Good morning. It’s Monday, August 17th.

You read. We listen. Let us know what you think by replying to this email.

Anthropic is juggling a lot right now. Over on the dev side, they dropped a guide to taming Claude Code token costs. Since agentic coding drains context fast, keeping sessions clean with /clear, /context, and direct file tags makes a huge difference. Prompt caching cuts repeat input costs down to 10 percent, but swapping model modes or effort levels blows away the cache entirely. To help users benchmark outputs, Anthropic is also testing a side-by-side model comparison interface right inside Claude chats.

Anthropic is internally developing a model that outperforms Mythos 5, with no current plans for public release. Separately, a safety issue led to the exposure of 133 million contractor requests over nearly a year after biological and chemical weapon classifiers were accidentally left off between May 2025 and April 2026.

Meanwhile, Anthropic's recently launched global text watermarking system and upcoming detection API tweak token probabilities using secret-keyed green and red word lists similar to Google's SynthID Text. Anthropic swears writing quality remains untouched, but critics warn Anthropic's new global text watermarking alters token probabilities, potentially degrading Claude's writing quality despite company denials. Proving that point, an open-source tool built to strip those exact watermarks hit 10,000 GitHub stars within days.

Facing pushback, CEO Dario Amodei framed the backlash around a broader ‘crisis of trust,’ insisting heavy compute requirements naturally centralize AI power and that true credibility will come from curing diseases rather than slick marketing campaigns.

Silicon Valley's biggest voices are fighting over who gets to control the future of AI. On one side, Replit CEO Amjad Masad and Elon Musk point to an 18x jump in intelligence output per joule over 16 months. They argue rapid hardware and algorithm gains will bring AGI to standard devices, shattering the idea that AI must remain tied to massive data centers.

Anthropic CEO Dario Amodei disagrees, arguing scaling laws inherently concentrate power among deep-pocketed labs. Amodei defends his call for strict pre-deployment testing, claiming structured rules protect open models and smaller rivals rather than entrenching incumbents.

David Sacks isn't buying it. He slammed Amodei's proposed federal approval agency as a "DMV for AI" designed to lock in regulatory capture. Sacks argues Anthropic is leveraging former policy officials to push rules that bottleneck open-source competitors, handing gatekeeping power to a bureaucracy tied to frontier labs.

As Sacks frames it, the core debate comes down to a fundamental divide: Amodei believes frontier AI is too dangerous to distribute, while his critics believe it is far too powerful to centralize.

OpenAI is now making way more money from corporate clients than everyday users.. CFO Sarah Friar confirmed enterprise sales officially passed consumer revenue, pushing annualized revenue to $40 billion after a 20 percent spike in July. Business accounts jumped 32 percent to hit two million users as corporate buyers focus on real work per dollar over raw token counts.

To give its AI agents better context, OpenAI is testing an opt-in feature called Computer History in its macOS app. It logs clicks, keystrokes, and active apps so ChatGPT can recall what you worked on earlier without taking screenshots.

Putting those browser tools to work, Ethan Mollick used GPT-5.6 Sol in Codex to take over Chrome and recover 5,302 old 𝕏 bookmarks, pulling down saved threads on prehistoric handaxes and biological systems.

Developers working on huge codebases can also unlock a full one-million-token context window in Codex using GPT-5.6-Sol. Engineer Tibo shared that adding a 1,000,000 token limit to ~/.codex/config.toml with auto-compaction at 900,000 tokens lets the model hold massive amounts of history.

Chert builds conversational iMessage and FaceTime AI agents that handle customer interactions and integrate with CRMs.

Outcome uses AI to turn creator content into interactive, personalized sales funnels that deliver customized action plans.

Blume is an open-source framework that builds fast, AI-ready documentation sites from plain Markdown files.

Nuphos is an AI-native DevOps workspace that helps engineering teams investigate, manage, and automate cloud infrastructure safely with AI agents.

DeepSeek Harness is an open-source agent runtime that uses a modular plugin architecture to turn language models into task-executing agents.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.