This week (February 16–20, 2026) had no single superstar announcement, but four major stories together draw a clear direction: security becoming a product feature, the default slot becoming a battlefield, price-performance becoming the language, and compute becoming the moat. Three of the four stories are about distribution and economics rather than raw capability — a sign of where the competition has moved. Here is the week, story by story.
ChatGPT Gets a Lock
On February 16, OpenAI introduced Lockdown Mode and Elevated Risk labels to ChatGPT, defending against prompt injection and data-exfiltration attacks. When an assistant browses and handles data on the user’s behalf, every piece of external content is a potential attack carrier — building detection and restriction into the product itself, documented and versioned, is the inevitable path for anything at this scale. Labeling plus lockdown is defense-in-depth, shipped as a product feature.
Claude Sonnet 4.6 Takes the Default Slot
On February 17, Anthropic released Claude Sonnet 4.6 and made it the default model for free and Pro users, with a knowledge cutoff of May 2025. The default slot means: most users never choose a model, the product chooses for them — the most expensive square on the board at every generational change, and the most contested. Free-tier users get the newest model on day one; the acquisition play is unmistakable.
Gemini 3.1 Pro Speaks Through Pricing
On February 19, Google released Gemini 3.1 Pro in preview: 2x reasoning versus Gemini 3 Pro at the same price, on the gemini-3.1-pro-preview API endpoint. As capability gaps between labs blur, Google moved the comparison axis to capability per dollar — the most direct way to monetize infrastructure scale. For API teams, that arithmetic lands directly in unit economics.
Meta and NVIDIA Sign a Million-GPU Deal
In the same week, Meta announced a long-term infrastructure partnership with NVIDIA in its official newsroom, involving millions of Blackwell GPUs. Compute lock-in is now a moat in frontier competition — a multi-year commitment counted in millions raises everyone else’s cost of entry one more notch. Capacity plans that assume renting at scale on short notice are now betting against the direction every hyperscaler is locking in.
The Week in One Look
- Security entered the release cadence: Lockdown Mode and Sonnet 4.6 shipped in the same week — defense is no longer an appendix to the docs but part of the version
- Default placement and price were the week’s two levers: Anthropic grabbed the default slot; Google set the price-performance tone with double-at-the-same-price
- The unit of infrastructure scale jumped again: from model races to GPU contracts, the cost floor of competing keeps rising
- Platform defaults set the pace: when an entire user base gets a new model in one switch, vendor confidence — and vendor risk — both scale with it
- Watch the second-order effects: risk labels reshape what enterprise buyers expect from every assistant vendor, and preview pricing reshapes every rival’s pricing page
- Three notes for developers: evaluate preview endpoints before committing, run regression tests when default models change, and put injection defense on your agent requirements list
Sources
- This Week in AI — SD Times
- Claude Sonnet 4.6 — Anthropic
- Gemini 3.1 Pro — Google
- ChatGPT Lockdown Mode — Help Net Security
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.