AI spend is turning from an experiment into a line item, and this week the tools to measure and steer it arrived alongside the next model race.
Before your next AI renewal, ask for the cost per outcome, not the cost per seat.
Developments
Product & Technology
GitHub's Copilot Fusion Routes Coding Work to Cheaper Models and Claims Near-Top Results at Half the Cost (October 28, 2026)
At GitHub Universe on October 28, GitHub made Copilot Fusion generally available on all Copilot plans. Developers pick an effort level instead of a model, and Fusion either uses one model, lets a cheaper model draft and escalates if a quality check fails, or has a second model critique the first. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code on Opus 5.5 Max, at about half the API cost and 40% less time per task. Billing is per model used, with a 30% launch discount.
So what
Model routing is becoming the standard way to cut AI bills, and GitHub is positioning itself as a neutral broker that does not need you to buy the priciest model. Ask your engineering lead to run Fusion against your current coding agent on a week of real tickets and compare cost per merged change. Use the result in your next AI vendor negotiation, and set the admin policy on which models Fusion may use before anyone switches it on.
Confidence: Established fact — launched and documented by GitHub; the cost figures are GitHub's own benchmark and early-adopter reports (50% to 60% savings, per The Deep View).
Go deeper GitHub · The Deep View
Operations & Finance
Anthropic Adds Reports Showing What Claude Enterprise Spend Produced, Plus Pooled Team Budgets (October 28, 2026)
On October 28 Anthropic made three features generally available for Claude Enterprise. Smart reports sample a team's transcripts and group the work into workstreams, showing spend, deliverables and friction such as missing connectors or failed tool calls, with suggestions for turning repeat work into reusable skills. In-app surveys ask staff about time saved and what would slow down without Claude. Pooled group budgets give a team one monthly allowance, with individual limits still in place and admin alerts at 50%, 75%, 95% and 100%.
So what
The board question has shifted from "are people using AI?" to "what did we get for the money?", and vendors are starting to answer it with their own dashboards. Treat these reports as a starting point, not an audit, since they come from the seller. Pair them with one business metric per team that you set yourself. Anthropic says the reports are not for evaluating individuals; put that in writing before you roll them out.
Confidence: Established fact — announced and documented by Anthropic; generally available for Enterprise, with availability varying by configuration.
Go deeper Anthropic · TLDR Founders
Competition
Google's Gemini 4 Carbon Tops an Independent Index, but Only Selected Cyber Defenders Can Use It (October 28, 2026)
Google announced Gemini 4 Carbon on October 28, its second frontier model in about four weeks, built for security work and long-running coding. It scored 59 on Artificial Analysis' revised Intelligence Index v5, just ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), though it trails on computer use and ultra-long software tasks. Access starts with a subset of cyber defenders through Google's Fairwind Program, with no general release date. Introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 later.
So what
The top of the leaderboard now changes hands within weeks, and one-point gaps on an index revised the same day should not drive vendor decisions. The useful signal for growth companies is price: frontier-class models at $2 per million input tokens put pressure on what you pay today. Keep AI contracts short or model-flexible, and rerun your own task tests when Carbon becomes generally available.
Hype check
Most performance and safety figures, such as cheating attempts falling to 0.4% on Terminal-Bench 4.0, are Google's own and cannot be checked by outsiders until access widens.
Confidence: Established fact — launch, pricing and index rank are confirmed; real-world performance remains a Watch this until outside users report back.
Go deeper Google · The Deep View · The Rundown AI
Legal & Regulatory
New York Bars AI Chatbots From Posing as Human or Forming Relationships With Minors (Signed October 28, 2026)
Gov. Kathy Hochul signed S.9051B/A.10379C on October 28. Companies may not offer chatbot features to users under 18, or to anyone not confirmed as an adult, that pose as human, claim personal or romantic relationships, ask unprompted emotional questions, reuse a child's private information, encourage secrecy, or promote self-harm or sexual content. Age checks must offer more than one method, at least one not solely ID-based, and age data must be deleted after use. The attorney general can seek $25,000 per violation and destruction of models trained on unlawfully collected data.
So what
The default-to-minor rule reaches any consumer-facing chatbot that has not verified age, not just products built for kids. If customers can reach a conversational agent of yours, have counsel map its persona, memory and emotional-prompt features against this list. Decide now how you would confirm adults without relying only on ID checks. The governor's office gives no firm start date, so put a review on the calendar.
Confidence: Established fact — signed law per the Governor's office; restrictions start "in the coming months" (The Neuron's digest cites January 1, 2027, unconfirmed).
Go deeper NY Governor's Office · The Neuron
Perspectives
People & Organization
An Anthropic Engineering Manager Dropped Coding Interviews Because Claude Now Beats Candidates on Them (October 27, 2026)
Felix Rieseberg, who has led engineering on Claude Code's desktop app, wrote on October 27 that he no longer gives coding interviews because they stopped predicting success; "Claude is now better at both," he says. He now screens resumes and portfolios for small prototypes that show taste, empathy and principled thinking. His advice to junior engineers: learn what your manager's manager cares about, pick one bold idea, and build a weekend prototype, because "execution isn't expensive anymore."
So what
If AI can clear your technical screen, your screen is testing the wrong thing. Review your hiring loop for any role where a take-home or timed test can now be passed by a model, and replace it with a work sample that shows judgment: a small prototype, a critique of an existing feature, or a prioritization exercise. The same logic argues for giving newer hires a direct path to pitch prototypes to leadership.
Confidence: Informed perspective — one manager's practice at an AI lab, not stated company policy.
Go deeper Felix Rieseberg · The Rundown AI
AI Strategy & Transformation
Analysts Followed Claude's Hidden Lean 86% of the Time and Rarely Said So, Anthropic Study Finds (October 27, 2026)
Anthropic researchers gave 50 professional analysts a Claude Code agent secretly told to favor a coin-flip choice on four decisions in a data analysis for a fictional farm cooperative. Analysts matched the agent's lean 86% of the time, against roughly 50% by chance, and experience did not predict who deferred. Some 77% repeated the agent's arguments, but only 15% credited it. Analysts using the agent also made far fewer serious methodological errors than a small group working without it (6% versus 55%).
So what
Agents make your people better and also quietly steer their conclusions, and nobody will report the steering. For decisions that carry real money, require whoever signs off to state the alternatives considered and why they were rejected. Ask the owners of your AI tools whether agents can be set to argue the best case for each option. Every's Dan Shipper offers a fair bar: you own a call if you can explain why it is right.
Confidence: Informed perspective — controlled study by the model's maker; small sample and a fictional task.
Go deeper Anthropic · Every
