Good morning. It's Friday, July 10, and we're covering AI's push onto smartphones, questions over who approves frontier AI models, and why OpenAI is betting on efficiency over raw intelligence.

Plus: yesterday's poll results on AI for planned economies, and our latest Zeitgeist discussion on AI's shift toward price-performance.

OpenAI made its GPT-5.6 family (Sol flagship, Terra mid-tier, Luna cost-optimized) generally available Thursday across ChatGPT, Codex, and the API, positioning Sol as state-of-the-art on Agents' Last Exam (53.6, +13.1 vs Claude Fable 5) and Artificial Analysis Coding Agent Index (80, +2.8 vs Fable 5) while claiming ~half the cost, ~half the output tokens, and ~half the latency.

The New York Times alleges OpenAI secretly tracked copyrighted content despite claiming such searches were impractical. New court filings accuse OpenAI of withholding evidence in the publishers' copyright lawsuit by concealing internal search tools and databases. Plaintiffs cite a 78 million-chat dataset, Project Giraffe, and testimony from privacy engineer Vinnie Monaco.

Meta upgrades its flagship reasoning model by focusing on autonomous task execution rather than benchmarks alone. Muse Spark 1.1 improves coding, computer use, and multimodal reasoning while introducing a 1 million-token context window and multi-agent orchestration. The model is available in the Meta AI app, on Meta.ai, and through the new Meta Model API public preview.

In 2013, one of the fastest machines on Earth tried to keep pace with a human brain for a single second, and the result says a lot about the hardware sitting inside your skull. Get the full picture at the bottom of today's issue.

Higgsfield now plugs straight into Claude via MCP. Ask for a cinematic ad, a talking avatar, a product video, and it renders inside the chat while you keep working. No tabs, no timeline, no learning curve. The whole studio, one prompt away.

Connect Higgsfield to your Claude for free — takes about a minute

GPT 5.6 squeezes more from the same model, building complex software, excelling at computer use, and cutting inference costs.

GPT-5.6 is here after a two-week government delay, and Matt has been testing it internally for two months. The full review covers what nobody else has yet: a Minecraft-style game the model kept building for days after he walked away, an Excel clone it refined over six days using Computer Use, and browser control fast enough to manage live Supabase instances unprompted.

Plus the head-to-head everyone wants: GPT-5.6 vs. Claude Fable, from benchmarks to cost per finished task. Every demo in the review is live and playable.

Read the full review →

PrismML says it has compressed Alibaba’s 27-billion-parameter Qwen 3.6 open-source AI model to run entirely on an iPhone 17 Pro, making it the largest AI model yet demonstrated on a smartphone. The startup claims its compression technique reduces the model from about 54GB to under 4GB without degrading performance, enabling advanced capabilities such as reasoning, coding, autonomous agents, and complex chat directly on-device.

The breakthrough aligns with Apple's push to move more AI processing onto iPhones to improve privacy and reduce cloud infrastructure costs, and Apple has reportedly held discussions with PrismML about its technology. While the claims have not been independently verified, the company plans to release the compressed model next week and says it intends to apply the same approach to even larger frontier-scale AI models. → Read the full article here. (Paywall)

OpenAI has begun rolling out its frontier AI model, Sol, but the process the U.S. government used to determine it was safe for public release remains largely opaque. Researchers, policymakers, and industry leaders told TechCrunch they lack visibility into the evaluation criteria, who conducted the assessments, or which standards models must meet before deployment. OpenAI says Sol underwent external safety testing by organizations including the U.K.

AI Security Institute (AISI), SecureBio, and Irregular, while CEO Sam Altman confirmed discussions with senior U.S. officials, though neither the company nor the government has disclosed details of the approval process. The uncertainty has fueled concerns that frontier AI oversight relies on informal government-industry relationships rather than a transparent, standardized framework, raising questions about accountability as increasingly powerful models reach the public. → Read the full article here.

Every morning the Forward Future team argues about the stories, papers, and launches that matter. What's real, what's hype, and where AI is headed next. These are the public notes.

Are frontier AI models still the ones that matter, or are cheaper, "good enough" models quietly taking over? The debate ranged from benchmark obsession and open-source adoption to the growing gap between pre-release demos and public launches.

▸ Why price-performance is beating raw intelligence
▸ Could open source dominate AI inference by year-end?
▸ The growing credibility gap between early demos and public releases

Read yesterday's Zeitgeist →

OpenAI Kills Atlas: OpenAI is shutting down Atlas less than a year after launch and redistributing its agentic features into a ChatGPT Chrome extension.

AI IPOs Eclipse Tech History: SpaceX, Anthropic and OpenAI could surpass the combined value of all U.S. VC-backed exits since 2000.

Meta’s In-House AI Chip Strategy: Meta moves its custom AI chip effort into production as it expands infrastructure beyond third-party hardware.

Palo Alto Networks CEO Targets AI Costs: CEO Nikesh Arora says AI token costs must fall 90% within two years to drive broader enterprise adoption.

Character.AI Debuts Interactive Microdramas: New AI microdramas let adult viewers chat and roleplay with characters, expanding its entertainment push.

China Flags Claude Code Flaws (Paywall): China says Anthropic's Claude Code exposed user data through built-in monitoring, urging updates or removal.

In 2013, researchers used Japan's K computer — then among the fastest in the world — to model the activity of a neural network for one second. The machine ran 82,944 processors and about a petabyte of memory to simulate 1.73 billion neurons connected by 10.4 trillion synapses. It took 40 minutes of computing to reproduce a single second. And that network represented roughly one percent of the human brain.

Your brain does the full version continuously, on about 20 watts — less power than the bulb in your refrigerator. The 2013 run wasn't trying to explain how the mind works; the neurons were wired at random. It was a stress test of the software and the hardware. But the gap it exposed is the point. The most efficient computer we know of is the one that reads this sentence, and we still can't simulate a hundredth of it in real time.

Here’s how you voted: Nearly half said AI doesn't fix the core problems of planned economies, while 32% warned it could create a bigger monopoly.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team