Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models

October 1, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
October 01, 2026

HEADLINES

Trump Pulls Dario to the Podium as Labs Sign a Superintelligence Accord

A White House summit meant to produce the Accord on Super Intelligence was upstaged when President Trump pulled Anthropic CEO Dario Amodei forward to address the press. Amodei echoed the president's line that "whoever wins AI wins," calling for the industry to win, and win safely: a striking moment for a CEO who once told staff Anthropic was targeted for refusing to give "dictator-style praise" to Trump. The accord itself is a one-page commitment to build internal controls for frontier model testing and deployment and to partner with external auditors, which Trump called "morally binding." Mark Zuckerberg framed it as "a start," not the only thing the industry will ever do. Labs sitting in rooms with government beats labs viewing their role purely through competition, but the independent auditor piece still needs to be built out in practice.

"Super Intelligence" is Official, and America.gov Arrives

Following the summit, Trump signed two executive orders: one formally swapping "AI" for "super intelligence" in federal terminology, and a far more functional one establishing America.gov as an AI-powered single portal for government services. Unveiled in an Apple-style keynote with Marco Rubio playing Steve Jobs, the planned features (passports, post-marriage name changes, Medicare enrollment) amount to something like MCP for government, with agencies given 90 days to integrate and the functionality slated for next year. For now it is a knowledge base searching across 29,000 government websites. Privacy concerns spiked after Trump said it "cannot in theory be hacked," but Pliny the Liberator found the chatbot's guardrails unusually airtight.

FTC Opens a Probe Into Rogue Agents

According to agency sources speaking with the New York Post, the FTC opened an investigation into OpenAI and Anthropic a few weeks ago covering the Hugging Face incident and the dozens of rogue agent events disclosed since, and is now drafting civil investigative demands. Third-party safety lab METR can reportedly expect one as well. This is the existing-law approach long argued for by former FTC Chair Lina Khan, and cosigned by David Sacks in a case of strange bedfellows, but the FTC source's partisan framing on the record has critics questioning how serious the probe is. The optimistic read is trust but verify: accord signatories that fall short could face Section 5 enforcement.

MAIN STORY

Gemini 4 Argon, Sonnet 5.5 and What Models You Should Be Using Right Now

Google entered 2026 looking strong, until vibe coding and agentic harnesses like Claude Code and Codex reshaped the race and left it outside the top-lab conversation for most of the year. Gemini 4 Argon, its first frontier model in more than six months, posts benchmarks that say Google is back, yet almost no one can actually use it. And the race it is rejoining has changed: models now compete alongside harnesses, personal agents and product UX.

IS GOOGLE BACK?

Gemini 4 Argon
Google is so back, at least on the benchmarks.
Gemini 4 takes state of the art on the Vals Index, Vals Finance Agent and Harvey's legal agent benchmark, ties for third on the Artificial Analysis Intelligence Index, and pulls coding, its biggest weakness all year, into the frontier ballpark without leading it. Access is limited to "trusted cyber defenders" on cybersecurity grounds, with API and Ultra subscribers next, an echo of Google's announce-now, ship-later Gemini launch in December 2023.

The Benchmark-to-Reality Gap
"We literally can't even use the model."
Google's benchmark track record invites caution, and Bloomberg reports some employees with access say the model struggles on certain coding tasks, which Google denies. Logan Kilpatrick counters that new Gemini revisions now go through thousands of engineers for weeks before release, and even the optimists note a great model is only half the battle without a competitive coding harness and personal agent.

Claude Sonnet 5.5
An incredible model that you probably shouldn't use.
Sonnet 5.5 continues Anthropic's return to form, jumping to 70.6% on Terminal-Bench 4.0 (up from 10.3%, ahead of Opus 5.5) and scoring 56 on the AA Intelligence Index, second only to Opus 5.5. But it's token hungry enough that its max-settings AA run cost more than Opus 5.5, so a technical marvel that matches Fable 5 four months later may find its best home as an Opus subagent.

Muse
3 million weekly users in just over three weeks.
Meta's personal agent now has 1 million daily users, up from 500,000 weekly at the end of week one, outpacing Codex's roughly three-month climb to 3 million. That raises the real question: does great UX on a less-than-frontier model beat a frontier model wrapped in weaker UX? Some fans are already noticing the model's limits.

OpenAI Dots
"A plumber in a house where you're not allowed to touch the pipes."
Early gripes about OpenAI's new agent center on friction, above all that it's meant to manage work in ChatGPT and Codex without enough access to those tools and data. The UX-versus-model question is nowhere near settled.

DoorDash
"Order my usual protein bowl to the office."
DoorDash's agent now takes orders by text, matching a phone number to the customer's profile, usual order, address and payment, with agentic grocery orders already carrying 50% higher basket values. With Amazon blocking Muse agents, build-your-own versus partner is wide open, and DoorDash is hedging by building its own agent while keeping its platform open to third-party ones.