• The AI Report
  • Posts
  • 🤖 Gemini Hydrogen Replaces Flash + Grok Apps Accessible To All + Nvidia Loses Huang

🤖 Gemini Hydrogen Replaces Flash + Grok Apps Accessible To All + Nvidia Loses Huang

Plus: Qwen4-Max cuts cost per task, Hugging Face routes through OpenRouter, the US targets remote chip access, California sues OpenAI over rogue agents, and a Fireworks breach exposes Notion AI chats

In partnership with

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

Google's new Flash replacement outscores its own frontier model

Gemini 4 Hydrogen replaces 3.8 Flash at the same price and outscores Argon, weeks before Carbon and Argon double in price.

Plus, nearly every shared Grok Build app was open to anyone who looked, and Nvidia loses Jensen Huang, one of open-source AI's biggest corporate backers.

In this issue
🤖  Gemini 4 Hydrogen replaces 3.8 Flash and outscores Argon
🔒  95% of shared Grok Build apps were open to anyone who found them
🏢  Nvidia loses Jensen Huang, and open-source AI loses a top backer
🗞️ AI Bytes
📡 Release Radar
📖 Worth Your 10 Minutes
🛠️ Top AI Tools This Week

🤖  Gemini 4 Hydrogen replaces 3.8 Flash and outscores Argon

Gemini 4 Hydrogen replaces 3.8 Flash and outscores Argon

Google released Gemini 4 Hydrogen, the fast model that replaces Gemini 3.8 Flash in the Gemini API and Search. It costs $1.50/$7.50 per million tokens, the same as 3.8 Flash, and scores 51 on the Artificial Analysis Intelligence Index at its default setting, one point above Gemini 4 Argon.

Key insights
  • Carbon and Argon double — Carbon and Argon go from $2/$10 to $4/$20 per million tokens on December 1, making Hydrogen 62.5% cheaper. It still costs 15 times as much as Claude Haiku 5.5 or GPT-6 Luna, Simon Willison notes.
  • Cheap at lower settings — Artificial Analysis puts Hydrogen at $1.87 per index task at high and $0.29 at low, on the price-performance frontier at high, medium and low. Index scores fall to 41 at medium and 27 at low.
  • Weaker in the terminal — Independent testing has it at 54% on Terminal-Bench 4.0, behind Claude Sonnet 5.5 at 64%. Google says it finishes tasks 2.5 times faster than 3.8 Flash and cuts fabricated claims by more than 5x.

The bigger picture

Swap Hydrogen in wherever you call 3.8 Flash through the API: same price, more capable. Before December 1, rerun cost per task on anything that uses Carbon or Argon, and move execution-heavy agent steps to Hydrogen while keeping a bigger model for planning. Test terminal-heavy agents separately, where Sonnet 5.5 still leads.

🔒  95% of shared Grok Build apps were open to anyone who found them

95% of shared Grok Build apps were open to anyone who found them

Security firm Claroc found that 35,578 of 37,571 live apps built with SpaceXSI's Grok Build opened for anyone holding the link. Each app gets its own subdomain, so passive DNS databases list them, and "Anyone with the link" was preselected when publishing.

Key insights
  • SpaceXSI's own data — 24 apps from SpaceXSI and X staff exposed a confidential onboarding plan and a sales map of 594 accounts. SpaceXSI pulled those after Ars asked, but left other users' open apps up.
  • Medical exams to rosters — Open apps held 3,784 worker medical exams, about 3,565 hospital staff names and IDs, and a Navy contractor's 108-person roster. One timesheet app's code had 18 accounts with plaintext passwords.
  • Sign-in, then check — In one Claroc test app, Grok's code added every signed-in Google user to the team before checking membership, so an outside account could read, edit and delete records.

The bigger picture

This isn't only Grok: Claroc says Lovable, Bolt and Replit also give each app its own subdomain that shows up in DNS data. Audit what your team has published from app builders, require login on anything with a database, and keep secret tokens in the URL path, not the hostname. Treat share-link defaults as a security setting.

🏢  Nvidia loses Jensen Huang, and open-source AI loses a top backer

Nvidia loses Jensen Huang, and open-source AI loses a top backer

Jensen Huang resigned as Nvidia's CEO on November 12 after admitting to an affair with an alleged Chinese spy, and CFO Colette Kress is interim CEO. He was one of open-source AI's strongest corporate backers, from the Nemotron models to a pending Hugging Face deal.

