Good morning. It’s Friday, August 14th.

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

OpenAI previews 14× faster GPT-5.6 Sol for real-time agents

OpenAI is previewing an invite-only Ultrafast Mode for GPT-5.6 Sol, an API tier backed by Cerebras wafer-scale hardware that cranks inference speeds up to 14x standard processing. By packing model weights into 44GB of on-chip SRAM, the setup bypasses the memory bottlenecks that slow down GPU clusters, topping out at 750 tokens per second.

In benchmark testing, Sol Ultrafast tore through 2,500 questions in Humanity's Last Exam in 11 hours and 11 minutes, far faster than Claude Fable 5, which took over 78 hours with similar accuracy.

The limited preview targets tighter feedback loops for time-sensitive tasks like live voice, real-time coding, financial research, and incident response, with early testing underway at Jane Street, Podium, Basis, and Rogo.

On desktop, OpenAI is rolling out Computer History for ChatGPT Work and Codex users on macOS. Building on its Chronicle experiment, the opt-in feature tracks interaction events like typing, clicks, and app switches rather than recording constant screenshots or audio. That local event log drops token overhead while letting ChatGPT remember what you were debugging or where you left off on a pull request. Users can pause tracking, exclude specific apps or sites, and delete their timeline at any time.

The product updates arrive during another shakeup in the front office. Chief Revenue Officer Denise Dresser is leaving after just nine months, right behind former COO Brad Lightcap, former AGI lead Fidji Simo, and former CMO Kate Rouch. Wiz president and COO Dali Rajic is stepping in as the new CRO to steady revenue operations.

Making matters trickier, OpenAI's own recent 69-page internal report found no statistically significant link between heavy ChatGPT use and higher revenue per employee.

Grok 4.6 ties GPT-5.6 Sol at half the price

SpaceXAI just dropped Grok 4.6, delivering top-tier performance at less than half the cost of its main rivals. The model notched an Intelligence Index score of 61 on Artificial Analysis, tying GPT-5.6 Sol Max while passing Kimi K3.

Trained on synthetic reasoning data and reinforcement learning, Grok 4.6 focuses heavily on agentic workflows, wrapping up long-horizon coding, research, and CAD jobs on GDPval-AA v2 in about 53 steps which is roughly half the 103 turns needed by Claude Opus 5.

Pricing is where SpaceXAI is really picking a fight. At $2 per million input tokens and $6 per million output, Grok 4.6 undercuts top rivals by more than 60%. But there is a catch for heavy users: once your prompt crosses 200,000 tokens in its 500k context window, those rates double to $4 and $12..

Supported by an official safety-tested model card, the launch includes double usage quotas for week one and distribution through xAI’s API, Cursor, OpenRouter, Vercel, Cloudflare, and Cognition's Devin platform. Musk highlighted that the model works best inside Grok Build and teased Grok 4.7, claiming a dataset built on SpaceX engineering data will push it past every model on the market.

At the same time, 𝕏 open-sourced the ranking and filtering code for its ‘For You’ timeline on GitHub to gather public critique. The codebase pairs with a pilot label transparency feature on a new Under the Hood dashboard. Eligible accounts in a randomized test group (at least a year old with ten-plus monthly posts) can view and download data on any visibility-limiting labels applied to their posts or profiles.

DeepSeek open-sources Claude Code rival that can rewrite its own code 

DeepSeek is launching DeepSeek-V4-Pro, pairing major agentic upgrades with a substantial price hike. The model jumped 11 points on the Vals Index to claim the second spot among open-weight models at $0.14 per task. Available on the web, mobile, and API via Expert Mode, V4-Pro offers adjustable reasoning effort, native OpenAI Responses API support, and one-click Codex integration.

To complement the model update, Alongside the model release, DeepSeek open-sourced DeepSeek Harness v0.1 under an MIT license as a direct alternative to Claude Code. Developed with Peking University researchers, the framework runs on a system called Cordis, which treats every piece of the setup, from tools and sandboxes to models and file systems, as interchangeable plugins.

Cordis lets self-updating AI agents modify their own code safely by automatically undoing broken changes, recalculating dependencies, and swapping modules on the fly without crashing the system.

To manage heavy server loads as it prepares for a potential IPO, DeepSeek is abandoning flat rates for dynamic peak and off-peak API pricing starting August 16 at 16:00 UTC, with off-peak usage discounted by 50%.

Input and output rates for V4-Pro jump from $0.435/$0.87 per million tokens to $0.66/$1.98 off-peak and $1.32/$3.96 during peak hours, while V4 Flash output moves to $0.66 off-peak and $1.32 at peak.

Model & Product Releases

Benchmarks & Evaluation

Security & Cyber

AI Research & Self-Improvement

Developer Tools & Agents

Policy & Regulation

Business & Markets

Data, Content & Publishing

Labor & Economy

Robotics

Nuphos is an AI-native DevOps workspace allowing teams to safely collaborate with autonomous infrastructure agents.

Skilldocs is a collaborative markdown editor where teams co-edit agent instructions in real time and sync feedback to AI agents.

Lettertrace is an open source AEO tool tracking brand visibility across Claude, ChatGPT, and Gemini using personal API keys.

Kane CLI is an agentic terminal tool running natural language browser tests that output cryptographic evidence packages.

Dograh is an open-source voice platform with visual workflows, local execution, and MCP-driven AI agent builds.

Thank you for reading today’s edition.

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.