xAI cut frontier-model pricing pressure, Amazon blocked Meta’s shopping agent, and Novo Nordisk showed what a narrow internal assistant can save. Today’s issue spans the model market, agent access and one working deployment.
The expensive end of the model market has a new price problem. xAI released Grok 4.7 on Sunday with a larger base model and longer reinforcement training.
Its training also put more weight on tasks that take hours rather than minutes.
The model starts at $2 per million input tokens and $6 per million output tokens, the same rate as Grok 4.6.
It is available through the Grok API, Cursor, Grok Build, third-party coding tools, routers and cloud platforms.
xAI’s own tests put the trade-off in view:
Grok scored 46.3% on CursorBench, ahead of GPT-5.6 Sol at 41.7% but behind Fable 5.1 at 51.8%.
On a terminal-work benchmark, Grok reached 38.0%, while Fable reached 57.9%.
xAI says the model allows 3.3% of risky cyber prompts through while rarely blocking legitimate security work.
The company is pitching twice the speed at half the price of comparable models. Those are vendor benchmarks, not a verdict across every production workload, but the pricing is live now.
Founders context: Re-run a representative week of coding or document work against Grok before moving anything. Compare total task cost, correction time and failure rate, not benchmark rank alone.
The fight over who owns an AI-assisted purchase has reached the checkout. Amazon cut Meta’s Muse agent off from shopping on Amazon.com.
Amazon says Meta neither asked permission nor made the agent identify itself while browsing customer accounts.
Muse launched on 8 September and can use a service’s public API or operate its website through a browser.
Meta says credentials sit in secure storage and remain hidden from the model. Amazon argues that the agent still handles sensitive account data without the retailer’s consent.
The commercial tension is harder to miss:
Muse reached number one in Apple’s US free-app chart a week after launch.
Early users reported switching insurance, filling grocery carts and finding checkout discounts.
Amazon made more than $68bn from advertising last year, a business tied to shoppers viewing its pages and sponsored products.
Amazon’s block cites its Conditions of Use. That matters after a US appeals court rejected its anti-hacking case against Perplexity in August while leaving contract claims open.
Founders context: Treat browser access as a fallback, not a distribution contract. Agents that identify themselves, use approved APIs and preserve an auditable consent trail have a better chance of surviving the next platform block.
Novo Nordisk’s rare-disease analytics team did not start with a company-wide chatbot. One senior analyst built a narrow assistant around the questions his team already answered from healthcare claims and prescription data.
The tool turns plain-language questions into database queries, runs against data refreshed each day and returns a first pass for an analyst to inspect.
Roughly three analysts in the US and India use it for launch analysis, market monitoring, KPI reviews and brand planning.
Dataiku reports the early operating effect:
First-pass answers arrive in under one minute.
Routine reviews and exploratory analysis take 50% to 70% less time.
The team projects a 50% overall efficiency gain and about half a full-time role of capacity reclaimed without another hire.
The human checkpoint stayed in place. Analysts validate the query, refine the interpretation and decide what reaches the commercial team.
The figures come from a Dataiku customer case study and are not independently audited. The workflow is unusually specific, with one expert, one recurring backlog, live internal data and a defined review point.
📡 Further Context
🧠 Jev introduces decision models. The small model turns unstructured text into typed scores for routing, ranking and classification, with output priced free and input at $0.042 per million tokens.
📐 OpenAI forms a mathematics advisory group. The independent group will give mathematicians a route into OpenAI’s research after the company said its systems had resolved more than 100 open problems.
💻 Google opens pre-orders for its Gemini-first laptop. Googlebook starts at $899, pairs Android apps with desktop Chrome, and gives Gemini screen access only when the user calls it.
🧑 IBM finds a judgement gap around AI at work. Seventy-one per cent of surveyed HR chiefs called supervising and overriding AI essential, while only 29% of employees ranked judgement as important.
🛡 UN panel calls for stronger agent safeguards. The new scientific panel wants governments to install protections before uncertainty is resolved as autonomous systems move into infrastructure, finance and public services.
🤖 AI Tool Context
Grok 4.7: Frontier model for coding and knowledge work, starting at $2 input and $6 output per million tokens.
Jev: Low-latency decision model for classification, ranking and routing rather than prose generation.
V7: Application platform that turns company files into source-linked context for multi-step agents.
Muse: Meta’s personal agent for email, calendar, shopping and other browser-based tasks.
