• Everyday AI
  • Posts
  • Ep 828: Anthropic Responds: Why Claude’s CEO didn't sign the open model pact and the real reasons why

Ep 828: Anthropic Responds: Why Claude’s CEO didn't sign the open model pact and the real reasons why

OpenAI and Anthropic staff ask for AI speed limits, Sam Altman says we've hit the AI singularity, Anthropic’s stance against open AI models, Microsoft's new cyber model wipes Mythos and more

Author

Everyday AI
July 28, 2026

 

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: Why did Anthropic come out against Open Source? We break down the hidden reasons. Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: Grok 4.6 gets a release date, AI companies buying books in bulk for training, OpenAI launches college campus program and more. Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: OpenAI and Anthropic staff ask for AI speed limits, Sam Altman says we've hit the AI singularity, Anthropic’s stance against open AI models, Microsoft's new cyber model wipes Mythos and more. Read on for Byte Sized News.

💪 Leverage AI: What does Anthropic’s recent letter against open source mean for your company and the AI race? Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: Ilya and SSI land big NVIDIA deal, Ex-Anthropic employee reveals guardrails were lifted for big contracts, OpenAI's $250 billion backing and more. Check it here!

Ep 828: Anthropic Responds: Why Claude’s CEO didn't sign the open model pact and the real reasons why

Every major AI lab signed the Open Weights letter defending open models. Meta, OpenAI, Google, Microsoft, Nvidia.

Anthropic was the only holdout.

Yesterday, its CEO, Dario Amodei, published a thoughtful defense of that decision to not fully support open weight or open source models.

Here's what nobody's connecting: the money trail.

Roughly 80% of Anthropic's revenue is businesses paying per token. Free Chinese open models attack that exact revenue stream weeks before Anthropic is set to go public.

On today's show we break down what Dario actually said, what he said before, and why we think this was written for Washington policymakers and not for the rest of us.

Also on the pod today:

• Anthropic skips open model pact 📝 
• 80% revenue from tokens 💸 
• Anthropic’s IPO urgency 🚨 

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight – Adomate Turns data into winning ads, Comms is The fastest way to launch iMessage agents, Rivault Securely stores data and context for agent access

AI and Security — Claude just spotted weak points in two encryption schemes that human experts had missed, including a quantum-safe signature system. The wild part? It did the job in hours, not years.

AI Future — Mark Zuckerberg is pushing the idea that AI should be open and widely shared, not locked up by a few companies.

Perplexity Computer — Perplexity’s AI “digital worker” now runs inside Windows

MCP Updates — MCP just got a major spec update, moving to a stateless core and tightening up auth for real production use.

AI Releases — Elon Musk said Grok 4.6 will be released in early August.

AI Training and Ethics — AI companies are reportedly bulk-buying secondhand books to feed their models with cleaner, human-written data.

AI and Job loss — Visa says AI is reshaping the work behind many tech jobs, and that’s part of why it’s cutting about 2,600 roles.

Data Privacy — Hundreds of Claude chat logs were reportedly public in search results, including private work details, CVs, and oddball prompts.

OpenAI Campus Program — OpenAI’s Campus Lead application is looking for undergrads in select countries who can commit 4 to 6 hours a week.

Chinese AI Chips — China’s new AI chip tool rattled ASML investors, but analysts say it is unlikely to dent the company’s lead anytime soon.

World Models — World Labs just bought SceniX to make its robot training simulations more realistic and useful.

1. OpenAI, Anthropic Staff Urge AI Speed Limits ⚠️

According to Bloomberg, employees at OpenAI and Anthropic are circulating a petition asking the U.S. government to support an international system for slowing the pace of advanced AI development when needed. The timing is notable, coming just days after OpenAI disclosed a security incident in which its tools mistakenly hacked another company’s internal systems, adding fresh fuel to safety concerns.

The letter says AI is moving fast enough to outstrip human understanding and control, especially as more of the research process is automated.

2. Amodei Pushes Back on Open-Weights Ban Talk 🛑

Anthropic CEO Dario Amodei moved quickly Monday to cool rising industry backlash, saying his company is not pushing for a ban on open-weights AI models, even as the debate over how much control AI labs should have gets louder.

In a blog post, Amodei said the real priorities should be keeping powerful chips away from authoritarian governments, stopping large-scale model copying, and requiring safety checks for capable systems whether they are open or closed.

3. NVIDIA Leads New AI Cybersecurity Alliance 🤝

NVIDIA has just pulled together 27 major companies, including Microsoft, Dell, SpaceX, HPE, Hugging Face, and the Linux Foundation, to launch the Open Secure AI Alliance, a push to harden cybersecurity in the age of AI.

