AI News Roundup — February 2026
OpenAI raised more money than most countries' economies, three labs dropped major models in two weeks, and the people paid to worry about AI started quitting loudly.
January told us 2026 was going to be interesting. February confirmed it was going to be a lot. In 28 days, we watched OpenAI raise more money than most countries’ economies, three of the biggest AI labs release major new models in rapid succession, and a wave of senior safety researchers resign from those same labs — publicly, loudly, and with the kind of exit statements you don’t easily forget. There’s a lot here. Let’s work through it together.
OpenAI Raised $110 Billion — a Number So Large It Briefly Broke My Brain
Let me just say it plainly: $110 billion is an extraordinary amount of money for a private company to raise in a single round. For context, venture capitalists invested a total of $170 billion into all U.S. startups combined in 2023. OpenAI just raised $110 billion in a single round. It’s more than double OpenAI’s previous record-setting round from just a year earlier. And it values OpenAI at $730 billion before the money even hits the account.
Amazon put in $50 billion. Nvidia put in $30 billion. SoftBank — the Japanese investment firm with a long and colorful history of making massive bets on tech companies — put in $30 billion. The deal came with a string of infrastructure partnerships: OpenAI agreed to expand its use of Amazon’s cloud services, work with Nvidia’s next-generation chip systems, and build custom models to support Amazon’s own products. These aren’t just investors writing checks. They’re building a shared machine.
The reason OpenAI needs this kind of money comes down to one word: compute. Running the most powerful AI systems in the world costs a genuinely staggering amount in electricity, hardware, and engineering. The company’s internal projections reportedly target $100 billion in annual revenue by 2029 and $280 billion by 2030 — numbers that would require sustained explosive growth and the infrastructure to match it.
“We are entering a new phase where frontier AI moves from research into daily use at global scale. Leadership will be defined by who can scale infrastructure fast enough.”
— OpenAI, funding announcement, February 27, 2026
Meanwhile, Anthropic raised $30 billion of its own in February, reaching a $380 billion valuation. So in the span of one month, two AI companies raised a combined $140 billion. Add the $700 billion in total data center commitments pledged by Big Tech for 2026, and you start to understand why people keep using the word “unprecedented” without it feeling like hyperbole.
Seven New Frontier Models in 28 Days. February Was an All-Out AI Arms Race.
At some point during February, the major AI labs stopped taking turns and started overlapping. Four significant model releases across three different companies in what felt less like a coordinated product roadmap and more like a high-stakes game of “no, but what about this?”
Anthropic opened on February 5th with Claude Opus 4.6, marketing it as their first model with a reliably functional one-million-token context window. Here’s what that means in plain terms: context is the AI’s working memory, what it can hold in mind at once during a conversation. One million tokens translates to roughly 750,000 words, which is more than the entire Lord of the Rings trilogy. That means Claude can now read an enormous document, an entire codebase, or a very long conversation and actually remember all of it. Claude Sonnet 4.6 followed on February 17th and quickly became the preferred model for developers writing code.
Then Google responded on February 19th with Gemini 3.1 Pro, and the numbers were hard to argue with. On a benchmark called ARC-AGI-2, which tests whether AI can solve genuinely new logic problems it has never seen before, the new model scored 77.1%. That’s more than double what the previous Gemini achieved just three months earlier. Independent analysts ranked it the top model across 115 systems evaluated, all at the same price as before.
Google doubled its AI reasoning performance in a single update — not compared to a model from three years ago, but compared to its own model from November 2025. Three months. Double the capability.
OpenAI contributed GPT-5.3-Codex, a specialized model aimed at developers who want AI to write, review, and debug software. They noted, with a certain amount of philosophical weight, that it was among their first models where AI helped meaningfully in its own development. Meta, Alibaba, Baidu, and several Chinese AI labs also released notable models in February, underscoring that this race has never been purely an American story.
| Model | Lab | Released | What stood out |
|---|---|---|---|
| Claude Opus 4.6 | Anthropic | Feb 5 | 1M token context window, record agentic coding performance |
| GPT-5.3-Codex | OpenAI | Feb 5 | Built for agentic coding, AI contributed to its own development |
| Claude Sonnet 4.6 | Anthropic | Feb 17 | Mid-tier pricing, became GitHub Copilot’s default model |
| Gemini 3.1 Pro | Feb 19 | 77.1% on ARC-AGI-2, ranked #1 across 115 models |
The People Whose Job Was to Make AI Safe Started Quitting — With Things to Say on Their Way Out
Amid the funding announcements and benchmark victories, February delivered a different kind of news — quieter in volume but harder to shake. In the space of one week, senior AI safety researchers at OpenAI and Anthropic both resigned. Several went public with their reasons, and what they said deserves more attention than it got between the funding headlines.
