AI News Roundup: The Week of The Great Model Reset (Feb 20, 2026)

Welcome to Optijara's weekly AI roundup. Every Friday, we cut through the hype to tell you what actually matters for your business, your code, and your future.

What were the biggest AI releases this week?

The week of February 13–20, 2026, was defined by Anthropic's release of Claude Opus 4.6 and Sonnet 4.6 (which introduced a massive 1M token context window in beta), OpenAI's aggressive retirement of legacy models like GPT-4o and o4-mini from ChatGPT, and Google's expanding global education push.

We are seeing a clear shift from "more models" to "better, deeper thinking" models. The era of the lightweight chatbot is ending; the era of the reasoning agent is here.

1. Anthropic Dominates with Claude 4.6 Series

What is Claude Opus 4.6?

Claude Opus 4.6 is Anthropic's new flagship reasoning model, released alongside Sonnet 4.6. The new Sonnet 4.6 notably features a massive 1M token context window in beta and significant upgrades to agentic planning and coding.

While OpenAI has been tweaking parameters, Anthropic just dropped a hammer. The new Sonnet 4.6 isn't just a minor update; it's a full overhaul of the model's skill set.

1M Token Context (Beta): You can now dump entire codebases or legal archives into Sonnet 4.6.
Agentic Capabilities: Improved ability to plan multi-step tasks without getting lost in the weeds.
VS Code Integration: Claude Code received a major update, making it a true pair programmer rather than just a chatty sidebar.

Why this matters for Optijara clients: If you're building complex RAG (Retrieval-Augmented Generation) pipelines, the 1M context window on a mid-tier model like Sonnet changes the economics of development completely.

2. OpenAI's "Great Reset": Retiring the Legacy Models

Why did OpenAI retire GPT-4o and GPT-5 Instant?

OpenAI has officially retired GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and GPT-5 (Instant & Thinking) from ChatGPT to consolidate users onto the new GPT-5.2 architecture. This move appears to streamline their offering and encourages adoption of their most capable reasoning engine.

It’s a bold cleanup. Usually, legacy models linger for years. OpenAI is effectively burning the boats. On February 4, 2026, they also restored the "Extended Thinking" capability to GPT-5.2, fixing a regression that had limited the model's reasoning depth in January.

What’s gone from ChatGPT:

GPT-4o (The former flagship)
GPT-4.1 & 4.1 mini
o4-mini
GPT-5 Instant & Thinking (The early V5 releases)

What’s staying: GPT-5.2 is now the default standard. If you were building on the specific quirks of 4.1, it's time to migrate.

3. Google's Global Education Play

Who gets free Gemini upgrades in 2026?

Google is offering free upgrades to its most advanced Gemini models for students over 18 in select countries, including Indonesia, Japan, the UK, and Brazil (with rollout phasing across US/Korea) through July 2026. This program aims to lock in the next generation of power users on the Google ecosystem.

While the model wars rage at the high end, Google is playing a distribution game. By giving students in key emerging and developed markets free access to their best tools, they are betting that familiarity will breed loyalty (and future enterprise contracts).

Also notable: Gemini for Google Workspace is rolling out enhanced usage reporting for admins starting February 16, 2026, giving enterprises better visibility into how their teams are actually using AI.

4. Gemini 3.1 Update (Feb 19, 2026)

What changed with Gemini 3.1 this week?

Google announced Gemini 3.1 Pro on February 19, 2026, and began rolling it out across developer and consumer surfaces, including Gemini API (preview), Vertex AI, the Gemini app, and NotebookLM.

The practical signal is clear: Google is positioning 3.1 Pro as its upgraded reasoning baseline for complex workflows, not just chat. For teams, this matters most in planning-heavy tasks like multi-step analysis, coding copilots, and knowledge synthesis where reliability across longer task chains is more important than one-shot responses.

5. The Rise of "Mixture of Experts" for Coding

What is the MiniMax model?

MiniMax is a new "Mixture of Experts" (MoE) model released this week that achieves competitive coding benchmarks at a fraction of the inference cost. MiniMax claims their pricing structure could allow enterprises to run approximately 4 continuous autonomous instances for $10,000/year.

This is the trend to watch. We aren't just getting smarter models; we're getting cheaper smart models. The MiniMax release proves that specialized MoE architectures can compete with massive dense models for specific tasks like coding and operations.

FAQ: Your Questions, Answered

Is GPT-5.2 better than Claude Opus 4.6?

It depends on the use case. Early benchmarks suggest GPT-5.2 holds an edge in pure reasoning depth for logic puzzles and math, while Claude Opus 4.6 is often preferred for creative writing and massive context handling (1M tokens).

What happened to GPT-4o?

GPT-4o has been retired from the ChatGPT interface as of February 2026. OpenAI has replaced it with GPT-5.2 to streamline their model lineup, though the API generally remains available for developers.

Can I still use GPT-4 via API?

Yes, for now. While the models have been removed from the ChatGPT consumer interface, OpenAI typically supports legacy API endpoints for a longer sunset period to avoid breaking enterprise applications.

What is a "Mixture of Experts" model?

A Mixture of Experts (MoE) model is an AI architecture that uses multiple specialized sub-models (experts) and only activates a fraction of them for any given prompt. This makes the model faster and cheaper to run than a traditional dense model of the same size.