Expert controls
The dials the apps
keep locked.
The consumer apps pick one default and hide the rest. Big-AGI exposes the parameters each provider actually forwards, on your own key.
Turn the reasoning up.
The same model runs fast and cheap or slow and thorough, depending on one setting. Pick a provider and drag it. This is the real control from the app.
Reasoning effort
Balanced
OpenAI · effort = medium
OpenAI reasoning effort, from none to extra-high. Same prompt, same key. You set how hard it works, before it answers.
The frontier features, wired in.
Not just parameters. Big-AGI reaches the newest provider capabilities as they ship, the kind the consumer apps take months to expose.
Gemini
PreviewText to video
Turn a prompt or an image into a short 720p clip with sound, through Gemini Omni Flash.
Gemini
PreviewManaged agents
Run the Antigravity agent in a Google-hosted Linux sandbox, or Deep Research agents that return long-form reports.
Anthropic
Document Skills
Give Claude xlsx, pptx, pdf, and docx skills for real document work.
Anthropic
Fast Mode
Roughly 2.5x faster inference on supported Claude models, at premium pricing.
OpenAI
Precise image editing
High-quality image generation with precise input editing and an uncompressed output mode.
OpenAI
Located search
Attach an approximate location to web search so results match where you are.
xAI
Handle-scoped X search
Search X filtered to specific handles, with a recency window from a day to a year.
Perplexity
Academic mode
Restrict Perplexity grounding to academic sources.
OpenAI · Anthropic · Gemini · xAI
Code sandboxes
Server-side code execution for data analysis, files, and charts.
What each provider exposes.
Every provider forwards a different set of knobs. Big-AGI surfaces each one at its native range, and only for the models that accept it.
| Control | OpenAI | Anthropic | Gemini | xAI | Perplexity |
|---|---|---|---|---|---|
| Reasoning effort | none to xhigh | low to max | minimal to high | — | — |
| Thinking budget | — | 1024 to 65536 | 0 to 24576 | — | — |
| Verbosity | low / med / high | — | — | — | — |
| Web search | context depth | dynamic filter | grounding + recency | X + web + interval | date + academic |
| Code sandbox | yes | yes | yes | yes | — |
| Image generation | 4 modes | — | ratio + size | — | — |
Read from each provider live, per model. This is the vendor-level picture.
Four groups of control.
Reasoning
The axis the consumer apps hide. Set effort from none to extra-high, or hand a model a raw thinking budget in tokens. Big-AGI only offers the levels a given model actually accepts.
Web and tools
Turn web search on with a recency window, pre-filter results, run a hosted code sandbox, search X with a handle filter, or attach document skills. Each is a real toggle, not a hidden default.
Output
Steer verbosity separately from reasoning, request structured outputs, or drive native image generation with aspect ratio and size. The shape of the answer is a control, not a surprise.
Sampling
Temperature at its real range, forwarded straight to the provider. No invisible clamping, and nothing added to your prompt.
Deeper than a settings panel.
Set a control once, or scope it. Each layer overrides the one before it, so a persona can carry its own reasoning depth and a single message can still override everything.
Stop settling for one default.
Bring your own key, open the workspace, and turn the dials the apps keep for themselves.
BIG-AGI
Resources
© 2026 Token Fabrics·Built with passion in San Diego