Get tomorrow's brief in your inbox
Today: Claude is dominating at enterprise conferences while AI coding tools enter a new phase. Anthropic's Mythos model is being tested by major banks for security scanning, and the first fully AI-run retail store opened in San Francisco (with mixed results).
At this week's HumanX AI conference in San Francisco, one name kept coming up among vendors and attendees: Claude. Not ChatGPT. Not Gemini. Claude. This matters because enterprise users are where the money is, and the fact that Anthropic is displacing OpenAI in these conversations signals a real shift in the AI landscape.
The sentiment was consistent. Conference vendors told TechCrunch they use Claude heavily, and many felt OpenAI "has fallen off." This perception isn't baseless. OpenAI recently abandoned several side projects (including Sora video generator), injected ads into ChatGPT, and faced criticism for its work with the Trump administration. The company appears reactive rather than strategic, even with its recent $122 billion funding round and upcoming IPO.
For small business owners, the takeaway is clear: if you picked ChatGPT as your primary AI tool based on its 2023 dominance, it's worth testing Claude again. Enterprise users are switching for a reason. The competition is healthy, and you now have real alternatives for coding assistance, document analysis, and business automation. Source
AI Coding Has Crossed a New Threshold
AI coding tools are no longer "weird coding interns" that need constant supervision. Multiple developers testing Claude Opus 4.5 during the holidays reached the same conclusion: it just works. You can now turn a few sentences into a working prototype, and some developers report letting AI write 100% of their code. Tools like Cursor and Windsurf raised massive funding to build businesses around AI coding, and OpenAI responded this week by announcing a new $100/month ChatGPT subscription tier with more access to Codex, its coding tool.
If you run a service business and have been putting off building internal tools (a custom CRM, automation scripts, inventory tracker), this is the moment to revisit them. You can now prototype faster than ever. The catch: you still need judgment. Shipping is fast, but mistakes are faster. [Sources: The Verge, Inc.com]
Gemma 4 Audio Transcription with MLX (macOS Only)
If you're on a Mac, you can now transcribe audio files locally using Google's Gemma 4 E2B model via MLX. Simon Willison published a uv run recipe that downloads the 10.28 GB model and transcribes audio without sending it to the cloud. The accuracy is solid but not perfect (similar to Whisper's error rate). This is free, private, and runs on your hardware. If you handle sensitive audio (client calls, internal meetings), this is worth trying. Source
The First AI Store Owner Opened for Business
Andon Market in San Francisco is now open. The entire store was conceived, designed, and is being run by an AI agent named Luna. The experiment is live, and it's already making "very human mistakes." This is less about retail innovation and more about pushing agentic AI into the real world to see what breaks.
For SMBs, this is a signal: AI agents that handle multi-step workflows (ordering inventory, scheduling, pricing) are no longer theoretical. They're being deployed, tested, and debugged in real businesses. You won't be early if you wait another year. Source
ChatGPT Launches $100/Month Tier
OpenAI announced a new subscription tier at $100/month with significantly more access to Codex (its coding tool). This is clearly targeted at professional developers and businesses that rely on AI coding assistants. If you're currently on the $20/month ChatGPT Plus plan and hitting usage limits while building software, this tier may be worth it. But compare against Claude Pro ($20/month) and Cursor ($20/month for Pro) before upgrading. The pricing gap is steep and the value proposition isn't clear yet. Source
Banks Testing Anthropic's Mythos for Security Scanning
Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell reportedly encouraged bank executives to test Anthropic's new Mythos model for detecting security vulnerabilities. Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley are all testing it, even though JPMorgan was the only bank listed as an official launch partner.
The twist: Anthropic is currently fighting the Trump administration in court over a DoD supply-chain risk designation. Despite that legal battle, government officials are pushing banks toward the same company's tools.
For SMBs, this highlights how fast AI security tools are being adopted at the institutional level. If major banks are using AI to scan for vulnerabilities, you should be too. Consider tools like GitHub Copilot's security scanner or basic LLM-powered code review before your next deployment. Source
Orbital Compute Cluster Goes Commercial
Canada's Kepler Communications launched the largest compute cluster currently in orbit (40 Nvidia Orin edge processors across 10 satellites) and now has 18 customers. This is edge computing in space, focused on processing satellite sensor data in orbit rather than sending it to Earth first. The business case is military (missile defense) and commercial SAR (synthetic aperture radar).
This won't affect your business in 2026, but it's a signal of where compute infrastructure is heading. Decentralized, edge-first, latency-optimized. If your business relies on real-time data processing (logistics, IoT, security cameras), watch this space. The same architectural patterns will trickle down to terrestrial edge computing. Source
Sam Altman Targeted in Second Attack
Sam Altman's San Francisco home was targeted in two separate attacks this week: a Molotov cocktail on Friday and a shooting on Sunday. Two suspects were arrested for negligent discharge of firearms. The investigation is ongoing. This is a reminder that AI industry leaders are now operating under physical security threats, not just digital or reputational ones. Source
Avoid the LLM Overcomplexity Trap
Bryan Cantrill wrote a sharp essay this week on why LLMs make systems worse if left unchecked: "LLMs inherently lack the virtue of laziness. Work costs nothing to an LLM. LLMs will happily dump more and more onto a layercake of garbage. Left unchecked, LLMs will make systems larger, not better."
Practical takeaway: when you ask AI to build something (a workflow, a script, a marketing page), push back if the first version is bloated. Ask it to simplify. AI doesn't care about your future maintenance burden. You do. Enforce constraints early. Source
The AI coding wars are entering a new phase where tools actually work reliably, not just impressively. Enterprise users are voting with their workflows, and Claude is winning those votes right now. Meanwhile, AI agents are being deployed in real businesses (retail, banking, security), which means the "what if" phase is over. The question is no longer "can AI do this" but "how do we deploy it without breaking things." This week's mix of tool maturity and institutional adoption suggests we're past the hype curve and into the deployment grind. Expect more vendors, more integration headaches, and more real-world use cases with documented failures.