Gemini 4 Carbon takes the top spot, but only security teams get to touch it

Good morning 👋 Google just retook the No. 1 spot on the model leaderboards. It's handing the keys to almost nobody.

In today's issue:

• 🔭 Gemini 4 Carbon tops the leaderboard, but only cyber defenders get in
• 🧠 NVIDIA ships an open transcription model, OpenAI moves into ServiceNow
• 🔬 Why summarizing a group debate makes an LLM dumber, plus two more papers
• 🏛️ The White House starts a 180-day clock on robotics tariffs
• 🔗 FuriosaAI's valuation quadruples, Oracle caps its AI spending

🔭 THE ONE THING

🛡️ Gemini 4 Carbon takes the top spot, but only security teams get to touch it

Google DeepMind shipped Gemini 4 Carbon this week and it landed atop the Artificial Analysis Intelligence Index v5.0 with a score of 59, nosing out Opus 5.5 (58) and GPT-6 Astra (57). The pitch isn't general intelligence, it's security: 88.2% on SEC-bench Pro's vulnerability detection, 71% on CWE-bench patching, and a clean 0.0% attack success rate on Gray Swan's prompt injection benchmark. That's a model tuned to find bugs and resist being hijacked while doing it, and the numbers hold up under independent scrutiny. Google isn't opening it up, either: access runs through the Fairwind Program, cyber defenders only, with intro pricing of $2/$10 per million tokens that climbs to $4/$20 once the honeymoon ends. A one-point lead at the top of a benchmark is thin enough to shrug off; a frontier lab gating its best model specifically to defenders, not red-teamers, not the general API, is the part worth sitting with.

🧠 MODELS & RELEASES

• 🎙️ NVIDIA shipped Nemotron 3.5 Transcribe, an open-weight speech model that transcribes 40 languages, translates 31, and tells 13 languages' worth of speakers apart, running a Mamba-Transformer mixture of experts with just 4B active params. It's already first on the Hugging Face Open ASR leaderboard (5.33% WER, English), and the Granary v2 dataset behind it, 463,000 hours across 48 languages, ships open too. Hugging Face
• 🤝 OpenAI and ServiceNow are putting Codex inside ServiceNow's app-builder and a ChatGPT plugin in front of IT teams, so an admin can close an incident or run a security triage without leaving the chat window. Convenient for both sides: ServiceNow gets the OpenAI halo, OpenAI gets a few hundred thousand more enterprise seats to count. OpenAI

🔬 RESEARCH HIGHLIGHTS

• Memory kills the signal Feed an LLM group's debate through a summarizer before the final vote, and the two obvious compression rules both break on CrossProfile, a new 480-task benchmark run across Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models. Keep what's relevant and you readmit the decoys, since decoys are relevant by design, costing 23 to 31 points versus keeping rare claims instead. Keep what's rare and you're no better than chance unless the rare stuff happens to be true. Strip the who-said-it tags (models give extra weight to claims more agents repeated) and accuracy jumps 13 to 24 points. The full, unedited transcript scores worse than no notes at all. Rare, Relevant, and Still Wrong

• GEPA-Curator patches a quiet assumption behind self-improving agents: that every failed rollout is a lesson worth learning. Plenty are just a broken harness or a noisy judge. The fix filters those out and feeds the optimizer its own past wins as extra context. Across GoBrowse, ALFWorld, AppWorld, and tau-bench, with Gemini 3.1 Pro and Claude Opus 5 doing the optimizing, it beats vanilla GEPA by up to 20.7% and ACE by up to 21.2%. Curating the data mattered more than collecting more of it. GEPA-Curator

• Can an LLM find a new law of physics? LongChang drops agents into simulated worlds where one physical law is quietly broken, heavier objects falling slower, say, and lets them poke around with no hints. Twelve models tried. GPT-6-Astra found the altered law 16.7% of the time. The other eleven scored zero. Open-ended scientific discovery is still mostly theater. Can LLMs Uncover Hidden Physical Laws through Free Exploration?

🏛️ POLICY & REGULATION

• The White House issued a Section 232 proclamation on robotics and industrial machinery imports, but read the text before you panic about tariffs. Commerce already found the imports threaten national security after a 90-day investigation; what the proclamation actually orders is 180 days of negotiation between Commerce and USTR before any duty gets written down. No rate, no HTS codes, no quota, just a clock and a threat. Builder read: this is a warning shot for anyone sourcing robotics hardware or parts overseas. Price in tariff risk on your BOM now, because the legal trigger is already pulled.

🔗 QUICK LINKS

• FuriosaAI raised $620M in a Series D that quadruples its valuation to $2.9B, betting data centers want an inference alternative to Nvidia. Bloomberg
• Oracle told investors it won't take on more than $20B in new AI financing through fiscal 2030, the clearest ceiling yet on its spending spree. Finimize

See you tomorrow.

Pradeep Perugu