Presented by
Good morning, AI enthusiasts. More than 100 cybersecurity executives just signed an open letter demanding the Trump administration lift Fable 5's export ban, citing an irony that cuts deep: the model's guardrails were so restrictive they became a running joke among red teamers on launch day.
The same rules designed to make Fable 5 safe enough to deploy are now the primary evidence that the ban may not be justified. Has the government's most dramatic AI intervention been built on a misread of its own research?
In today's recap:
Cybersec community rallies to reverse the Fable 5 ban
Nadella's "token capital" framework for winning AI
Build expressive AI text output with Ink-2
OpenAI burned $38.5B in 2025, up nearly 8x
4 new AI tools, prompts, and more
CYBERSECURITY
100+ cybersec leaders challenge the Fable 5 ban
Screenshot
Recaply: More than 100 cybersecurity executives just signed an open letter urging the US to lift Fable 5's export ban, arguing the model isn't uniquely dangerous compared to GPT-5.5, Claude Sonnet, or Kimi 2.7.
Key details:
The letter argues Fable 5 is good at offensive cyber tasks. But so are GPT-5.5, Claude Opus, and Kimi 2.7. None of those models face any export restriction.
The letter has over 100 signatories, from CISOs to university professors. Fable 5's guardrails were so strict they became "a source of humor" among red teamers. The letter also cites Anthropic's 1,000+ hours of red team testing.
Katie Moussouris helped write multilateral export controls. She told CyberScoop the research behind the ban used a "multistep and manual process." She called the restriction "heavy handed."
Controls took effect June 13, pulling Fable 5 from users globally. The EU Commission says it's in contact with Anthropic. Senator Mark Warner wants a statutory framework with clear risk assessments.
Why it matters: The government's argument was that Fable 5 provides unique offensive uplift. The cybersecurity community's response is different. Fable 5's guardrails were so conservative the model became a punchline on launch day. Both claims can't be fully true. If the pushback forces a reversal, it sets a precedent. Export controls would then apply to software-based AI, not just hardware. That distinction matters for every AI company with international operations.
PRESENTED BY WISPR FLOW
4x more context into every prompt. Zero extra effort.
You think faster than you type. Which means every typed prompt leaves out the constraints, examples, and edge cases that would have made the output actually useful.
Wispr Flow turns your voice into paste-ready text inside any AI tool. Speak naturally — include "um"s, tangents, half-finished thoughts — and Flow cleans everything up. You get detailed, structured prompts without touching a keyboard.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Free on Mac, Windows, and iPhone.
MICROSOFT
Nadella reframes AI competition around token capital
Getty Images
Recaply: Microsoft CEO Satya Nadella just outlined a framework for which companies will win the AI era, arguing that proprietary AI systems built on foundation models ("token capital") create more lasting value than the models themselves.
Key details:
Nadella splits corporate value into two types. Human capital covers knowledge, judgment, and relationships. Token capital is the proprietary AI systems a company builds on foundation models. His argument is that both grow together.
"The last thing any of us want is a world where every company is ceding value to a few models that eat everything they see," Nadella wrote. He warned that such concentration would have no political sustainability.
The framework arrived one day before OpenAI's financials leaked, showing $38.5B in net losses on $13.07B in revenue. The timing makes Nadella's case against betting on model builders more pointed.
Nadella published the post on June 14, ahead of earnings season and as the Fable 5 ban raised questions about AI platform concentration.
Why it matters: Nadella's framework arrives at a moment when the AI cost story is breaking open. OpenAI grew revenue 3.5x in a year and still widened its losses nearly 8x. That math only makes sense if model builders capture all the value. Nadella's bet is the opposite, that builders on top of models capture more. For any company evaluating AI investments, that argument deserves a stress test.
GUIDES
Build expressive AI text output with Ink-2
Recaply: In this tutorial, you will learn how to set up Ink-2 by Catersia, connect to its WebSocket API, and use its native turn detection events to capture clean speech transcripts and build low-latency voice AI pipelines.
