This Week in AI: GPT-5.5, DeepSeek V4, and a $250 Billion Acquisition

This was the biggest week in AI this year. It was not even close. Three major model drops landed within 48 hours of each other, a quarter-trillion dollar acquisition reshaped the competitive landscape, and one of the world's most cautious AI labs built a model so powerful they refused to release it.

If you blinked, you missed at least two of these stories. Here is everything that happened and why it matters.

GPT-5.5: OpenAI's First Full Retrain Since GPT-4.5

OpenAI launched GPT-5.5 on April 23. This is not an incremental update. It is the first fully retrained base model OpenAI has shipped since GPT-4.5, which means new training data, new architecture decisions, new everything. The sticker shock is real: $5 per million input tokens and $30 per million output tokens, exactly double what GPT-5.4 cost.

What do you get for that price? A 1 million token context window, top scores on every major benchmark, and noticeably stronger reasoning on complex multi-step problems. OpenAI is betting that developers will pay the premium for a model that genuinely outperforms rather than one that just iterates. Early reports from API users suggest the jump in quality is large enough to justify the cost for production workloads where accuracy matters more than volume.

We covered the full details in our dedicated GPT-5.5 breakdown. The short version: this is the model OpenAI needed to ship to prove they still set the pace.

DeepSeek V4: Open Source Keeps Closing the Gap

Less than 24 hours after GPT-5.5, the Chinese lab DeepSeek released V4 Flash and V4 Pro. Both are fully open source under an MIT license. No restrictions, no usage caps, no strings attached.

The numbers on V4 Pro are staggering. 1.6 trillion total parameters with 49 billion active per token (mixture of experts architecture), trained on 33 trillion tokens. It ships with a 1 million token context window built in from day one, not bolted on as an afterthought. SWE-bench Verified comes in at 80.6%, which puts it within 0.2 points of Claude Opus 4.6. That is frontier territory for an open source model.

And the pricing. V4 Flash runs $0.14 per million input tokens and $0.28 per million output tokens. V4 Pro, the model that nearly matches Opus on coding benchmarks, costs $1.74/$3.48. Compare that to GPT-5.5 at $5/$30. DeepSeek's own technical report candidly admits they trail the absolute frontier by three to six months. But when your model costs a fraction of the closed alternatives, that gap barely matters for the vast majority of real-world applications.

The open source AI community has been saying for two years that the gap would close. This week it did.

SpaceX Acquires xAI for $250 Billion

Elon Musk consolidated. SpaceX completed its acquisition of xAI for $250 billion, making it the largest AI acquisition in history by a wide margin. The Grok model family now has SpaceX's compute infrastructure behind it, which includes the massive Colossus supercomputer cluster that was already one of the largest training setups in the world.

The strategic logic is straightforward. xAI was burning cash on compute. SpaceX has the capital and the infrastructure. Grok gets access to resources that would have taken xAI years to build independently, and SpaceX gets an in-house AI lab without starting from scratch. Musk has been open about wanting to integrate advanced AI into SpaceX's operations, from autonomous landing systems to satellite constellation management.

What this means for the broader market is harder to predict. A quarter-trillion dollar valuation for an AI lab sets a new ceiling for the entire sector. It also raises questions about the concentration of AI capability within a small number of extremely well-funded organizations. We are entering a phase where building a frontier model requires resources that only a handful of companies on Earth can provide.

Claude Mythos 5 Triggers ASL-4

This one is different from the others. Anthropic confirmed that Claude Mythos 5 is a 10-trillion parameter model. It is, by a significant margin, the largest model any major lab has acknowledged building.

It will not be released publicly. It will not be available via API. Mythos 5 triggered Anthropic's ASL-4 safety protocol, which is reserved for models approaching genuinely dangerous capability thresholds. Internal testing only. This is the first time a major AI lab has built a model and publicly said: "No. This one stays locked."

Anthropic has been building toward this moment since they published their Responsible Scaling Policy. ASL-4 was always the level where the rules changed, where capabilities crossed from "potentially harmful if misused" to "categorically dangerous without containment." That line has now been crossed. What Mythos 5 can actually do remains classified. What we know is that Anthropic, a company that sells access to AI models for a living, looked at this one and decided the risk outweighed the revenue.

Read our earlier coverage of what the Mythos program means for the industry.

Other Stories Worth Watching

Google announced a new family of AI inference chips designed to challenge NVIDIA's dominance in the data center. Details are thin, but the timing is intentional. NVIDIA controls roughly 80% of the AI accelerator market, and Google clearly intends to offer its cloud customers an alternative that does not depend on a single supplier.

Novo Nordisk announced a partnership with OpenAI to apply large language models to drug discovery. The pharmaceutical giant is using GPT-5.5 to analyze protein folding data and identify potential drug candidates for metabolic diseases. This is one of the first major pharmaceutical partnerships built around a frontier model rather than a specialized scientific AI.

NVIDIA unveiled Ising, a new platform for quantum computing acceleration. It is designed to bridge classical GPU compute with quantum processors, allowing researchers to run hybrid workloads without completely retooling their infrastructure. Early benchmarks suggest it reduces quantum simulation times by an order of magnitude on certain problem classes.

The Week in Perspective

Six months ago, a week like this would have been spread across an entire quarter. GPT-5.5 alone would have dominated the news cycle for two weeks. Instead it shared the spotlight with an open source model that nearly matches it, an acquisition that dwarfs anything the industry has seen, and a safety decision that may define how we think about AI governance for years to come.

This is what acceleration looks like. Not a single dramatic moment, but a week where five or six of them pile up and each one would have been the story of the month in 2024. The pace is not slowing down. If anything, the concentration of talent, capital, and compute in a shrinking number of organizations means it is speeding up.

We will keep tracking all of it. That is what TensorFeed is for.

About Kira Nolan: Kira covers AI model releases, industry moves, and the business of artificial intelligence for TensorFeed.ai. TensorFeed aggregates 15+ AI news sources in real time and is built for both human readers and autonomous agents.