My view

I do not need every new AI tool to do everything. I need to know which part of my work it can help with, what could go wrong, and how I will check it. My direction is simple: use AI to prepare and question the work, then keep responsibility for the numbers and the final view.

What supports it

The official product pages point to different roles. GPT-6 Astra combines reasoning with coding and computer use. Anthropic presents Opus 5.5 as an improvement for coding and knowledge work. Ask Jev is a separate public demo that uses a judgment model to handle short questions. These are not three interchangeable financial analysts. Source: OpenAI Anthropic Ask Jev / Wayne Sutton.

GPT-6 Astra: help me build the first version

OpenAI describes Astra as able to work across browsing, code, and professional tasks. For me, the useful test would be turning a public annual report into a traceable input table and a small working model. I would ask for the source page, currency, unit, and period beside every number. I would then check those inputs and run the calculations myself. Source: OpenAI.

Opus 5.5: help me question the work

Anthropic released Opus 5.5 on 22 September and describes improvements in coding and knowledge work. I would test it as a reviewer: which assumption drives the result, what evidence is missing, and which formula fails under a different input? That is a role I want to test, not a claim that it is better than Astra at every task. Source: Anthropic.

Ask Jev: learn the routing idea

Ask Jev is Wayne Sutton’s demo, not a TypeSafe-operated service. It uses TypeSafe’s Jev model for short judgments and can route a question to another model for a written reply. I would use made-up examples to learn how that handoff works. The site’s public questions appear on a shared wall, and its operator can see private asks too. I would keep real company and personal data outside it. Source: Ask Jev / Wayne Sutton.

My first finance exercise

I would map the company’s suppliers, buyers, and revenue sources first. Then I would set one question, such as how a change in payment timing affects cash. AI could help extract and organize the inputs, write a first calculation, and suggest checks. I would reconcile totals, test the assumptions, and write a short conclusion in my own words.

What I would keep learning

Accounting tells me whether a figure makes sense. Industry knowledge tells me what drives it. Statistics helps me judge a comparison, and coding helps me repeat the work. Better AI gives me more reason to learn those skills: I need them to give clear instructions and recognize a bad answer.

What could challenge this view

A stronger model can still use the wrong reporting period, miss a footnote, or build a formula that looks right but is not. Running the same question through two tools can help me find disagreements, but agreement alone does not prove the answer. I need the original source and a calculation I can reproduce.

What I will watch next

I would start with one company, one annual report, and one small model. I would record extraction errors, unsupported statements, formula failures, the minutes spent checking and correcting the work, and the cost per accepted result. I have not run that comparison yet, so this is a plan rather than a ranking.

Sources