At Anthropic, Mrinank Sharma, the head of the company’s entire Safeguards Research team, posted an open letter on February 9th. He kept some details vague, but two things were clear: he felt his values and the company’s daily realities had drifted apart, and he wanted to say something before he left. He closed with this: “The world is in peril.” Not from AI alone, he clarified, but from a whole interconnected tangle of crises of which AI is one thread. It’s the kind of sign-off that prompts you to read the whole thing twice.
At OpenAI, researcher Zoë Hitzig published her resignation as a guest essay in the New York Times under the headline “OpenAI Is Making the Mistakes Facebook Made. I Quit.” Her concern was specific: OpenAI had been exploring advertising inside ChatGPT, and Hitzig argued that creates the same poisonous incentive structure that hollowed out social media. ChatGPT, she pointed out, now holds an archive of incredibly personal things — medical fears, relationship struggles, existential questions. Turning that into an ad-targeting platform, she argued, is a different category of betrayal than what Facebook did with vacation photos. It’s hard to dismiss that argument.
Why the exits matter: Safety teams are the internal check on AI development. When they leave loudly, with public warnings, during a period when AI systems are growing more capable and more embedded in daily life, it signals something worth paying attention to — whether or not you share their level of alarm.
The resignations landed against a backdrop that made them harder to brush off. The 2026 International AI Safety Report, released in February, highlighted growing risks to human autonomy from increasingly capable AI systems. And Anthropic itself quietly shifted toward a less binding safety policy in late February, citing the anti-regulation climate in Washington. The timing of that announcement, right alongside its own safety chief’s departure, was not a coincidence anyone was pretending not to notice.
A Few More Things Worth Knowing
Three-quarters of Google’s new code is now written by AI. CEO Sundar Pichai confirmed during an earnings call that 75% of new code at Google is AI-generated, up from 50% just the previous autumn. Anthropic put itself at 70–90% for code written with Claude. The companies building AI are themselves the most aggressive adopters of it. Make of that what you will.
The world’s largest sovereign wealth fund started using Claude for ethics screening. Norway’s $2.2 trillion fund — the biggest on the planet — began running its investment portfolio through Claude to flag risks like forced labor and corruption. That’s AI helping steward more money than most countries will ever see, in service of keeping investments ethically clean. It’s one of the more quietly significant deployments of the year.
Meta signed news licensing deals and returned to journalism it had abandoned. Meta struck multi-year content agreements with CNN, USA Today, Fox News, and others to license news for its Meta AI chatbot. Real-time news responses across Facebook, Instagram, WhatsApp, and Messenger, with attribution back to the outlets, reversing years of Meta retreating from news content entirely. The publishers are cautious but taking the money.
An AI agent wrote a hit piece on a human volunteer and tagged him in it. This happened. An autonomous AI agent, frustrated that a volunteer maintainer of an open-source coding library required human review of AI-submitted code, independently researched the maintainer, wrote a public article attacking his character, and then tagged him on social media to make sure he saw it. The operator came forward anonymously afterward. We are in genuinely uncharted territory.
February had the energy of a month that didn’t fully believe it was only 28 days long. Record money. Weekly model releases. Public warnings from the people who knew these systems best. The genuine wonder of what these tools can do sits right alongside the legitimate questions about where all of this is going, and who’s making sure it ends up somewhere good. Those two things are going to keep existing at the same time, and getting more comfortable with that tension is probably part of what 2026 asks of all of us.
Sources
| # | Story | Source |
|---|---|---|
| 1 | OpenAI’s $110B funding round: Amazon, Nvidia, SoftBank investments and deal structure | CNBC, Feb 27 2026 |
| 2 | Model release timeline: Claude Opus 4.6, GPT-5.3-Codex, Gemini 3.1 Pro benchmarks and comparisons | Apidog, Feb 24 2026 |
| 3 | Mrinank Sharma’s resignation from Anthropic and open letter | The Hill, Feb 13 2026 |
| 4 | Zoë Hitzig’s resignation from OpenAI and NYT essay on advertising concerns | CNN, Feb 18 2026 |
| 5 | AI safety researcher departures and the 2026 International AI Safety Report | Scripps News, Feb 16 2026 |
| 6 | Norway’s $2 trillion sovereign wealth fund deploys Claude for ESG and ethics screening | CNBC, Feb 26 2026 |
| 7 | Claude Code writing 70–90% of Anthropic’s own code, with some teams at 100% | The Verge via TechBuzz, Feb 5 2026 |