What We Changed in BrahmAI: Maths, Documents and Fairer Billing

Most product updates are a list of features. This one is mostly about things that were quietly wrong and are now right — which, if you use the product, matters more.

Here is what changed and why.

Maths now renders as maths

If you asked BrahmAI an engineering or maths question, the answer often came back looking like this:

[ t = \frac{P \cdot D}{2 \cdot (S \cdot E + P \cdot Y)} ]

That is a correct formula. It is also unreadable unless you happen to know LaTeX.

Models write maths in a typesetting notation that expects the display layer to turn it into an actual formula. Ours did not. So every answer involving an equation — every pipe thickness calculation, every beam design, every bit of statistics — arrived looking broken, even when the maths was right.

Formulas now render properly, as formulas. This affects a narrow slice of users enormously and everyone else not at all, which is exactly the kind of fix that tends to sit unfixed for a long time.

One detail we are quietly pleased with: prices in ordinary text still render as prices. The obvious way to add maths support also treats $12-$29 per month as an equation and mangles it. We checked before shipping.

Reading documents, and reading handwriting

BrahmAI reads attachments — PDFs, images, spreadsheets, text files — and this has been extended in two directions.

Models that cannot see images can still read your documents. Not every model is multimodal. NVIDIA's Nemotron family, for instance, is text-only by design. Rather than silently dropping your attachment, we convert the document to text first so the model reads its content. You get an answer either way.

Council sessions accept attachments too. More on Council below, but the short version: you can now send the same document to several models at once and compare what each makes of it.

There is an honest limitation worth stating. Reading a photograph of handwriting is genuinely hard, and different models disagree on ambiguous digits. For engineering work, where a misread number is worse than no answer at all, check anything that matters. We would rather say that than imply the feature is more reliable than it is.

Council: when you do not know which model to trust

Council remains the feature we would point to first if you are comparing BrahmAI to anything else.

You ask a question. Several models answer it simultaneously, side by side. Then a separate model reads all the answers and gives you a verdict on which is strongest and why.

The reason this exists: the honest answer to "which AI model is best" is that it depends on the question, and you usually cannot tell in advance. Council removes the guess. When the models agree, you can be more confident. When they disagree — which happens more than you would expect — that disagreement is itself the useful signal.

Billing that matches what you used

This is the least glamorous section and probably the most important.

We audited how BrahmAI charges for usage and found several places where the numbers did not match reality. All of them are now fixed.

Stopping a response mid-answer. Previously, hitting Stop meant the whole request was free — even if you had already received the entire answer. That was a bug in your favour, but a bug. Stopping now costs what was actually generated: stop after two words and you pay for two words.

Features that were never charged. Web search and memory extraction both make real calls to outside services on every use, and neither was being billed. They now are, at one credit each, and only when they actually return something. A search that finds nothing is free.

Plan changes. Switching between plans mid-month could, in certain sequences, hand out more credits than intended. Upgrades now grant your new allowance once per billing period, and downgrades keep the credits you already hold rather than resetting them.

Concurrent requests. Firing many requests at once could let them all pass a balance check that only one of them could afford. Requests in flight are now counted against your balance.

Some of these mean you are charged in situations where you previously were not. We think that is the right call. Billing that is quietly wrong in either direction is worse than billing that is predictable, and a product that cannot account for its own costs cannot stay affordable.

Security work you should not have to think about

We ran a full security audit of the platform and worked through the findings.

Without turning this into a list of things that were once wrong: the issues were fixed, verified against the live system rather than assumed, and covered by tests so they do not come back. Our automated test suite grew from 438 checks to over 630 in the process.

Two changes are worth mentioning because they affect you directly:

• Shared conversations are properly private. A shared link gives access to that one conversation and nothing else.
• Uploads are bounded. File size and type are enforced on our side, not just in the browser.

What has not changed

The lineup is now 40 models across 13 providers on a single subscription — Claude, GPT, Gemini, Grok, DeepSeek, Llama, Nemotron, Kimi, Mistral, Qwen, Command, GLM, and our own Nandini.

The free tier still exists, still needs no card, and still includes five models. If you have not tried the platform since these changes, that is the place to start.

What is next

More capable models arrive constantly, and keeping the lineup current is ongoing rather than a project with an end. Beyond that, the work we find most interesting is making it easier to know which model to use — Council is the first version of that idea, not the last.

If something in BrahmAI is wrong or awkward, tell us. Several items in this update came from exactly that.