AI Leadership Weekly · Issue #100 · Tuesday 8 September 2026 · 08:00 BST

Good morning.

Issue 100 arrives with a suitably busy model week. OpenAI introduces GPT-6 Astra, Anthropic updates two differently controlled model offerings, and Google presents a new weather-forecasting system.

The temptation is to put them on one scoreboard. The more useful approach is to ask three different questions: what new work could become feasible, under what operating terms, and where does a specialist system serve a decision better than a general assistant? A model name can make a headline. It cannot, by itself, write the business case.

Capability: GPT-6 Astra introduces another frontier-model option, with staged availability. Access: Anthropic distinguishes broadly available and restricted offerings. Specialisation: WeatherNext 3 targets forecasting rather than general office work. This week’s task is to connect an announced capability to an actual decision—and verify the terms under which you could use it.

https://openai.com/index/gpt-6-astra/

OpenAI · 3 September 2026

What happened. OpenAI launched GPT-6 Astra on 3 September, describing stronger multi-step work and computer-use capabilities. Access began with a limited set of organisations, with broader paid-plan and API availability planned over the following days. Enterprise access required administrator enablement at launch.

Why it matters. Our take: revisit tasks that failed for a clear capability reason, not every task in the portfolio. Perhaps a process required too much intermediate reasoning, or a long sequence repeatedly lost its place. A stronger model deserves a controlled second look at those cases. It does not automatically justify replacing a cheaper, satisfactory system already doing routine work.

What to do. Take a previously unsuccessful test with an agreed expected result. Confirm your organisation has approved access, then rerun the task with comparable inputs and a review of the full output. Record what improved, what still failed and the actual cost of reaching a usable result.

https://www.anthropic.com/claude-fable-and-mythos-5-1

Anthropic · 1 September 2026

What happened. Anthropic announced Claude Fable 5.1 and Mythos 5.1 on 1 September, describing a shared underlying model with different safeguards and access arrangements. Fable is generally available; Mythos is restricted to trusted access. The separate Enterprise Frontier Safeguards offering was planned for later in the autumn.

Why it matters. Our take: purchasing capability also means purchasing an operating arrangement. Who can use it, which activities are permitted, what information is retained and how review works can determine whether the service fits the organisation. A future customer-controlled option may be useful, but it cannot satisfy a requirement that must be met before today’s pilot starts.

What to do. Ask for the terms applicable to the exact product, account and deployment route being considered. Mark promised future features separately from current controls. Have the relevant internal owners decide whether the present configuration is suitable before introducing sensitive or business-critical work.

https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/

Google DeepMind · 3 September 2026

What happened. Google DeepMind introduced WeatherNext 3 on 3 September, using live geostationary satellite information and hourly refreshes. Some forecast variables reach five-kilometre resolution; that resolution does not apply to every output. The announcement includes routes into Google’s cloud services.

Why it matters. Our take: the business value of a forecast lies in the decision it improves. Logistics, staffing and asset-management teams may care about different locations, horizons and tolerances. A more detailed prediction is not automatically a better operational plan. The useful experiment connects a forecast to a specific choice, then compares that choice with the existing approach.

What to do. Choose one weather-sensitive operational decision and involve the person who currently makes it. Define the location, timing and consequences of being wrong. Evaluate relevant historical or controlled examples before relying on a new feed for live decisions, and preserve an appropriate fallback.

27 August: Anthropic’s hardware-interface proposal remains a research preview, not an established interoperability guarantee. Anthropic · 27 August 2026

26 August: OpenAI’s postmortem gives buyers concrete containment questions to ask about agent experiments. OpenAI · 26 August 2026

25 August: New wellbeing research funding should be read as a research commitment, not a settled result. Anthropic · 25 August 2026

27 August: Support for scientific teams adds another example of access being shaped around a particular user group. Anthropic · 27 August 2026

Further dated updates, including recent context worth keeping in view.

2 September: Google launches Gemini 3.8 Flash; the Cyber variant remains a restricted-access offering. Google · 2 September 2026

2 September: IBM publishes a US K–12 readiness study; its findings should not be relabelled as UK workforce evidence. IBM · 2 September 2026

3 September: Google announces voice features for Gmail, Docs and Keep, with product-specific rollout conditions. Google Workspace · 3 September 2026

3 September: OpenAI announces subsidised cybersecurity access and support for frontline defenders, initially in the US. OpenAI · 3 September 2026

Minutes 0–4: Find a use case previously rejected because the technology could not produce an acceptable result. Write down the actual failure, not a general impression that the model was weak.

Minutes 4–8: Check whether this week’s announcement addresses that failure and whether an approved version is available to you. A promised release does not count as a test environment.

Minutes 8–12: Prepare the original example and the acceptance criteria. Keep the information appropriate for the tool, and include a difficult case rather than only the demonstration that almost worked.

Minutes 12–15: Assign a bounded retest. Decide in advance whether success would justify more evaluation or a deployment decision. Those are different thresholds; clearing the first should not silently be treated as clearing the second.

Sometimes a new model changes what can be built. Sometimes a better interface or a clearer process changes more. At Data Wave, we keep those possibilities separate so that the next investment responds to a real constraint. Ambition is useful when it leads to a specific test, a usable service and an honest account of the work still required.

Which previously rejected use case deserves another look—and what evidence would make you change your mind?

Brought to you by Data Wave — your AI & Data Team as a Subscription.

Explore Data Wave