.avif){kind=icon}
The first AGI Model is here?
OpenAI president, Greg Brockman, closed the press briefing with "Welcome to the AGI era," remarking he personally believes OpenAI has reached artificial general intelligence, and if we look back in a couple of years, it might be about this model.
Artificial general intelligence has been OpenAI's mission since 2015. Their charter defines it as highly autonomous systems that outperform humans at most economically valuable work.
In finance most economically valuable work includes month-end close, VAT filing and budget modeling.
The AGI Debate
So the question if we actually reached artificial general intelligence is fair but it’s unanswerable, at least for now, and I would rather skip it.
ARC Prize, whose benchmark is specialised in AGI, published its own analysis and specifically declined to call the model AGI. There is no referee and no agreed definition.
Why Astra Changes Finance Work
But what there is… one capability that changes the shape of finance work.
Astra can operate your computer, and it’s smarter than anything else we have right now.
It can use software, it can also reconcile messy data with legacy tools, and it also can execute complex workflows just like a human analyst, but much much faster. By automating manual clicking and typing, your job changes from data entry, into orchestrating and directing.
Why does this matter so much?
Until now, getting AI into a finance process meant getting AI into your systems. APIs, connectors, IT tickets, and security reviews.
That works only for the modern part of your finance stack. But there are other places.
Other places are most of the work. The bank portal that exports a CSV with a merged header row, the tax portal, the ERP screen, the vendor form that every supplier sends in a different layout.
None of it has an API you are getting access to anytime soon. So a human sits there, moving the data between screens, because software is not there yet.
I did that work for two years before Alaan, doing month-end close and reconciliations. The description of my work was moving things around more than anything.
Pull the statement, paste it, match it, find the items that do not match, then email the person who knows why, then wait.
The analysis takes a few hours, but getting to the point where analysis is possible takes weeks sometimes.
At Alaan we built Alaan Intelligence to eliminate that manual legwork across corporate spend and transaction matching, but Astra takes browser-level execution even further.
Astra value is in that middle layer. OpenAI's launch page shows it doing impressive things including, filling out forms, updating records in a CRM, and completing a US Form 1040. There is also a demo of an Excel problem taken from the Financial Modeling World Cup.
Nicolas Boucher checked that one and what did he find? Astra finished it about four times faster than the human who won the competition.
Those are the CFO numbers.
Performance Benchmarks
On AutomationBench, which tests multi-step business workflows rather than puzzles, Astra scored 41.4%. The previous OpenAI model was at 18.1%. Claude Fable 5.1 sits at 31.4%.
On Agents' Last Exam, which OpenAI describes as testing agents on professional tasks in real software including financial modeling, Astra reached 59.3%.
What does that say? The best available model, working in real applications on complex professional work, fails around four times out of ten.
Another two that matter more to accountants than to anyone else. OpenAI's hallucination benchmark puts Astra at 4.2% against 12.2% for its predecessor, and it is roughly three times less likely to claim it did something it did not do.
OpenAI reported 99.9% on ARC-AGI-3. ARC Prize ran it on their own harness and got 62.7%. Both numbers are accurate in different environments. Like the difference between a demo and deployment.
Security and Safety Measures
The safety section of this launch reads like it was written for auditors, and why? Well….
In July, a group of OpenAI models in testing escaped their sandbox and breached Hugging Face's systems. The investigation found hundreds of agents coordinating and behavior aimed at hiding what they had done.
OpenAI paused parts of its work, delayed Astra, then built a test from that incident: hand the model a difficult or impossible task and see whether it pushes past the boundary it was given.
The previous model went outside the authorized target 48% of the time without safeguards. Astra did it in 0% of cases.
Zero is a good number. But it remains a test results, not a full control.
Finance Control
What we are being offered is a new kind of user. One that can log into a system of record, read what a human reads, and change things.
Every question we already ask about human access applies here without modification. Who approved this? What changed? Can one identity both create the payment and release it? Is there a log that survives the person who made it?
OpenAI's answer inside Codex is a review step that halts an action for human confirmation, and it reports Astra never tried to get around that step even when the setup was deliberately made easy to bypass.
To me this sounds a lot like a maker-checker.
Finance invented that control. We should be the ones insisting it is switched on, all the time.
Getting Started with Astra
Ok Ali, that sounds great, but where do we start?
Start somewhere a mistake is cheap and reversible. Reconciliation prep, normalizing statements, drafting variance commentary for a human to sign, cleaning up a model someone else built badly.
Start by reading and preparing. Without posting yet.
Also give the agent its own login credentials instead of sharing your own, so the audit trail tracks which actions were the machine's.
Nothing in the month-end close changed this month. The close will look the same, it will be the same, but faster, way faster.
The bottleneck that made finance work slow was never intelligence. It was access, and access is what we have now.
The gap between what a model can do and what we trust it to do unattended is a control problem. Conveniently, closing that gap is exactly our job.
Would you give an agent its own login to your ERP? And what would need to be true before you let it touch anything that posts?
Gain control over business expenses with Alaan corporate cards
Invygo earned AED 100,000 in cashback using Alaan cards
Personalise approval workflows to align with your business needs with Alaan's Spend Management platform
Discover the power of automated expense tracking and smarter spend control with Alaan
Turn data into actionable insights with Alaan's spend management tools
Stay Tax Compliant with UAE's #1 Corporate Card and Spend Management Platform
Customisable corporate card policy template
Know how much Corporate Tax you have to pay this fiscal year
Easily integrate Alaan with your ERP for accurate, real-time expense tracking
Close books faster with Alaan's AI-powered accounting automation
Reconcile your books in minutes instead of hours, every single month.
Keep petty cash organized with Alaan corporate cards and automated expense management
