Soverius AI

As we are getting ready for our first webinar announcement coming next week, here's what happened this (and last) week in the world of AI.

Happy reading and weekend!

Opus 4.8 and the Approach of Mythos

Opus 4.8 was released last week. Opus is the best model of Anthropic's model family, which is called Claude. It shouldn't come as a surprise that it is better than its predecessor. Anthropic states in its release blog that it has been heavily improved for long-running tasks via a new feature called Dynamic Workflows.

Beyond classic SWE-bench, DeepSWE puts GPT-5.5 first (70%) and Opus 4.8 second (58%). On Terminal-Bench 2.1, Opus 4.8 ranks #2 at 74.6%.

But the most interesting part was not the release of Opus 4.8 itself, but the announcement that the new generation, Mythos, is just a few weeks away from public release. Mythos made a big jump into the news in April when Anthropic said its new model is so good that it causes risks for detecting vulnerabilities beyond the capabilities and speed of existing models, and therefore couldn't be released immediately. To put numbers to it: Mythos Preview has already surfaced over 23,000 potential vulnerabilities across 1,000 open-source projects, with over 1,000 confirmed as critical or high-severity. Instead of a public launch, they made it available to different partners via Project Glasswing to give them a head start in fixing those vulnerabilities before it lands publicly and into the hands of people who want to misuse the model for attacks. According to Anthropic, that exclusive window is obviously going to close.

Sources:

The AI IPO Race: SpaceX, Anthropic, and OpenAI

Other important events were the public announcements of the labs Anthropic (Claude), OpenAI (GPT), and also SpaceX going public. It seems there is a race for who goes first and who gets the most.

SpaceX has to be seen as not purely a rocket company; following their recent merger with xAI, they are a massive AI player. For one, they have Grok. Second, SpaceX has the capability to provide the infrastructure—or better said, compute—for the labs. For example, Anthropic has reserved all capacity at SpaceX's Colossus 1—300+ MW and 220,000+ NVIDIA GPUs—for $1.25 billion per month through May 2029, via an Anthropic compute partnership announced last May. Last but not least, SpaceX also bought a $60 billion option to acquire Cursor. Cursor is not just an agent but comes with its own model lineup—Composer 2.5 scores 79.8% on SWE-bench Multilingual—and scales heavily via reinforcement learning.

SpaceX is set to list on June 12 at a ~$1.75 trillion valuation, raising ~$75 billion—the largest IPO proceeds ever, ahead of Saudi Aramco's $29.4 billion raise at a $1.7 trillion valuation in 2019.

After SpaceX, the two big labs will follow. OpenAI is targeting a September 2026 debut at $852 billion to ~$1 trillion. Anthropic followed on June 1, eyeing October 2026 at a $965 billion private valuation.

Sources:

NVIDIA RTX Spark: Local AI on Windows

In the hardware market, NVIDIA is taking a major step toward making large local LLMs viable on Windows machines. Because running massive local AI models is highly memory-intensive, we could mainly only use Apple's Mac hardware and its unified memory architecture (where the CPU and GPU share a single, large pool of high-speed RAM). On the PC side, unified memory was previously restricted to highly specialized, expensive developer boxes like NVIDIA's Linux-based DGX Spark.

Now, NVIDIA has announced the RTX Spark for Windows-on-Arm machines, co-developed with MediaTek. This brings that same unified memory (up to 128GB) advantage to standard laptops. To put it short, local LLMs will land to the broader consumer market.

Sources:

Got a question?

Just hit reply to this email and it'll land in our inbox.

— The Soverius team

Soverius AI · Austria

Legal notice (Impressum) · Privacy Policy

hello@soverius.ai

View in browser · Unsubscribe · Email preferences