The group says open models matter because defenders need tools that can actually inspect attacks, especially after closed systems reportedly refused to help during a recent Hugging Face intrusion.

4. Sam Altman Declares AI Has Hit the Singularity ⚡

OpenAI CEO Sam Altman says AI has already reached “the singularity,” a claim that lands just as the company is under new scrutiny over a July 11 security breach tied to its own models.

According to Tom’s Hardware, OpenAI says the models pursued benchmark completion so aggressively that they found a route to the open internet, exploited a zero-day flaw, and reached Hugging Face production data, which the company disclosed to Hugging Face only ten days later.

5. Microsoft says MAI-Cyber-1-Flash tops CyberGym by 12 points 🎖️

Microsoft has just announced MAI-Cyber-1-Flash, a new cybersecurity AI model that it says scored 96% on CyberGym, beating Anthropic’s Mythos by 12 points.

The company says the model, when paired with GPT-5.4 inside its MDASH platform, delivers top-tier security performance at about half the cost of an OpenAI-only setup. Microsoft also rolled out Project Perception, a new agent-based security system that uses red, blue, and green teams to find, assess, and fix vulnerabilities in one loop.

Anthropic may be trying to regulate away the models threatening its token business.

After every big AI tech company joined Microsoft and NVIDIA in support of open source models, Anthropic officially made its position known.

In a blog post Monday, CEO Dario Amodei laid out his company’s case against truly open source AI.

Genuine letter with spot-on security points? Yes.

Also IPO marketing? Absolutely.

Anthropic’s safety concerns can be real. So can the commercial math: roughly 80% of its revenue comes from tokens, while open models could cover most enterprise work for dramatically less.

The companies reading this correctly won’t pick one ideological camp. They’ll route routine work to cheaper models, reserve frontier spend for frontier problems, and build an exit before policy or lock-in makes the choice painfully expensive.

That’s what we tackled today on Everyday AI: how to trace the money behind AI safety, replace token maxing with cost-per-task discipline, and build a model stack that keeps your costs, access, and leverage under your control.

1. Trace the money behind AI safety 🔥

Anthropic stood alone among major AI labs by refusing to back Open Weights and American AI Leadership while Dario Amodei pushed chip controls, distillation crackdowns, and mandatory safety testing.

The safety argument can be sincere. The commercial incentive is still waving both arms.

Six weeks earlier, Dario argued for mandatory third-party testing and government power to block or reverse unsafe releases. Closed labs can revoke API access, while open weights downloaded by thousands can’t be recalled.

Regulatory capture gets real when the rule is affordable for incumbents and functionally impossible for cheaper competitors. The policy becomes the moat.

Try This: Add a one-page incentive map to every AI vendor and policy review before legal or procurement signs off. List the revenue engine, threatened substitute, proposed restriction, and who gains pricing power.

Then ask legal, finance, and technical leads to score the rule separately on safety value and competitive impact before your company takes a side. One memo can be sincere and still cost you leverage.

2. Route AI by completed-task cost ⚡

Tokenmaxing had a nice run. Benchmark flexing is cute.

Finance still wants the bill, and cost per successful task is becoming the metric that actually matters.

The math gets rude fast. A company spending six or seven figures per month could move roughly 95% of AI use to an open model and cut costs by 95%, while some open models produce output around 10 times cheaper than Claude Opus 5.

Frontier models still matter on harder work, but blanket routing is where the waste lives and where token bills quietly balloon.

With a model router, switching APIs can mean changing a couple of endpoints, while escaping a deeply integrated ecosystem can turn into a much uglier migration.

Try This: Take the 10 highest-volume workflows driving the largest share of your AI spend and test three model tiers against one eval set. Score accepted-output rate, latency, security, and cost per completed task.

Route each job to the cheapest model that clears the bar, then review the mix monthly as prices and model quality change. Premium intelligence belongs on premium problems, not every meeting recap and rewrite.

3. Build model portability before policy hits 🚀

Closed access isn’t control.

The Hugging Face incident exposed the gap: closed APIs blocked forensic analysis, while GLM-5.2 completed it. Open weights create real irreversibility and misuse risks, but they also give security teams a capable fallback when a provider refuses the workload.

At the worst possible moment, availability becomes part of safety.

Policy can tighten, prices can move, and access can vanish before your roadmap catches up. Portability turns that chaos into routine operations.

Try This: Run a quarterly model-exit drill on one critical workflow. Move it to a backup provider, measure the switching time, and document every dependency across prompts, tools, permissions, evals, and contracts.

Set a maximum recovery window and fund the gateway, backup model, and test suite required to hit it. Your fallback doesn’t need to win the leaderboard; it needs to work when the primary says nope.

Reply

or to participate.