Hello Project AI enthusiasts,
OpenAI’s release of GPT-6 Astra has made cyber capability and monitorability an immediate operating issue. This week also brings Caterpillar’s $100M workforce response to autonomy, cheaper Claude use with fewer safety interruptions, a reported move towards outcome-based AI pricing, and Anthropic’s attempt to give agents a common language for physical equipment. Together, they make one point clear: AI strategy is becoming an operating model decision.
Coming up
We're doing another open AMA — Thursday 17 September, 5.15pm UK, live on LinkedIn. No slides, no script, bring the awkward questions.
Featured This Week
GPT-6 Astra: OpenAI ships a model it rates Critical for cyber risk
OpenAI released GPT-6 Astra on 3 September and rates it ‘Critical’ for cyber risk. Project Flux reports that the model found previously unknown software vulnerabilities without guidance and can sustain software work for roughly 40 minutes. The release is staged, and OpenAI itself acknowledges that Astra’s recurrent reasoning makes its behaviour harder to monitor.
That combination matters to delivery firms. The practical question is no longer whether an AI assistant can draft, analyse or automate. It is whether the firm has access controls, escalation paths and continuity plans for tools whose capabilities and risks can change between product releases. Treat frontier-model access as a controlled service, especially where code, client data or connected equipment are involved.
Editor’s Picks
This week's must-read stories on AI, project delivery, and infrastructure:
Caterpillar’s $100M answer to autonomy: retrain the operator
Caterpillar plans to spend $100M over five years training its workforce in AI, autonomy and robotics, while developing AI assistance for equipment. With 118,000 employees, 1.6 million connected assets and years of automated-mining experience, it is treating capability as part of the deployment programme. For contractors, the useful signal is that plant autonomy changes operator roles before it removes them.
Claude Fable 5.1: cheaper tokens, fewer safety blocks
Anthropic’s Fable 5.1 release cuts typical workload costs by about 25%, according to the company, while cache reads fall to $0.25 per million tokens. Project Flux’s useful caveat is that the saving depends on how much a workflow reuses context. Anthropic also reports fewer false-positive safety interventions, which may reduce abandoned runs in long document, spreadsheet and research workflows.
OpenAI tests outcome pricing: a warning for the billable hour
OpenAI is reportedly piloting outcome-based pricing with a small group of large customers, charging when the AI completes a task rather than for tokens consumed. The company has not confirmed the reported terms. Even so, the experiment is relevant to consultancies and delivery firms that still price primarily by inputs: vendors are testing the same shift from effort to results that clients may eventually expect from professional services.
Anthropic’s Model Hardware Standard: a common language for machines
Anthropic has opened a research preview of the Model Hardware Standard, a model-agnostic specification intended to let agents discover and operate programmable devices. It is plumbing rather than a robot, but that distinction is important. The company says integrations that take weeks or months can fall to hours or minutes. Site sensors, plant and testing equipment need that interoperability before agentic workflows can move beyond reporting.
AI in AEC and Projects
Here are some intriguing reads, specifically for AEC and project delivery professionals:
Deere’s construction and forestry sales rose 18% to $3.6bn in its latest quarter, while its order book runs into fiscal 2027. Data centres and large infrastructure projects are now an equipment-demand signal, not only a real-estate story.
ABC found that 36% of contractor respondents use drones, and 97% of those users rely primarily on DJI. Progress capture and inspection teams should map this vendor dependency before regulatory or procurement conditions force a change.
QualiT is positioned as a quality tool that combines AI with human judgement to capture project-specific decisions. Its proposed link to Building Safety Act information requirements makes accountable review, rather than automation alone, the central test.
JE Dunn has created a dedicated power generation unit under Ryan Wilson, citing AI infrastructure, onshoring and ageing assets. The contractor’s portfolio already includes more than 220 power projects, showing how the grid constraint is reshaping delivery capability.
Curated Links
Selected stories worth reading before you get back to delivery:
Cursor-agent breach exposes a social-engineering control gap. Reuters reports that hackers induced a coding agent to assist malicious operations. Treat prompts, permissions and agent logs as security controls, not interface details.
Unitree G1 security research raises a robotics data question. A public security assessment reports telemetry transmission and exploitable provisioning flaws. Firms trialling connected robots should include network architecture and patch status in their site assurance.
WeatherNext 3 brings higher-resolution AI forecasting to Google products. Google says the model refreshes hourly with live satellite data. Site teams should watch for practical planning uses while retaining local meteorological advice for official warnings.
World Labs introduces Atlas for spatial world generation. The company’s early-access world model accepts text, images, video and 3D inputs. Its value for AEC depends on whether spatial fidelity survives real project data and review processes.
GE Appliances shows how AI is used on a factory line. An NPR report shows AI-assisted inspection, autonomous parts delivery and real-time operations data working alongside people, not replacing them outright.
Tool of the Week
AutoRFP.ai for bid and tender responses
AutoRFP.ai uses semantic search across a content library to help teams assemble and tailor tender responses. In a vendor-published case study, Perk reports doubling RFP volume while maintaining headcount and saving 949 hours. Treat those figures as customer claims, then test the product on a limited set of approved historic bids. Check source provenance, answer accuracy and review ownership before loading live client material.
Governance
ChatGPT becomes a ‘Very Large Online Search Engine’ under the DSA
The European Commission designated ChatGPT a ‘Very Large Online Search Engine’ on 31 August. The designation is VLOSE, not VLOP, and gives OpenAI four months to meet the relevant additional Digital Services Act obligations by January 2027. European firms using ChatGPT in client-facing services should monitor the resulting product and policy changes, particularly where risk assessment and user safeguards affect their own controls.
Also This Week
Nate Amidon shares his unique journey from a C-17 pilot to founder of Form 100 Consulting. [Podcast]
One More Thing
Dyson’s CameraJet puts machine learning in a toothbrush
Dyson’s new CameraJet uses a 100,000-pixel macro camera and machine learning to identify gaps between teeth while brushing. The company says it scans 28 live images each second and trained the system on 470,000 dental images. A Wi-Fi-connected toothbrush is a generous interpretation of personal AI, but it is also a reminder that computer vision is moving into ordinary physical products.
Quick Win
Put a live Google Drive file beside ChatGPT
ChatGPT’s Library can surface connected Google Drive files and keep a Google Doc, Sheet or Slide open beside a conversation. Where supported and authorised, ChatGPT can update the source file; the initial experience excludes Shared Drives. Start with a non-sensitive monthly cost report or risk register. Ask for a comparison against the previous version, review the proposed changes, and keep the original record intact until a responsible project team member approves the update.
Event of the Week
COMIT Conference 2026, 9 to 10 September, London
COMIT’s two-day conference at the Institution of Mechanical Engineers focuses on AI, industrialised construction, advanced work packaging and robotics. The agenda is built around productivity in UK construction, which makes it a practical venue for teams comparing technology claims with delivery methods. The event falls within this edition’s seven-day window.
📊 Hiring: Power BI & Project Data Reporting Specialists
Pace Global is recruiting data reporting talent for major project delivery programmes across London, Manchester and Birmingham.
If you know your way around Power BI and you've worked on the data behind real projects — schedules, cost, performance — they want to talk to you.
£45k–£70k · permanent · project controls
See the roles → https://www.paceglobal.org/hot-jobs
Till next time,
Project Flux
All content reflects our personal views and is not intended as professional advice or to represent any organisation.