Google announces Gemini 4 Argon — its first frontier model in six months — with benchmarks that scream comeback, but no actual release to test them against. Add Sonnet 5.5's token-hungry brilliance, Muse's blistering growth, and Dots' friction-filled debut, and the real question comes into focus: in 2026, how much does the model alone even matter?
Benchmarks say Google is so back — but the race isn't about benchmarks anymore.
Gemini 4 Argon posts state-of-the-art scores across agentic knowledge work and closes the coding gap that kept Google out of the conversation all year. But you can't use it yet, Google's benchmark track record invites skepticism, and in the time they've been away the race has fundamentally changed: harnesses, user experience, and personal agents now matter as much as raw model capability. Sonnet 5.5 and the Muse-versus-Dots fight make the same point from different angles — being good on the model is no longer enough.
The president invited basically every big AI CEO to the White House — timed, the show notes, to steal the week from OpenAI DevDay. The subject: the heads of the frontier labs signing an accord on superintelligence. David Sacks's label for the gathering stuck.
— Dario Amodei, Anthropic CEO, to the press after the White House meeting. Trump pulled Dario Amodei forward to speak to the press — a striking choice for a CEO who previously told staff Anthropic had been targeted for refusing to give 'dictator-style praise to Trump.' Dario insists he's saying what he's always said; X spent the day psychoanalyzing whether this was deference or finally playing the game.
The labs committed to developing internal controls for frontier model testing and deployment, plus external auditors for verification. Trump called the document 'morally binding,' floated a 10-member industry oversight committee, and said he's seeing 'tremendous self-policing.' Zuckerberg framed it as a start, not the end state.
The media response was a Rorschach test on Trump and tech CEOs — skeptics like Chuck Todd see pure self-policing with no public input. NLW's middle position: it's genuinely better that these leaders are regularly in rooms with government rather than viewing everything through competition — but the 'independent external auditor' needs to actually get defined in practice.
The first makes Trump's terminology change official: the federal government will now say superintelligence, not artificial intelligence. The second, far more functional, establishes America.gov as an AI-powered single portal for government services.
Marco Rubio played Steve Jobs in an Apple-style keynote: provide your information once and the agent tracks down and completes forms across departments — passports, name changes, Medicare enrollment, coming next year. For now it's a knowledge base searching 29,000 government websites, and departments have 90 days to integrate.
Amid security concerns — not helped by Trump's 'it cannot in theory be hacked into, and when they figure out a way to do that, we'll end it' — jailbreaker Pliny the Liberator found the guardrails airtight: the chatbot flags anything resembling personal information and refuses the query until it's removed. 'Never seen that before.'
Per agency sources speaking to the New York Post, the investigation covers the Hugging Face incident and dozens of disclosed rogue-agent events, with civil investigation demands being drafted and plans to compel executive testimony. Third-party safety lab Meter can reportedly expect a demand too.
The investigation leans on a long-running thread: existing product-safety law already covers dangerous, unvetted, or defective AI products. Lina Khan's post making that case was co-signed by David Sacks — yet another example of the AI safety debate creating very strange bedfellows.
Matt Stoller calls it a protection racket — Trump investigates and clears the labs, creating hurdles for other probes. A former FTC public affairs director puts the credibility at '0.0%' after the White House love fest. The optimistic read: trust-but-verify, where falling short of the accord's commitments could trigger real FTC enforcement under Section 5. As with all policy right now, bring a whole bowl of salt.
After a year firmly outside the frontier conversation — no Gemini 3.5 Pro ever shipped — Google unveiled its first frontier model in over six months. Demis Hassabis claims frontier performance across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. The catch: it's announced, not released.
Gemini 4 scored 68.9% on the VALS index, beating Fable 5.1, GPT-6 Astra, and current leader Opus 5.5 — with bigger gaps on Automation Bench and VALS Finance Agent. On Harvey's legal agent benchmark it more than tripled the leading score, at 19.6%.
Argon is new state of the art on DeepSWE at 77.9%, but sits 10 points behind Astra on FrontierSWE and bottom of the pack on Terminal Bench 4.0. Two readings: failing at the single most important use case is a problem — or, given coding was Gemini's biggest weakness for a year, just being in the frontier ballpark is a big deal.
