The next AI product is the workflow

OpenAI’s September 10 introduction of ChatGPT for Financial Services makes a useful statement about where AI products are heading. The tailored ChatGPT Work offering combines GPT-6 Astra with built-in financial datasets and tools for producing research and client materials. The company says data from providers including Daloopa, PitchBook, LSEG News and Crunchbase are indexed and hosted on its infrastructure, with granular citations intended to make claims easier to trace.[1]

The significant product idea is the connection between evidence and a finished piece of work. A model that can reason well still needs the right figures, the relevant document version and a way for someone to examine what it did. Putting those elements together changes the competitive question from the quality of a chat response to the reliability of an entire working process.

OpenAI says Morgan Stanley and Evercore helped shape the product, with initial attention to investment banking and equity research. That is evidence of its intended audience and development process. It is not an independent measurement of time saved or error rates. The launch makes claims about more reliable data access; those claims still need to be tested against difficult, ordinary work.[1]

Availability is part of the news

The following day, OpenAI updated its GPT-Rosalind announcement to say the specialized life-sciences model was leaving research preview and becoming available globally to eligible organizations through a trusted-access program. The notice says published pricing takes effect October 5. This is a September change to access, attached to an article originally published in June; it should not be mistaken for a wholly new model launch on September 11.[2]

The underlying product emphasizes scientific workflows, including evidence handling, analysis and the production of research artifacts. OpenAI publishes its own evaluation results for those tasks. Such evaluations can help describe what the company is optimizing, but performance on a benchmark is not equivalent to a validated laboratory result or a demonstrated improvement across an organization’s research program.[2]

Anthropic is pursuing a related audience. On August 27, it announced an initial 10,000 free or discounted one-year subscription seats for scientists, alongside an expansion of its AI for Science program. The company describes Claude Science, launched in June, as combining research tools, auditable artifacts and access to computing resources. Eligibility and model access vary; the announcement is not a promise that every researcher receives unrestricted access to every model.[3]

What would make the products convincing

These announcements suggest that the useful comparison is becoming more concrete. Can a system retrieve the right source, preserve its meaning, perform the analysis and leave a result another person can inspect? A persuasive demonstration would show the path from input to output, including the points where an expert intervened. A polished final document alone cannot answer those questions.

There are several ways a product can fail while appearing successful. It can cite the correct company and the wrong reporting period. It can produce a clean table while silently mixing definitions. It can run an analysis that is internally consistent but answers a different question from the one requested. These are illustrative evaluation cases, not allegations about the products announced this week.

The corresponding test should include messy inputs, revised documents, ambiguous requests and a reviewer who did not watch the work being produced. It should count the time required to inspect and correct the result. It should also establish what happens when a source is missing: does the system disclose the gap, ask for help or fill it with a plausible answer?

For now, the clearest verified development is a change in product packaging and availability. Vendors are putting models closer to domain-specific data and tools, and giving specialized users more routes to access them. The larger claim—that this reliably improves professional work—will require evidence from use. That is the distinction worth carrying into the next round of launches.