AI Browser Agent — Models

Which AI model each Prompteryx model runs on, what the Prompteryx harness adds on top, and how the other models (Gemini, Claude, GPT) are run.

Prompteryx models: a well-known model inside a harness we built

A Prompteryx model is not a new AI model. Each one runs on a model from Google (shown below and, in smaller grey text, next to its name in every model list). What makes it a Prompteryx model is the harness around it: the instructions, tools and checks Prompteryx built and tuned for driving a browser. We measured each harness on a fixed set of real website tasks and kept the version that finished the most tasks correctly.

A Prompteryx model costs the same per step as the model it runs on — the harness itself adds no charge. When Google retires a model, we move the Prompteryx model to Google’s replacement and this page and the grey text change with it (for example, Prime+ moved from Gemini 3.7 Flash to Gemini 3.8 Flash on 12 Oct 2026).

Prompteryx model	Runs on	What the Prompteryx harness adds	Best for
Prompteryx Prime	Gemini 3 Flash	Our benchmark-winning instructions, plus page tools: it can read the page's text, check a value really changed and act on an element directly instead of by pixel. It also switches between tabs and follows our rules for pages that try to block automated browsing.	The best all-rounder — anything where the answer must be right.
Prompteryx Core	Gemini 3 Flash	The same benchmark-winning instructions (very careful with forms: one field at a time, then checked), looking at the screen only.	The default. Cheaper and faster than Prime; strong on forms.
Prompteryx Swift	Gemini 3.5 Flash-Lite	The same instructions as Core on Google's lightest model.	Simple, high-volume tasks where speed and price matter most.
Prompteryx Prime+	Gemini 3.8 Flash	Exactly Prime's harness on Google's newer, generally available Flash model.	Prime's approach on a newer model; slightly pricier per step.
Prompteryx Jev (beta)	Jev + Gemini 3.5 Flash-Lite	Works from the page's structure (its accessibility tree) instead of screenshots. Jev, our own decision model, picks every step; Gemini 3.5 Flash-Lite writes any text it types, writes the answer, and looks at the screen when Jev is stuck (a pop-up the page structure doesn't show).	Very fast, text-heavy pages. Cloud Browser only.

What the harness does on every step

Tuned instructions — how to plan, keep track of progress, fill forms carefully and check its own work before saying it is done.
Screenshots at the right size — the AI Vision setting picks how sharp each screenshot is, which is the main thing that decides cost.
Several actions per step when it is safe (typing a whole form, then one look), so a task takes fewer, cheaper steps.
A final check (Prime, Prime+, Core and Swift) — at the end the run is checked against what you asked, so a run that only looked finished is reported as unfinished.
Safety stops — before a purchase, a sign-in or sending a form it pauses for your OK (you can change that per schedule).
Page tools (Prime and Prime+) — reading the page’s text and checking values directly, switching tabs, and set rules for pages that try to block automated browsing.
Everything else you see — the live view, recording, parallel agents, structured output and the credits shown for each step.

The other models

Gemini 4 Argon (and, for admins, the plain Gemini Flash models) — Google’s own computer-use approach with no Prompteryx tuning. Useful to compare against the Prompteryx models.
Claude Fable 5.2, Claude Fable 5.1, Claude Opus 5 — Anthropic’s models on our vision loop: each step they see a screenshot and choose a click, typing or scrolling. Fable 5.2 is the newest and most capable; Fable 5.1 is the previous Fable.
GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra — OpenAI’s models on OpenAI’s own computer-use tool.

These are named exactly as their makers name them, so there is no grey text next to them.

Where you see which model ran

Model lists — the AI Browser Agent toolbar, the chat’s browser picker, a schedule’s browser settings and the Visual Studio AI Browser Agent node show the Prompteryx name with the model it runs on in grey (“Prompteryx Core · Gemini 3 Flash”).
Progress → Used in this session — names the chat model and the browser model that ran, e.g. “Gemini 3.8 Flash, Prompteryx Core · Gemini 3 Flash (browser)”.
This reply: N credits under a reply — its breakdown lists the browser model the same way, with the credits it used.

More about each setting: AI Browser Agent — Settings.

Prompteryx models: a well-known model inside a harness we built

Prompteryx model	Runs on	What the Prompteryx harness adds	Best for
Prompteryx Prime	Gemini 3 Flash	Our benchmark-winning instructions, plus page tools: it can read the page's text, check a value really changed and act on an element directly instead of by pixel. It also switches between tabs and follows our rules for pages that try to block automated browsing.	The best all-rounder — anything where the answer must be right.
Prompteryx Core	Gemini 3 Flash	The same benchmark-winning instructions (very careful with forms: one field at a time, then checked), looking at the screen only.	The default. Cheaper and faster than Prime; strong on forms.
Prompteryx Swift	Gemini 3.5 Flash-Lite	The same instructions as Core on Google's lightest model.	Simple, high-volume tasks where speed and price matter most.
Prompteryx Prime+	Gemini 3.8 Flash	Exactly Prime's harness on Google's newer, generally available Flash model.	Prime's approach on a newer model; slightly pricier per step.
Prompteryx Jev (beta)	Jev + Gemini 3.5 Flash-Lite	Works from the page's structure (its accessibility tree) instead of screenshots. Jev, our own decision model, picks every step; Gemini 3.5 Flash-Lite writes any text it types, writes the answer, and looks at the screen when Jev is stuck (a pop-up the page structure doesn't show).	Very fast, text-heavy pages. Cloud Browser only.

What the harness does on every step

Tuned instructions — how to plan, keep track of progress, fill forms carefully and check its own work before saying it is done.

Screenshots at the right size — the AI Vision setting picks how sharp each screenshot is, which is the main thing that decides cost.

Several actions per step when it is safe (typing a whole form, then one look), so a task takes fewer, cheaper steps.

A final check (Prime, Prime+, Core and Swift) — at the end the run is checked against what you asked, so a run that only looked finished is reported as unfinished.

Safety stops — before a purchase, a sign-in or sending a form it pauses for your OK (you can change that per schedule).

Page tools (Prime and Prime+) — reading the page’s text and checking values directly, switching tabs, and set rules for pages that try to block automated browsing.

Everything else you see — the live view, recording, parallel agents, structured output and the credits shown for each step.

The other models

Gemini 4 Argon (and, for admins, the plain Gemini Flash models) — Google’s own computer-use approach with no Prompteryx tuning. Useful to compare against the Prompteryx models.

Claude Fable 5.2, Claude Fable 5.1, Claude Opus 5 — Anthropic’s models on our vision loop: each step they see a screenshot and choose a click, typing or scrolling. Fable 5.2 is the newest and most capable; Fable 5.1 is the previous Fable.

GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra — OpenAI’s models on OpenAI’s own computer-use tool.

These are named exactly as their makers name them, so there is no grey text next to them.

Where you see which model ran

Model lists — the AI Browser Agent toolbar, the chat’s browser picker, a schedule’s browser settings and the Visual Studio AI Browser Agent node show the Prompteryx name with the model it runs on in grey (“Prompteryx Core · Gemini 3 Flash”).

Progress → Used in this session — names the chat model and the browser model that ran, e.g. “Gemini 3.8 Flash, Prompteryx Core · Gemini 3 Flash (browser)”.

This reply: N credits under a reply — its breakdown lists the browser model the same way, with the credits it used.