At $1.99 per task on the AA benchmark run versus $3.26 for Astra and $7.63 for Fable 5.1, Argon looks close to the Pareto frontier — but that's with a 50% launch discount of unstated duration. At full price it's more expensive than Astra, and it sits in an uncomfortable middle ground: more than twice the cost of GPT-61 Sol without a clearly better benchmark story.
Argon tied Grok 4.7 and Astra at 68% on the CWE bench cybersecurity benchmark — reason enough, Google says, to limit access to 'trusted cyber defenders' while engaging with the US government's voluntary pre-release testing program, before gradually expanding to API and Ultra subscribers. For the old heads, announcing without shipping echoes the original Gemini announcement of December 2023.
One camp — Ethan Mollick's 'it's a three-way race again,' Nathan Lambert's more-labs-at-the-frontier optimism — celebrates the benchmarks. The other remembers Gemini 3.1 Pro's official numbers 'practically destroying Opus 4.6 across the board' before reality intervened, and wonders why no one is skeptical given Google's track record.
Logan Kilpatrick — who NLW says deserves Google's MVP for weathering this year's PR challenges — engaged with the skeptics directly: new Gemini revs now go through thousands of software engineers for weeks before release, which should 'close the benchmark to reality gap by a real margin.'
Per people with direct access, the model does well on benchmarks but struggles with certain coding tasks when employees actually put it to work. Google denies the reporting, and another Bloomberg source says the gripers are a minority against a 'large consensus internally' that Gemini 4 is the frontier.
In the time Google has been away from the top, the competition became about more than models: Peter Yang's take is that Google 'cooked' on Gemini 4 but now needs to compete on the coding harness, Antigravity, and personal agent Spark. Frontier scores are table stakes, not victory.
After the disliked Opus 5 and a token-hungry Sonnet 5, Anthropic claims Sonnet 5.5 is 30% faster and 30% cheaper — and the benchmarks jump: 70.6% on Terminal Bench 4.0 versus Sonnet 5's 10.3%, even beating Opus 5.5. Artificial Analysis puts it second overall at 56 on the intelligence index, ahead of Fable 5.1 and GPT-6 Astra.
On AA's max-settings run, Sonnet 5.5 cost $7.60 per task — 27% more than Opus 5.5 and over twice Astra — because it burns significantly more tokens than Sonnet 5 at the same per-token price. In Anthropic's defense, turning settings down to extra-high cut costs by two-thirds, closer to their 30%-cheaper claim.
Builder Kun Chen finds Sonnet is 'definitely not the same as Opus just fifty percent cheaper' — Opus thinks strategically about ambiguous problems while Sonnet works tactically — so he runs Opus for planning, Sonnet for implementation, and Fable when things go sideways. Theo's verdict: an incredible model you probably shouldn't use standalone, because its token hunger means it's rarely more cost-effective than Opus except as a sub-agent.
Three weeks in, Muse has 3 million weekly active users and 1 million daily prompters — up from 500K weekly at the end of week one. Codex took about three months to reach the same mark. A tiny fraction of Meta's half-the-planet distribution, but a strong early signal for the personal agent form factor.
With OpenAI's Dots out, the question raised by Muse's success gets a direct answer. Early Muse fans are noticing the underlying model 'isn't smart enough' and say their loyalty is to intelligence, not form factor — while Dots users complain it has so much friction 'it feels like it's not meant to be used' and lacks sufficient access to ChatGPT and Codex data. NLW isn't close to declaring a winning strategy.
Customers can now text natural-language orders — 'order my usual protein bowl to the office' — and the agent matches the phone number, looks up the usual order and address, and handles payment. With agentic grocery orders carrying 50% higher basket value, DoorDash is hedging the big platform question: partner with category-leading agents or build your own. For now, it's doing both.
NLW sides with Nathan Lambert: Google doing well is better for consumers and better for competition, so he's firmly in the camp hoping this release is great. But the verdict has to come from hands-on use — not 'dining from the table scraps of self-reported benchmarks and gripey employees talking to Bloomberg.'