Bancada

Bancada · early access

Every model.
One subscription.
One history.

GPT, Claude, DeepSeek, Qwen, Llama and Hermes in a single workspace — the conversation survives the model switch. Built in Brazil, priced in BRL.

Access rolls out in groups, first come, first served.

The math you already do

People who work with AI don't subscribe to one model. They subscribe to three.

ChatGPT for one thing, Claude for another, an open model for the rest. Each on its own bill, each with its own history — and the context dies with every tab switch.

The arithmetic, checked Aug 3, 2026:

ChatGPT Plus
~US$ 20/mo
Claude Pro
~US$ 20/mo
Sum
~US$ 40/mo
Unified history
none

Standard pricing as of the date above — check their pages; it's your math. And yes, those subscriptions include more than chat. If chat is what you use, this is the math.

Bancada exists so that math works out differently: one place, one subscription, every model. Final pricing ships with the launch — the one public commitment: less than the sum.

What's inside

The models, by name.

No “best-in-class models” hand-waving — the actual list. It grows, and the waitlist gets the changelog.

Frontier

The top closed models, for work that demands the ceiling.

  • GPT (OpenAI)
  • Claude (Anthropic)

Open

Fast, cheap and improving monthly — the daily workhorses.

  • DeepSeek
  • Qwen
  • Llama

Evil IA

the less-restricted mode

The house mode: less-restricted open models like Hermes, for authorized security research, red-teaming and technical work that consumer filters get in the way of. Under the acceptable-use policy — responsible freedom, not absence of rules.

  • Hermes
  • and other open models

How it works

Three steps, no magic.

  1. Pick a model

    Or don't — switch mid-conversation and the context comes along.

  2. Work

    Chat, file upload, long context. One persistent history, yours.

  3. Compare

    The same prompt on two models, side by side. Stop guessing which one is better at what.

What we use it for

Built by people who run several models a day.

Compare before you trust

The same question on GPT and Claude, side by side, before that answer lands in your codebase.

Code and long context

Upload the file, discuss the diff, keep the thread — without restarting the conversation every session.

Route by cost

Simple question to an open model, hard problem to a frontier one. You choose.

Security, done right

Defensive analysis, CTFs, authorized red-teaming — with Evil IA, minus the consumer-filter fights.

Made here

Priced in BRL, paid by Pix, with an invoice.

Foreign platforms bill in dollars, on international cards, no invoice. Bancada is Brazilian: subscription in reais, Pix, a proper invoice your company can deduct — and support in your language, from the people building the product.

  • Priced in R$
  • Pix
  • Invoice (NF)
  • Support in Portuguese

Questions technical people ask

The answers you'd go digging in the footer for.

Do my conversations train models?

No. Your conversations stay in your history, for you — they don't become training data, ours or anyone's. That's the first line of the privacy policy, not the fine print.

Where does my data live?

In your history, encrypted in transit, exportable any time. Deleted means deleted. The honest detail: prompts sent to frontier models pass through the providers' APIs under their API terms — which don't use API content for training by default.

Does “less restricted” mean no rules?

No. It means open models without consumer-product filters — useful for authorized offensive security, research and sensitive technical work. The acceptable-use policy applies to everything: no illegal activity, full stop. Responsible freedom is a feature; anarchy isn't.

Will there be an API?

It's on the Dev plan, after launch. The waitlist votes on roadmap order — it's public.

What will it cost?

Pricing ships with the launch, in BRL. The public commitment: less than the sum of the two subscriptions above. Cancel anytime, no phone call, no retention flow.

Why a waitlist?

Because opening to everyone at once, with real inference costs underneath, is how products die in month one. We open in groups, measuring cost and talking to every user. Slow on purpose.

Who's building

Two brothers, in public.

One builds the infrastructure — servers, GPUs, model integration. The other builds the business — product, community, the page you're reading. The construction is open: architecture, real GPU costs and product decisions, published as they happen. If that sounds like the right way to build, you'll like the inside.

The first group gets in first.

No spam: one email at launch, a changelog if you want it. That's all.