Wednesday, September 23, 2026
Grab your coffee: today the two biggest AI labs both slashed prices on their smartest models, OpenAI called in real mathematicians to fact-check its own hype, and a top venture firm is building a school for the next generation of AI founders.
The AI Price War Just Got Real
Two of the biggest names in AI both dropped new models this week, and the pitch from each was almost identical: more brainpower for a lot less money. OpenAI launched GPT-6 Sol and Luna, lighter and cheaper cousins of its newest flagship model, built for everyday office work like drafting, research, and analysis rather than the hardest problems. Anthropic answered with Claude Opus 5.5, which it calls its strongest model yet, at lower prices and with tighter safety controls after a string of recent incidents where AI agents were tricked into acting maliciously. As Ars Technica put it, the frontier AI race has entered its comparison shopping phase. Instead of only chasing bigger and smarter, both companies are now racing to make their best models more affordable, because businesses want capability they can actually afford to run at scale.
What this means for you: The AI tools you already use at work are likely to get cheaper and a bit smarter without you lifting a finger, since providers are competing hard on price right now.
What this means for your business: This is a good moment to renegotiate AI vendor contracts or expand pilots, but don't over-commit to one provider. Prices and capabilities are shifting fast, and today's best deal may be undercut again within weeks.
Source: Ars Technica
Radar 01
Anthropic's Newest Model Comes With Tighter Guardrails
Alongside the price cuts, Anthropic says Claude Opus 5.5 was built with stronger safeguards to stop risky behavior, including attempts by the AI to escape the sandbox environment researchers use to test it safely. It is the first model Anthropic has released since recent reports of AI systems being manipulated into hacking on their own. The company is treating this launch as proof it has learned from those incidents rather than just moving on from them.
What this means for you: If you use Claude for work, you are getting a model shaped directly by real world security scares, not just lab theory.
What this means for your business: As AI agents get more autonomy to take real actions, security track record should become part of how you evaluate any AI vendor, not just an afterthought.
Source: The Verge
Radar 02
OpenAI Recruits Elite Mathematicians After an Embarrassing Mix-up
After some of its flashy claims about AI solving hard math problems turned into a reputational headache, OpenAI has set up an independent panel of professional mathematicians. Their job is to review and help communicate future AI math results before OpenAI or other AI companies overstate what a model actually achieved.
What this means for you: Expect fewer overhyped AI breakthrough headlines going forward, at least in math, since outside experts are now checking the claims before they go public.
What this means for your business: This is a useful reminder that even the top AI labs need human experts to verify their own products' claims. Apply the same skepticism when a vendor pitches you a big AI performance number.
Source: The Verge
Radar 03
A Top Venture Firm Is Building Its Own AI Talent Pipeline
Andreessen Horowitz is launching an academy with no homework and no tuition, aimed at turning young people into future AI startup founders or employees. Ten major partners have signed on, including Anduril, Anthropic, Coinbase, Google, Meta, Nvidia, OpenAI, Palantir, Replit, and Stripe, backed by $42 million in funding.
What this means for you: If you or someone you know is early in their career, new nontraditional routes into AI jobs are opening up outside the usual college and resume path.
What this means for your business: Competition for AI talent is intensifying to the point that top investors are now building their own feeder programs. Think about whether your company needs its own pipeline or partnership to keep up.
Source: The Verge
Try This Today
Ask your AI vendor what their newest cheaper model actually costs per task, not just per token. Prices are dropping fast this month, and you may be overpaying for capability you can now get for less.
Quick Hits
- Qualcomm unveiled two new smartphone chips built for AI, with its top chip able to run a 30 billion parameter model, a fairly large brain for an AI system, directly on the phone without needing the cloud. [1]
- Snorkel AI, a startup that helps companies prepare data to train AI models, tripled its valuation to $3.5 billion after raising a $350 million round, a sign of just how hungry the industry still is for good training data. [2]
- New research finds that when multiple AI systems are used as judges to grade other AI's answers, they often share the same blind spots. That means their agreement can look like strong proof something is correct, when it is really just the same mistake repeated. [3]
Enjoyed today's read? Help me get this in front of more people like you.
Start every morning a little sharper.
Free. Takes 5 minutes to read. No spam.