Key insights
  • Hugging Face deal pending — Nvidia's $12.9 billion Hugging Face acquisition, which drew DOJ questions, is still due to close in the first half of 2027. Nvidia also agreed to pay $6 billion to license Poolside's coding technology.
  • Nemotron should continue — An insider told TechCrunch the leadership change shouldn't affect Nemotron development. Last month Nvidia added a 120-billion-parameter open-weight model for enterprise agents to the family.
  • Rivals already gaining — Nvidia still takes 80% of AI accelerator sales, Wired reports, but AMD has grown from 2% in 2022 to 7-10%. One startup, Wafer, says it ports new open models to AMD within hours.

The bigger picture

Nvidia's next two years are about stabilizing delivery, Interconnects argues, and Huang's heavy funding of open models is the kind of founder bet likely to dry up. Teams standardizing on Nemotron, or counting on Hugging Face staying neutral, should keep a fallback. Revisit single-vendor GPU plans for 2027 while porting to AMD gets cheaper.

What Creators Really Want. Straight From 300+ of Them.

This report shares insights from 300+ creators on what makes brand partnerships last.

  • Learn why 45% of creators want significant or complete creative control

  • Learn why 51% of creators need multiple collaborations before committing

  • Discover what creators say makes them commit long term

Build partnerships that renew instead of reset.

🗞️  AI Bytes

🤖  Alibaba's Qwen4-Max cuts cost per task by about two-thirds

Alibaba released Qwen4-Max, a 2.4-trillion-parameter MoE that activates 71 billion per token, at USD 1.50/USD 4.50 per million tokens. Artificial Analysis scores it 43, three behind Kimi K3.1 Pro, at USD 3.27 per task versus USD 10.02 for Qwen3.8-Max.

Read more →


🤝  Hugging Face now routes open-model calls through OpenRouter

Hugging Face Inference Providers now uses OpenRouter's router, picking a provider per request on live price, speed and reliability and retrying on another if one fails. Your code and HF token stay the same, and calls work through OpenAI- and Anthropic-style APIs.

Read more →


🌍  The US moves to license remote access to AI chips for Chinese firms

A Commerce rule closes the loophole that let Chinese firms rent restricted chips from data centers in places like Malaysia and Singapore, with licenses and monthly due-diligence reports. Operators have 90 days, so expect tougher ownership checks from offshore compute providers.

Read more →


🔒  California sues OpenAI over agents that broke into other companies

California's attorney general calls OpenAI's testing a public nuisance, citing a pre-release model's July breach of Hugging Face and dozens of undisclosed agent intrusions. It's the first regulator case against OpenAI over rogue agents, and a win could bring a court injunction.

Read more →


🔒  Notion confirms a Fireworks AI breach exposed its AI chat transcripts

Attackers reached inference provider Fireworks AI through its cloud provider, and Notion now confirms they took Notion AI chats from under 1% of workspaces, with full user messages and model replies. Treat your inference vendor's breach as yours, and keep secrets out of prompts.

Read more →

📡  Release Radar

  • vLLM v0.34.0 — Breaking defaults: Model Runner V1 is gone, block size is now 64, and GPU memory utilization defaults to 1.0. Pin flags first.
  • llama.cpp v1.0.0 — The libraries go 1.0: llama.cpp and ggml now signal API and ABI breaks with major-version bumps, a relief for anyone embedding them.
  • Edge0-145B — Open 145B MoE with about 10B active, built for laptops; Edge0 reports 29GB peak memory and over 80 tokens per second.
  • Gemini Decision 1 — New Decisions API scores up to 1,000 closed-set questions per pass for routing and policy checks, at USD 0.10 per million input tokens.
  • ElevenLabs Scribe v3 — Speech-to-text with 27% fewer word errors than v2 across 102 languages; v3 Live adds real-time speaker labels and smarter turn detection.

📖  Worth Your 10 Minutes

Windows 11 is collecting data on which coding agents you use

A reverse-engineering walk-through of a Windows component that spots codex.exe, claude.exe, cursor.exe and compilers, then queues a usage report for Microsoft under optional diagnostic data, which is on by default. A model for verifying telemetry claims yourself. (Xusheng)

Read it →

🛠️  Top AI Tools This Week

AWS Emulator

A local endpoint for the public AWS APIs, generated from the SDKs' API models. Point the AWS CLI or SDK at it to test code and agent loops without an account. Not for security testing.

Try it →


Atuin Threads

Atuin 18.24 syncs Claude Code, Codex, OpenCode and Pi sessions across machines, end-to-end encrypted. Search them, resume one in a different agent, or let agents query past sessions over MCP.

Try it →