You don't pick a model — you set priorities (speed, quality, cost)
and Bright Wrapper selects the best available model. Each use case below shows the recommended priority
configuration and which model it currently resolves to. Models may change as better options become available;
your priority weights stay the same.
Security note: keep Gemini credentials server-side in BrightWrapper. Do not reuse a browser-exposed
Google Maps key (
AIza...) for Gemini. Frontend keys must be restricted to Maps/Places/Routes only, and
Gemini calls should go through BrightWrapper APIs from your backend.
See
Agent Guide for the full key separation checklist.
📚 Model Catalog
Live from /api/model — hover model name for description
Loading model reference...
📄
Text Summarization
Condensing long articles, documents, or transcripts into shorter summaries.
Works well for articles up to 100K+ words thanks to large context windows.
{"speed": 2, "quality": 1}
Speed-biased → currently selects gemini-3-flash-preview.
Fast and cheap. Handles massive context windows (1M tokens) without breaking a sweat.
Quality is solid for straightforward summarization.
For high-stakes summaries (legal, medical), flip to {"quality": 2} to get
claude-opus-4-6 instead.
$ Low cost (~$0.50/M input, $3/M output)
10K words
~1¢
50K words
~4¢
Time
5-15 sec
Max input
~750K words
📋
Structured Data Extraction
Extracting specific fields from unstructured text into JSON. Examples: pulling contact info from emails,
extracting product specs from descriptions, parsing invoices.
{"speed": 2, "cost": 1}
Speed + cost biased → currently selects gemini-3-flash-preview.
Excellent at following JSON schemas reliably. Fast enough for batch processing thousands of documents.
The structured output mode ensures valid JSON every time.
For ambiguous inputs requiring judgment, use {"quality": 2} to get claude-opus-4-6.
$ Low cost
Batch size
6 parallel
100 docs
~10¢
Time (100)
~30 sec
API
/batch
🎨
Image Generation
Creating images from text descriptions. Supports various styles from photorealistic to illustration to technical diagrams.
Currently powered by Google's Gemini image generation
(gemini-3.1-flash-image-preview).
gemini-3.1-flash-image-preview
Best balance of speed and quality for most apps. Generates coherent scenes quickly while
keeping costs predictable for production image workflows.
$ ~2.4¢ per image (1K size)
Aspect Ratios
1:1
Square — avatars, icons, thumbnails
4:3 / 3:4
Classic landscape / portrait
16:9 / 9:16
Widescreen / tall — banners, hero images, stories
3:2 / 2:3
Photo proportions
5:4 / 4:5
Near-square, social media
21:9
Ultra-wide — cinematic banners, panoramic headers
Sizes & Pricing
1K
~2.4¢ — default, good for web use
2K
~4.8¢ — higher resolution
4K
~9.6¢ — print quality
Other Image Models (Industry)
Google Imagen 4
Latest Google dedicated image model (imagen-4.0-generate-001). Same aspect ratios minus 21:9.
OpenAI gpt-image-1.5
Latest OpenAI image gen. Limited to 1:1, 3:2, 2:3 only.
Stability Ultra
SD3.5 Large. Widest ratio support incl. 21:9 and 9:21. Not currently integrated.
Batch size
4 parallel
10 images
~24¢
Time each
5-10 sec
API
/batch
📊
Generate Infographics
Creating annotated diagrams with interactive hotspots. The system generates an image and
uses vision analysis to identify clickable regions automatically.
gemini-3.1-flash-image-preview
*Infographics are a two-step pipeline:
1. Image model generates the image
2. A vision-capable model analyzes the image to detect clickable hotspot regions
You're billed for both steps.
Generates the base image, then a vision model analyzes it to identify hotspot locations.
The image model handles visual generation while vision analysis accurately identifies interactive regions.
$ ~3-4¢ per infographic
Per run
1 at a time
10 infographics
~35¢
Time each
10-20 sec
Includes
auto hotspots
💬
Chat & Conversational AI
Building chatbots, customer support agents, or interactive assistants.
Requires good instruction-following and conversational coherence.
{"speed": 1, "quality": 1}
Balanced → currently selects gemini-3-flash-preview (best speed/quality ratio).
For customer-facing chat where polish matters, use {"quality": 2} to get
claude-opus-4-6 — strongest conversational depth but 10x the cost.
gpt-5.3 sits in the middle at {"quality": 1, "speed": 0.5}.
$ Low cost (balanced) — $$$ Premium (quality)
100 messages
~$0.10-2
Latency
streaming
History
200K-1M
API
/chat
🎤
Audio Transcription
Converting speech to text. Supports various audio formats and handles accents, background noise,
and multiple speakers reasonably well.
whisper-1
Industry standard for transcription (OpenAI). Handles diverse accents, background noise, and technical
vocabulary well. Currently the only transcription model available — and it's genuinely excellent.
$ ~$0.006/minute
1 hour audio
~36¢
10 hours
~$3.60
Max file
25 MB
Time
~1:10 ratio
💻
Code Generation & Analysis
Writing new code, explaining existing code, debugging, or refactoring.
Includes tasks like generating boilerplate, writing tests, or code review.
{"quality": 2}
Quality-biased → currently selects claude-opus-4-6.
Strongest production code generation with excellent understanding of project context. Extended thinking mode
helps with complex reasoning. Produces clean, idiomatic code and catches subtle bugs.
For simple scripts or bulk formatting, use {"speed": 2} to get gemini-3-flash at 1/10th the cost.
$$$ Premium (~$5/M input, $25/M output)
Alternatives
gemini-3-flash
Use {"speed": 2} for simple scripts, formatting, and boilerplate. 1/10th the cost of Opus.
1K lines
~5-15¢
Complex refactor
~$1-5
Context
200K tokens
Extended thinking
yes (Opus)
🌐
Research with Web Search
Answering questions that require up-to-date information from the internet. The model searches the web
during generation, synthesizes results, and returns a structured answer. Useful for market research,
fact-checking, competitive analysis, or any task where training data may be stale.
{"quality": 2} + "web_search": true
Quality + search → currently selects claude-opus-4-6 with live web access.
The model decides when to search and synthesizes multiple sources into a coherent answer.
All three providers support search natively: OpenAI (web_search), Gemini (google_search),
Anthropic (web_search). For quick lookups where depth doesn't matter, use
{"speed": 2} to get gemini-3-flash with search instead.
$$$ Premium (model cost + search overhead)
Per query
~5-30¢
Time
5-30 sec
API
/chat, /structured
Param
web_search: true
🔎
Semantic Search & Embeddings
Converting text into vector embeddings for similarity search, clustering, or RAG (retrieval-augmented generation) systems.
text-embedding-ada-002
OpenAI's embedding model. 1536 dimensions provides sufficient semantic resolution
for typical search, clustering, and RAG tasks. Currently the only embedding model available.
$ ~$0.10/M tokens
1K docs
~5¢
100K docs
~$5
Dimensions
1536
Time (1K)
~10 sec
For model comparisons and the agentic coding ecosystem, see
AI Landscape.