QUANTUM BEAT 11-06-26
Your daily AI briefing. Four stories, no fluff, straight talk.
ANTHROPIC OPENS THE VAULT: THE MODEL THEY WERE AFRAID TO RELEASE IS NOW IN YOUR HANDS
So here is the thing about Anthropic. For months they have been quietly running what they call Mythos, their absolute top-shelf AI model, through a controlled list of partners. Government labs. Critical infrastructure operators. A very short and very serious list of people who had signed enough paperwork to get a clearance at a mid-tier spy agency. The rest of us just had to read about it and feel left out.
That just changed. Anthropic dropped Claude Fable 5, which is essentially the public-facing version of Mythos, and they want you to know it is not the full thing but it is close enough to make them nervous. The guardrails are real. Ask it anything touching cybersecurity, biology, chemistry, or what they call distillation, and it falls back to the older Opus model like a bouncer pointing you to the exit. They stress-tested it with more than 1,000 hours of jailbreak attempts and nothing cracked it universally. That is either very reassuring or a challenge, depending on the kind of person you are.
What makes this release interesting is the timing. Anthropic spent the week before publishing a plea to the global AI community to pump the brakes, warning that AI systems might soon improve themselves without human approval. And then, with the same week not quite over, they turned around and released their most powerful model to anyone with a subscription. You have to admire the confidence, or the cognitive dissonance, depending on your mood.
There is a catch with the pricing that every enterprise buyer is going to hate. Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is double what Opus costs. And Anthropic is now requiring a 30-day retention policy on all traffic, even for companies that previously had zero-retention agreements. They say it is for safety monitoring only, not for training. Whether or not that is fully reassuring depends on how much you trust a company that just filed for a near-trillion dollar IPO to stick to its word on data privacy.
The early reviews from enterprise partners are good. Analytics company Hex says it got a 90 percent score on its core analytics benchmark. Vibe-coding platform Base44 says it is exceptional at one-shotting full apps. Genspark says it beat every other model in their evaluations. That is a strong opening week.
The thing worth watching here is what happens on June 23. That is when Anthropic pulls Fable 5 from flat subscription plans and starts requiring usage credits. They say they will put it back eventually as a standard feature, but the window is closing fast if you want to get a free look at what the most capable public AI model in the world can actually do. Go poke it before they start charging you per token for the privilege.
GOOGLE JUST MADE AI CHEAPER THAN A STREAMING SERVICE AND THE WHOLE INDUSTRY IS SWEATING
Google just did something that sounds boring until you think about what it actually means. They cut the price of their entry-level AI subscription, Google AI Plus, from $7.99 a month to $4.99, and doubled the storage that comes with it at the same time. That is a 37 percent price cut with a better product. On the same day. With no warning.
You can read the full breakdown here. The short version is that Google just made a capable AI subscription cheaper than Netflix Basic. The $4.99 tier includes video generation through Omni Flash, the Google Flow creative studio, and NotebookLM, which is genuinely one of the most useful AI research tools built in the last two years. That is a lot of product for five bucks.
Now here is why this matters beyond just being a good deal for consumers. The AI subscription market has been operating at premium prices partly because nobody had to compete on price yet. OpenAI charges $20 for ChatGPT Plus. Anthropic charges $20 for Claude Pro. Nobody was really undercutting anyone in the US market. Google had already been running sub-$5 plans in India for months, but that was a developing market play, the kind of move big tech companies make when they want growth numbers without hurting domestic margins.
Bringing it to the US changes the math for everyone. One investor interviewed for the story compared this to what happened to internet infrastructure companies during the web era. Cisco, Lucent, Akamai. These companies were enormously valuable during the buildout phase. Then the infrastructure became a commodity and the value collapsed. The argument is that the same thing is now starting to happen to AI model providers. Google has the distribution, the bundling power, and the vertical integration to price at $4.99 all day long. OpenAI and Anthropic do not have those same structural advantages.
The part that should make Anthropic executives nervous is not just the price cut itself. It is that Anthropic has not responded. No budget tier in India. No localized pricing anywhere. Just the flat $20 Pro plan. They are about to go public and the company they want to be worth $1 trillion is now competing with a $4.99 option that has a decent feature set and Google’s search integration baked in.
The next logical move is obvious. OpenAI will respond, probably with a budget tier of their own. That puts pressure on Anthropic to follow. And suddenly the entire model-as-a-service pricing structure starts looking a lot more like cell phone plans than software. Which is fine for users and brutal for margins. Watch for OpenAI to make a move within the next few weeks. This race just got started.
