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 Haiku 5.5 launched October 7 as the company's fastest and cheapest model, and the first Haiku with adjustable effort levels. It costs $0.10 per million input tokens and $0.50 per million output for requests under 100,000 tokens, down from $1 and $5 for Haiku 4.5. Requests above that size cost $0.50 and $2.50. Anthropic also cut Sonnet 5.5's cache-read price from $0.20 to $0.10 per million tokens, and says monthly API credits, up to $500 shared across a Team plan, are rolling out this week.
- OpenAI (ChatGPT): GPT-6 Sol began rolling out in ChatGPT on October 7 for Plus, Pro, Business, and Enterprise subscribers, and GPT-6 Luna started reaching Free and Go users the next day. It arrived with Intelligent UI, a redesign in which ChatGPT answers with interactive elements instead of text alone.
Beyond the Platforms
ChatGPT now answers with buttons, forms, calculators, and charts
OpenAI says Intelligent UI lets responses include graphics, tappable buttons, forms, charts, and interactive tools such as calculators and bill splitters. In OpenAI's demo to reporters, examples included a personal savings calculator and a multi-day hiking map. Users can turn the visuals down if they find them excessive. The examples come from OpenAI's own demonstration, and TechCrunch reported no independent testing.
Why it matters: an interactive calculator looks more authoritative than a paragraph of text, and it is just as capable of being wrong. If staff start using these in-chat tools for client numbers, fee estimates, or financial scenarios, the verification step still applies. The output being easier to use does not make it any more reliable.
Industry Spotlight
For Legal Practices
A federal judge ordered government lawyers to explain how they described the Army's AI use
In a bid protest over a roughly $450 million Army contract at White Sands Missile Range, the Court of Federal Claims ruled against the protester, TRAX International, on September 22, finding it could not tie any actual errors to the Army's AI tool, though Judge Carolyn Lerner wrote she "cannot entirely rule out the use of AI." The ruling's AI story is what drew attention. The government first said an analyst tested the tool on three proposals, that evaluators saw only one output, and that nothing went into the contract file. After the judge demanded specific answers, the analyst admitted he had also sent the AI evaluations of two proposals to two evaluation board members and the contracting officer. Law360 reported on October 8 that the judge has ordered four government attorneys to explain why they should not be sanctioned for misstatements about the Army's AI use. The order is sealed, so the specific allegations are not public, and the sanctions question is still open.
Why it matters: this is not a hallucinated-citation case. It is a case about what lawyers tell a court regarding how AI was used, and the account changed under questioning. California's SB 574 and the 10th Circuit proposal both turn on disclosing AI use accurately. The practical lesson for any firm is to find out, before you file anything that describes AI involvement, exactly who used which tool on what, and to write it down while it is still easy to reconstruct.
For Financial & CFO Firms
A practitioner checklist for what to settle before client tax data touches AI
A piece published October 12 in CPA Practice Advisor, written by the co-founder of an AI implementation firm for service businesses, lays out eight steps. Among them: write a one-page policy naming the approved tools and plans, get a separate AI consent from clients worded with counsel, offer a no-AI path for clients who decline, get vendor commitments in writing on retention and model training, confirm one client's documents cannot inform work on another, require human review of all output, and log AI use per return. It builds on the IRS Office of Professional Responsibility's June 24 alert, which calls for secure, enterprise-approved AI and warns against uploading sensitive information to public tools. The author notes that the IRS has not said whether sending return information to an AI tool counts as a third-party disclosure under Section 7216, and recommends treating it cautiously.
Why it matters: this is one practitioner's recommendations, not IRS guidance, and the consent wording is a question for your own counsel. But the open question is real. Where the rules are unsettled, the defensible position is a written policy, a short approved-tools list, and a vendor answer on training you can show someone. Per the author, a vendor that will not confirm it does not train on your data has given you your answer.
For SaaS & Engineering Teams
A model priced for volume changes which work is worth automating
Anthropic's own figures for Haiku 5.5 put it well below Sonnet 5.5 on hard coding tasks: 39.2 percent on Terminal-Bench 4.0 at maximum effort, against Sonnet's 70.6 percent, and roughly 20 percent at the default medium setting. It scores 72.4 percent on an offline subset of OSWorld 2.1, a computer-use test, up from 15.7 percent for Haiku 4.5. These are vendor-reported results, and the model is aimed at classification, summarization, support, and other high-volume jobs rather than difficult engineering.
Why it matters: at these prices, tasks that were not worth automating last month, such as tagging every ticket or summarizing every call, now are, and each one adds another automated step touching real data. Last week's PixelLeak finding was that agents cause trouble by taking routes no one expected. More cheap, automated steps means more of those routes. Before you wire a small model into a pipeline, decide what it may read, where it may write, and who gets alerted when it does something else.
On the Horizon
The 10th Circuit's proposed AI filing rule closes for public comment on October 18. As covered in issue 7, it would require anyone filing an AI-assisted document in a federal appeal to confirm a person reviewed it. If adopted, it takes effect January 1, 2027, the same day as California's SB 574 on lawyer AI use.
Why get ahead of this now: the comment window is days away, and any firm that files in federal appellate courts has a say in how the rule is worded. Whether or not you comment, the rule is a preview of the verification duty most courts are moving toward.
Try This This Week
Write down, on one page, which AI tools your firm approves and what client information may go into each. Then send each vendor one written question: do you retain our prompts and files, and do you train on them? If you cannot answer both questions for every tool people are using today, that gap is the first thing to close.