2.5-flare, the default) and GPT Image 2.5 Sunburst (gpt-image-2.5-sunburst). They cost the same per token; Sunburst renders more carefully and so uses more output tokens per picture. The image task’s model list shows only these two, and the other tasks’ lists leave them out.

Every picture is one call, logged like any other with its cost, and counts against the monthly budget. The words, logo and button on an ad are added afterwards by a template on your computer, not by the AI.

The image task was added with ad creatives. Workspaces created before it get it on first start after updating.

The answer task was added with AI visibility. Workspaces created before it get it on first start after updating, on the provider their draft task uses.

Each provider also offers a top model that isn’t a default: Claude Fable 5.1 (claude-fable-5-1) and GPT-6 Astra (gpt-6-astra). Both cost 2.5 times as much as Opus. Try one for strategy only if the default plans seem weak to you.

See Costs for prices.

Changing models

1. Open Settings. The AI card lists each task with its model, effort and max tokens.
2. Click the pencil icon on a row. The dialog is titled Model for “task”.
3. Choose Model (the list has every text model from both providers, with each one’s price per million tokens; for the image task, the image models), Effort (low, medium, high, xhigh, max; called Quality for the image task) and Max output tokens (256 to 64,000).
4. Save. The change applies to the next job that runs this task. Nothing restarts.

Saving a model needs a key for that model’s provider. Without one, the dialog says “Add an Anthropic API key first (Change keys or budget)” or the same for OpenAI.

Higher effort and a stronger model mean better output, more tokens and higher cost. If you see “Output hit max_tokens” errors on the System page, raise that task’s Max output tokens.

How calls are made

Only background jobs call the AI

Pages in the app never call the AI directly. A button that asks for AI work, such as building a brief, drafting a reply or running an audit, adds a job to Startica’s local queue, and a background worker inside the app runs it. Scheduled jobs (hourly listening, the weekly replan) work the same way.

That means:

• A slow AI call never freezes the window. Pages update live as jobs finish.
• AI work only happens while Startica is running. Close the window and it keeps running in the tray or menu bar; quit it and nothing runs.
• Expensive AI jobs run once and aren’t retried automatically. If one fails, you rerun it from the page. Scoring is retried once; the test ping up to three times.

Structured outputs

Anything that goes into the database (scores, briefs, drafts, roadmap items, audit findings) is requested with structured outputs: the request includes a JSON schema, and the model’s reply must match it. Startica then validates it again against the same schema. If it doesn’t match, the call fails with “Response did not match the expected schema” instead of saving something half-parsed.

Research calls are the exception. They return prose notes, which then feed a structured strategy call. They’re streamed to avoid timeouts. If Claude pauses mid-research (a server-side tool-loop limit), Startica sends the turn back to continue, up to five times. AI visibility answers work the same way: the answer is kept as written, and a structured score call then reads it.

Web search

The research and answer tasks let the model search the web. The provider runs these searches, not your computer, and bills them per search on top of tokens.

• Claude gets two server-side tools: web search (up to 8 searches per call) and web fetch (up to 8 pages for the Intel brief, 4 for the GEO audit).
• OpenAI gets its built-in web search tool, which opens pages itself, so there’s no separate fetch tool. The same limits apply to its tool calls in total (up to 16 for the Intel brief, 12 for the GEO audit).
• AI visibility answers get web search only, up to 3 searches per answer (3 tool calls on OpenAI), and no fetch tool. Startica keeps the pages the answer cites and the searches it ran.

Prompt caching

The stable part of each prompt (the system prompt with instructions, the product brief and similar context) comes first, so both providers can cache it. When the same prefix is sent again within a few minutes, such as when scoring a batch of conversations for one product, it’s read from cache at a fraction of the normal input price. Volatile content, like the thread being scored, comes after the cached part. Claude caches the parts Startica marks; OpenAI caches automatically. Cache reads and writes are tracked separately in the cost log.

Refusals and cut-off answers

• On Claude Opus, Sonnet and Fable, Startica enables Anthropic’s server-side refusal fallback. Haiku doesn’t support it.
• If the model still declines, the call is logged as refused, and the job fails with “Claude declined the request” or “OpenAI declined the request”, with the reason when the provider gives one.
• If an answer hits the max output tokens, the call fails with “Output hit max_tokens” rather than saving a truncated result.

Batch APIs

Startica doesn’t use either provider’s batch API in the current version. All calls are regular API calls, made as jobs run.

Every call is logged

Each call, including failed and blocked ones, is written to the ai_runs table with:

Field	Example
Task and model	draft, claude-sonnet-5-5
Prompt version	reply-draft@2
Status	ok, error, refused, blocked
Tokens	input, output, cache read, cache write
Estimated cost	in USD, from Startica’s price table
Latency	milliseconds
Error	the message, if any

OpenAI reports a dated model ID (such as gpt-6-luna-2026-09-01), which is logged as is and priced like its model.

The System page shows the latest 20 under AI calls, plus this month’s total under AI spend this month. The same table drives the monthly budget, which covers both providers together.

Prompt versions

Prompts live in the code next to the feature that uses them. Each has a version string that’s logged with every call, so when a prompt changes you can tell old output from new.

Prompt version	Feature
intel-research@2, intel-synthesis@4	Intel brief (since intel-synthesis@3 it lists suggested communities strongest first, so the Getting started checklist can pre-tick the top ones)
opportunity-score@1	Scoring conversations
reply-draft@2	Reply drafts
voice-guide@1	Voice profiles
content-write@2, email-campaign@1, content-topics@2	Content Studio
roadmap-plan@3	Roadmap and replan
ad-plan@1, ad-copy@1	Ads: campaign plans and ad rewrites
ad-creative-ideas@1, ad-image@1	Ad creatives: suggested ideas and the pictures
visibility-answer@1, visibility-analysis@1, visibility-questions@1	AI visibility: answering a question, reading the answer, suggesting questions
opportunity-draft@1	Opportunities: fix drafts (titles and descriptions, JSON-LD, llms.txt, FAQ answers)
audit@3:<kind>, audit@3:geo-research	Audits
ping@1	Health check

These are the versions at the time of writing; the AI calls table shows the ones your copy uses. To change or add prompts, see Prompts and evals.