HACKERS BROKE INTO MICROSOFT’S GITHUB AGAIN AND THIS TIME THEY WERE HUNTING AI DEVELOPERS SPECIFICALLY
This one is embarrassing for Microsoft. And not in the small way. We are talking about a company that owns GitHub, the world’s largest code hosting platform, getting its own repositories on that same platform compromised. Twice. In the same month.
The story, first broken by 404 Media and confirmed by TechCrunch, is that hackers managed to break into more than 70 of Microsoft’s open source GitHub repositories and inject malware into the code. The specific targets were tools used by AI developers, including integrations for Claude Code, Gemini’s command line interface, and VS Code, which is the most popular code editor in the world. The malware was designed to steal passwords and credentials from developers when they opened the compromised tools in their AI coding environments.
The technical term for this kind of attack is a supply chain compromise. Instead of going after one target directly, you poison the code that thousands of developers are already trusting and downloading. It is like contaminating the city water supply instead of knocking on one door. The beauty of it, from a hacker’s perspective, is that the developer thinks they are running safe, vetted code from a company they trust. Microsoft, in this case. The owner of GitHub itself.
What makes this especially noteworthy is the targeting. Hackers specifically went after tools used by AI developers. People who work with AI coding tools tend to have access to cloud infrastructure, corporate systems, production environments, and in many cases the AI model API keys that companies are paying hundreds of thousands of dollars a month to use. If you can steal one AI developer’s credentials at a large enterprise, you potentially have access to a lot of valuable and sensitive infrastructure.
The bigger problem here is that this appears to be a re-compromise. In mid-May, Microsoft’s open source Durable Task project was breached and cleaned up, or so they thought. Security researchers say the June incident may be the same attackers getting back in through a gap that was not fully closed. That is not a great look for a company that just launched a major AI security product and has been loudly marketing its AI developer tools for the past two years.
Microsoft pulled the repos and went quiet, which is standard incident response procedure. If you are a developer using Microsoft’s open source tools in your AI workflow, now would be a good time to audit your environment, rotate your credentials, and check whether any of the affected packages were in your dependency tree. The security firm Cloudsmith published a detailed analysis if you want to understand exactly what was compromised. This one is ongoing.
DATADOG VETERANS JUST RAISED $7 MILLION TO BET THAT OPENAI AND ANTHROPIC ARE GOING TO STOMP ON YOUR COMPANY
There is a bet being made in Silicon Valley right now that not enough people are talking about openly. The bet is that the big AI labs, OpenAI, Anthropic, Google, are moving fast into every software vertical they can touch. Legal. Healthcare. Finance. Developer tools. And companies that built on top of those labs are about to find out what it feels like when your main supplier becomes your main competitor.
Two former Datadog engineers, Sajid Mehmood and Conor Branagan, just raised $7 million to build a startup called Niteshift around exactly this premise. The lead investor is Greylock’s Jerry Chen. The angels include Reid Hoffman, Datadog’s own founders Olivier Pomel and Alexis Le-Quoc, and a handful of other credible operators. These are not people who throw money at a cute slide deck.
The pitch is straightforward. Right now, if you want to build a company using AI coding agents, you essentially have to run your code through OpenAI’s Codex or Anthropic’s Claude. Your most sensitive assets, the code that runs your products, flows through systems built by companies that are simultaneously building competing products in your market. Mehmood calls it the retail apocalypse equivalent for software. Amazon was building the cloud infrastructure that e-commerce companies ran on while also putting those same e-commerce companies out of business. A lot of Datadog’s early customers were retail businesses that refused to build on AWS for exactly this reason.
Niteshift’s solution is to separate the coding agent from the infrastructure underneath it. Instead of selling you a model, they sell you a cloud platform for running coding agents that can switch between Claude Code, Codex, open source options, or anything else depending on what your project needs. You are not locked into any one model provider. And crucially, you are not building on infrastructure owned by a company with a competing agenda in your market.
The pricing model is also different from everything else out there. Niteshift charges like a cloud provider, per minute of usage, not per token. Which means as models get cheaper and more efficient, you benefit from that without renegotiating a contract. Most AI coding tools are still in the labor replacement business, selling you intelligence that replaces developer time. Niteshift is selling software infrastructure for the agents themselves.
The honest caveat here is that this is a $7 million seed round in a market full of well-funded competition. Cursor has hundreds of millions and was reportedly looking at a $60 billion buyout offer from SpaceX. Cognition, the company behind Devin, just raised $1 billion at a $26 billion valuation. Amazon Bedrock has essentially infinite distribution. Niteshift’s founding team has the credibility from Datadog to get the conversation started. Whether the timing is right for the market to reward infrastructure independence over raw capability is the real question. The idea is sound. The execution is what we will be watching.