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): Opus 5.5 shipped September 22, the first model in the new 5.5 family. Priced at $4 per million input tokens and $20 per million output, 20 percent below Opus 5, with cache reads down 60 percent to $0.20 per million. Anthropic says typical workload costs drop 40 percent and output speed is up more than 30 percent. It matches Fable 5.1 on most work and beats it on agentic coding, computer use, and visual chart recognition.
- OpenAI (ChatGPT): GPT-6 Sol and GPT-6 Luna shipped the same day, September 22. Sol is priced at $2 per million input tokens and $10 per million output, half of GPT-5.6's rate, with a 1.05 million token context window. OpenAI positions it as the working model for coding, agent tasks, and business workflow automation.
- Google (Gemini): Released a cybersecurity-focused Gemini 3.8 Flash variant for vetted defenders, part of a broader push this month alongside Anthropic and OpenAI to offer safeguarded cyber-focused models to trusted organizations.
Beyond the Platforms
Two labs cut prices and raised the bar on agentic coding, on the same day
Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol both launched September 22, each substantially cheaper than its predecessor while posting real gains specifically on agentic coding and computer-use tasks, the kind of work where a model acts rather than just drafts text for a person to review.
Why it matters: this is the same shift issue 5 flagged with GPT-6 Astra, now cheaper and from two labs at once. A model that operates software directly, at a lower price than last month's version, means the barrier to someone on your team spinning up unsupervised agentic AI work just dropped twice in one day. If you don't know what's already running against your systems, that gap got cheaper to open.
Industry Spotlight
For Legal Practices
Legal's largest litigation-support vendor just bought its way into AI-powered breach response
Epiq acquired Canopy, a data breach response technology company, on September 25, its third acquisition this year. Canopy's platform assesses exposed data, detects sensitive information, and identifies impacted individuals; its team joins Epiq to build out "Epiq AI for Cyber," available through Epiq's existing service platform.
Why it matters: when the vendor firms already use for e-discovery and incident response is racing to build AI-native breach-response tooling, that's a signal about how seriously data-exposure risk is escalating across the industry, not just a product announcement. That kind of tool finds the exposure after client data has already left the building through some ungoverned AI use nobody approved. A documented policy is what keeps a firm from needing it in the first place.
For Financial & CFO Firms
The AICPA and a former IRS commissioner just stood up a formal council on AI risk in tax
The AICPA and former IRS Commissioner Danny Werfel launched the Council on AI Risk in Tax (CART) this month, bringing in the National Association of Enrolled Agents, the Federation of Tax Administrators, and representatives from accounting, law, technology, government, and academia. Its first job is testing and refining an AI Risk Framework for Tax spanning 20 risk areas across four categories: information integrity, fairness and legitimacy, security and data, and institutional capacity.
Why it matters: this is the profession's own governing body formally organizing around what "responsible AI use" actually means in practice, not just in principle, meeting throughout the year rather than issuing one-time guidance. A firm that's already documented its own AI use and written a policy won't be starting from zero when CART's framework starts showing up in exam expectations. A firm that hasn't will be.
For SaaS & Engineering Teams
Coding agents got cheaper and more capable on the same day, which makes "we'll govern this later" a worse excuse than it was last week
A survey of 307 senior technology leaders, the kind of governance research this briefing has cited before, found 93 percent at least somewhat concerned about AI-generated code reaching production ungoverned, while only 8 percent describe their own organization's governance as strong. Opus 5.5 and GPT-6 Sol shipping the same day, both cheaper and both stronger specifically at agentic coding, only widens that gap: the tools got easier to adopt without anyone approving them, not harder.
Why it matters: the case for waiting on an AI use policy has always rested on the tools not being capable enough yet to matter. That argument gets weaker with every release, and this week it got weaker twice, from two different labs, in one day.
On the Horizon
OpenAI says an internal, unreleased model resolved more than 100 long-standing open problems across most branches of mathematics, after roughly 24 days of training. The company announced an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton's Institute for Advanced Study and including Fields Medalist Timothy Gowers, to help evaluate and release the results. Most of the 100-plus claimed results haven't been published yet, so outside mathematicians can't independently verify them, and 25 Fields Medal winners have separately signed an open letter warning that AI labs racing to claim credit for famous problems threatens the field's own scholarly norms.
Why get ahead of this now: nothing here is actionable yet, and the verification question is exactly the point, a claimed result and a confirmed one are not the same thing. Worth watching whether the advisory group's review process becomes a real model for how AI-generated claims get checked before anyone trusts them, in mathematics or anywhere else.
Try This This Week
If your team already uses an AI coding assistant, ask whoever leads engineering to name every tool that's actually approved and what data is allowed into each one, out loud, in under a minute. If nobody can do that, that's the gap to close first. Cheaper, more capable tools don't wait for a policy to catch up, they just make the ungoverned version easier to reach for.