AI Leadership Weekly · Issue #71 · Tuesday 17 February 2026 · 08:00 GMT

Good morning.

This week offers two quite different answers to “better AI”: more reasoning for difficult scientific and engineering problems, or faster interaction on coding work. Google and OpenAI are addressing those needs through different products and access conditions. Anthropic’s funding announcement completes the picture from the supplier side. Capital, capability and response time all matter, but they are not interchangeable buying criteria.

Gemini Deep Think expands. Google offered its updated reasoning mode to AI Ultra subscribers and invited interest in early API access, not general API availability.

Codex-Spark targets fast coding. OpenAI introduced GPT-5.3-Codex-Spark as a research preview for Pro users, initially focused on text-based, low-latency coding.

Anthropic raises fresh capital. Anthropic announced a $30 billion Series G funding round at a $380 billion post-money valuation.

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

Google · 12 February 2026

What happened. On 12 February, Google announced an updated Gemini 3 Deep Think aimed at science, research and engineering. It made the mode available in the Gemini app to Google AI Ultra subscribers and invited interest in early API access. That invitation was not general API availability, and the announcement did not establish that its outputs were independently validated engineering decisions.

Why it matters. Our take: There is a useful distinction between generating a promising approach and authorising a real-world change. Put a specialist reasoning tool beside someone who understands the problem, the evidence and the limits of a plausible answer. The first benefit to test might be a better set of hypotheses, rather than a finished design requiring less expertise.

What to do. Ask an experienced colleague to choose one unresolved but bounded problem. Supply the assumptions and evidence, then compare proposed approaches against their own assessment. Record which suggestions were useful, which were unsupported and what independent work would be needed before acting. Keep the trial away from live operational decisions.

https://openai.com/index/introducing-gpt-5-3-codex-spark/

OpenAI · 12 February 2026

What happened. OpenAI introduced GPT-5.3-Codex-Spark on 12 February as a research preview for ChatGPT Pro users. Built for responsive, interactive coding and served on Cerebras infrastructure, it launched with text-only support and separate usage limits. Its launch positioning concerns the pace of the interaction, not proof of a faster end-to-end software release.

Why it matters. Our take: A developer may value keeping a train of thought while making a small change more than receiving a larger answer later. Test that proposition without mistaking generation speed for delivery speed. The useful measure is how quickly a change becomes understandable, tested and accepted. A fast suggestion that creates a long investigation has moved the delay, not removed it.

What to do. Pick a familiar, low-risk maintenance task. Compare the full time from request to reviewed change, including tests and corrections. Ask the developer whether the interaction helped them stay in control. Do not extrapolate the result to a large migration or an unfamiliar codebase without another test.

https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation

Anthropic · 12 February 2026

What happened. Anthropic announced a $30 billion Series G funding round on 12 February and said the capital would support research, products and infrastructure. This was a company announcement about financing and intended investment. It was not an independent assessment of service resilience or a commitment to a particular customer’s capacity, pricing or support arrangements.

Why it matters. Our take: Keep the procurement question concrete: what service are you buying, on what terms, and what happens when it is unavailable? A major investment can change the conversation with a supplier, but it does not answer those questions on your behalf. Avoid allowing enthusiasm about company scale to replace an agreed recovery route for an important process.

What to do. For one important AI dependency, ask the service owner to show the support route, usage limits, recovery procedure and last fallback test. Identify what can continue without the supplier. Put any missing commitment into the commercial discussion rather than assuming that a financing headline supplies it.

27 January · Prism puts AI inside scientific writing. OpenAI launched Prism, a free research-writing workspace powered by GPT-5.2 for personal accounts; organisational-plan access was still forthcoming. Source

26 January · Microsoft introduces Maia 200. Microsoft unveiled an accelerator designed for AI inference, with initial deployment in its US Central datacentre region. Source

22 January · Anthropic publishes Claude’s constitution. Anthropic set out the principles intended to shape Claude’s behaviour. The document describes training intentions rather than guaranteed outcomes. Source

20 January · ChatGPT adds age prediction. OpenAI described age prediction for consumer accounts, with additional protections for suspected under-18 users and a route to correct mistakes. Source

More dated updates from the preceding 30 days.

5 February · Opus 4.6 is released. Anthropic launched Opus 4.6, including a one-million-token context window in beta and updated capabilities for longer assignments. Source

5 February · OpenAI introduces Frontier. OpenAI announced an enterprise platform for developing and managing agents, initially working with a limited group of customers. Source

4 February · More agents come to GitHub. GitHub opened a public preview of Claude and Codex coding agents for eligible Copilot users within its existing development environment. Source

2 February · Codex gets a desktop app. OpenAI launched a macOS app for managing parallel coding-agent tasks. Windows availability was not part of the initial release. Source

Minutes 0–4 · Choose three real requests. Take one routine request, one interactive assignment and one difficult analytical problem. Describe what an acceptable answer must accomplish. Do not classify them merely by document length or the seniority of the person asking.

Minutes 4–8 · Set the service expectation. For each request, agree an acceptable wait, a spending limit and a reviewer. Include the cost of checking the result. Name the circumstances in which a quick answer would be less useful than a careful one.

Minutes 8–12 · Define the escalation. State what sends work to a more capable tool or a person: missing evidence, disagreement, repeated failure or a consequential decision. Ask how the recipient will see the original brief and the reason for escalation.

Minutes 12–15 · Run and compare. Try the three requests and record accepted outcomes, elapsed time and review effort. Adjust the rule where it sends work to the wrong place. Keep the first version small enough that colleagues will actually use it.

Use capability selectively. The practical goal is not to put the most elaborate model behind every task; it is to make the whole service proportionate, understandable and useful to the people doing the work.

Which of our tasks genuinely needs deeper reasoning—and which simply needs a prompt, dependable answer?

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

Explore Data Wave