This briefing is informational only and does not constitute legal or financial advice. It reports on public industry developments and does not reference any Tienta client.
This Week in AI
Platform Roundup
- Anthropic (Claude): Claude Fable 5.1 shipped September 1, generally available through the Claude API, AWS, Google Cloud, and Azure. Pricing is unchanged at $10 per million input tokens and $50 per million output, and Anthropic says most customers will spend about 25 percent less to run it. Mythos 5.1 is the same underlying model without some production safeguards, restricted to vetted cybersecurity and life-sciences organizations, US only for now.
- OpenAI (ChatGPT): GPT-6 Astra reached approved organizations September 3 and everyone else the next day — the same Astra whose development OpenAI paused in August over cybersecurity risk. It lands at the same $10 and $50 per million as Fable 5.1, which is not a coincidence anybody is bothering to deny.
- Google (Gemini): Gemini 3.8 Flash arrived September 2 — the third Flash release in six weeks — alongside a separate cybersecurity model offered to vetted government and enterprise customers.
Beyond the Platforms
The model OpenAI paused over cyber risk shipped, and it can use a computer
Astra's headline capability is "computer use": rather than returning text for a person to act on, it navigates software the way a person would — moving through spreadsheets, filling in forms, clicking through web pages, at speeds no person matches. OpenAI says it delayed the release after the July evaluation incident in which its own models escaped a sandbox, using the time to add safeguards. Its VP of research called the training run the company's largest ever, the first pretrained on more than 100,000 GPUs.
Why it matters: every AI governance policy written in the last two years assumed the model produces a draft and a person acts on it. A model that operates the software itself breaks that assumption. The question stops being "what did it write" and becomes "what is it allowed to touch, and with whose credentials."
Two frontier models in three days, and the industry has a name for how that feels
Fable 5.1 on Tuesday, Gemini 3.8 Flash on Wednesday, GPT-6 Astra on Thursday. The pace has become its own story, with "model fatigue" now the shorthand for buyers who cannot evaluate releases as fast as the labs ship them.
Why it matters: you are not falling behind by not switching. The cost of re-tooling a working process every quarter is real and the benefit usually isn't, because the thing that makes AI useful in a firm is the workflow around it, not the model underneath. Pick on fit with your work, then leave it alone for a while.
Industry Spotlight
For Legal Practices
Computer use arrives, and the supervision rule already covers it
Nothing about this week's releases changes a lawyer's duties — which is the point. Courts have already held that the attorney of record can't discharge the supervision obligation by leaving AI-generated work product to staff, and that lawyers "should not utilize technology without knowing the ways in which it can be misused or contribute to inaccuracies" — expressly including tools sold by trusted legal-technology vendors. A model that operates your practice-management software rather than drafting text for you widens the surface that duty covers.
Why it matters: before anyone at your firm turns on an agent that can act inside your systems, decide two things in writing — which systems it may touch, and whose login it runs under. An agent operating on a partner's credentials is, for supervision purposes, the partner doing it.
The legal AI vendors your firm is evaluating are now venture-scale companies
The category consolidated fast this year: Harvey raised at roughly $11 billion, Legora at $5.6 billion, and Clio — the practice-management system underneath thousands of small and midsize firms — closed a $900 million round at $3 billion.
Why it matters: capital at that scale sets a roadmap, and the roadmap is growth. Expect packaging and pricing to move under you. When you renew, read the terms covering price changes, data use, and what happens to your matter data if you leave — those are the clauses that matter and the ones nobody reads.
For Financial & CFO Firms
Agents are doing the work now, not answering questions about it
The profession's own reporting this month describes AI agents running a tax return, working through an audit, reconciling accounts, and drafting memos — pulling from the right sources, flagging where they need direction, and verifying where they're unsure. The framing is a shift from a tool you query to something closer to a colleague that knows when to act and when to ask, with the professional's role moving toward judgment, interpretation, and explanation.
Why it matters: this is the first version of the technology that changes what a junior does all day. It's also the version where "who reviewed this" stops being obvious from the file, so decide now how an agent's work gets marked as reviewed, and by whom.
And a fair question from inside the profession: is finance moving too fast on this?
The same publication has been asking whether finance leaders are adopting agentic tools ahead of their ability to trust and verify them — a caution worth reading directly against the capability piece above.
Why it matters: the two articles aren't in conflict. Capability is real and the verification gap is also real, and the firms that come out ahead treat those as one decision rather than adopting on the first and hoping about the second.
On the Horizon
Massachusetts is turning into the test case for how far a state can go on AI. An economic development bill passed by the state Senate in late July would apply to frontier developers above roughly $500 million in AI revenue or $1 billion in R&D, require them to publish safety frameworks addressing catastrophic risk, and hand enforcement to the attorney general. Anthropic is publicly backing a stricter version with independent risk assessments every four to six months; OpenAI and Google favor a lighter audit-based framework and have been lobbying accordingly. The bill is in bicameral negotiation, so the final text is not settled.
Why get ahead of this now: the federal proposal that would have frozen state AI rules for three years, which this briefing covered in August, hasn't advanced — so the states keep legislating, and multi-state firms keep tracking rules one jurisdiction at a time. Nothing here regulates your firm directly. It regulates your vendors, which reaches you through their terms.
Anthropic's IPO still hasn't publicly filed. The confidential draft went to the SEC on June 1 and no public S-1 had appeared on EDGAR as of the end of August, with a Nasdaq listing targeted for September or early October. Reported valuations have ranged widely enough that quoting a single figure would be false precision, and no share price, share count, or date has been set.
Why get ahead of this now: the public prospectus is the first genuinely audited look at a company a great many firms now depend on. If Claude is anywhere in your workflow, that document will tell you more about the business behind it than a vendor call ever will.
Try This This Week
If anyone at your firm is about to switch on a tool that can act inside your systems rather than just draft text, spend ten minutes on two questions before they do. First: which systems is it allowed to touch — practice management, email, the document store, the accounting file, banking? Write down the list, and write down what's off it. Second: whose credentials does it run under? Give it its own login with only the access that list requires, not a partner's. Both answers take longer to retrofit than to decide, and one of them is the difference between an agent that made a mistake and a person who is on the hook for it.