Step-by-step:
Sign up at play.cartesia.ai to activate the 3-month free offer, then go to Settings → API Keys and generate your Cartesia API key
Install the Cartesia Python SDK with
pip install cartesia, then initialize the client:import cartesia; client = cartesia.Cartesia(api_key="YOUR_KEY")Connect to the Ink-2 WebSocket STT endpoint at
/stt/turns/websocketusing your API key and model IDink-2, then begin streaming 16kHz mono PCM audio in 200ms chunksListen for turn events in the WebSocket stream:
turn.startfires when the speaker begins,turn.eager_endfires when Ink-2 predicts speech has ended, andturn.enddelivers the final confirmed transcriptOn
turn.eager_end, send the current transcript to your LLM to generate a response, then route the response through Sonic-3.5 for audio playback to complete the full voice roundtrip
Pro tip: Use turn.eager_end instead of waiting for turn.end. Ink-2 fires it early while still finalizing the transcript, letting you start LLM inference immediately and cutting perceived response latency by 100-300ms in live conversations.
TOGETHER WITH NEO
Smarter browsing. Your data never leaves the room.
Most AI tools are a trade — your data for intelligence. Norton Neo breaks that deal. Powerful built-in AI, anti-fingerprinting, VPN, and ad blocking come standard. No setup. No add-ons. No compromises. Search, summarize, and write with AI that works inside your browser and stays there.
OPENAI
OpenAI's 2025 losses hit $38.5B, up nearly 8x
ED Zitron
Recaply: OpenAI just disclosed full-year financials showing $38.5B in net losses, nearly 8x the $5.09B it lost the prior year, with total spending reaching $34B as research and development alone hit $19.18B.
Key details:
Revenue grew from $3.7B to $13.07B, but costs outpaced the gain. R&D reached $19.18B, more than 5x OpenAI's total 2024 revenue, with $10.59B of that paid to Microsoft for model training compute alone.
Total net loss before noncontrolling interests was $60.35B, with $41.55B tied to a fair-value charge from OpenAI's conversion from nonprofit to for-profit. Net loss attributable to OpenAI directly was $38.53B, with roughly $50B in assets at year end, half in cash.
According to audited financials viewed by Ed Zitron and verified by the Financial Times, OpenAI paid Microsoft $17.2B total in 2025, with $10.59B for model training compute and $3.64B in liabilities to Microsoft remaining at year end.
These figures cover calendar year 2025, the same year OpenAI converted from nonprofit to for-profit. Its IPO has been reported as a 2026 target. OpenAI hasn't commented publicly on the figures.
Why it matters: OpenAI grew revenue 3.5x in a year and still widened its losses nearly 8x. That gap tells you something about the cost structure of the AI frontier. Building at scale means paying billions for compute while simultaneously investing almost $20B in R&D. Any builder choosing an API provider is choosing a platform whose sustainability depends on closing that gap. The longer it stays open, the more the math favors vertical consolidation.
TOOLS
Trending AI Tools
⚙️ Factory 2.0 - Autonomous software engineering platform expanding from coding agents to full software factories, in production at NVIDIA, EY, Adobe, and Palo Alto Networks
🎥 Seedance 2.0 Mini - BytePlus's video generation model with multimodal references, exceptional motion stability, and audio-video joint generation
🎤 Sonic-3.5 - Catersia's text-to-speech model streaming the first audio byte in 90ms, ranked #1 for naturalness across 40+ languages
🔊 Ink-2 - Catersia's speech-to-text model with native turn detection, ranked #1 for accuracy with sub-100ms transcript latency
NEWS
What Matters in AI Right Now?
Salesforce acquired Fin (formerly Intercom) for $3.6B, adding an AI customer service agent that resolves 76% of support queries without a human to its Agentforce platform.
Meta introduced AI Mode to Facebook search, powered by Meta AI (Muse Spark), delivering answers grounded in public Groups and Reels content rather than standard search links.
Sakana AI launched Marlin, its first commercial product, an autonomous research assistant that runs strategic research tasks for up to 8 hours and outputs structured summary slides and detailed reports.
Anthropic is facing a federal class-action suit in California over its Max 5x ($100/mo) and Max 20x ($200/mo) plans, with plaintiff Karl Kahn claiming a single 5-hour session consumed 15% of his weekly allotment.
SpaceX agreed to acquire Anysphere, the company behind AI coding agent Cursor, for $60B, with the deal expected to close in Q3 2026.
Catersia launched Sonic-3.5 and Ink-2, claiming the #1 spots on both the speech synthesis and speech-to-text accuracy leaderboards, with sub-90ms TTS latency and native turn detection built into Ink-2.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
