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

Gemini 4 Argon is Google’s first frontier model in more than six months, and its strongest results are in agentic knowledge work and legal tasks. It scored 68.9% on the VAALS Index, ahead of Fable 5.1 and GPT-6 Astra and a few points above Opus 5.5. On Harvey’s legal-agent benchmark, its 19.6% score was more than triple Fable 5.1’s. The results suggest a meaningful step forward, but not dominance across the board: Argon scored 77.9% on DeepSwee, compared with Opus 5.5’s 74.2%, while its 55% on FrontierSwee and 57.4% on Terminal Bench 4.0 left it behind Opus on the latter by nine points. Its 69.2% on OS World was also below Astra’s 72.6%.

On Artificial Analysis’s Intelligence Index, Gemini 4 Argon scored 53, tied for third with GPT-6 Astra and Fable 5.1. Its benchmark run cost $1.99 per task with Google’s 50% launch discount, but Google did not specify how long the discount would last. At full price, Argon would cost more than Astra; even discounted, it cost more than twice as much as GPT-6.1 Sol without a clear benchmark improvement. Google said cybersecurity concerns were why Argon would not initially be generally available. It scored 68% on a cybersecurity benchmark, so access is limited to trusted cyber defenders, with gradual expansion planned, beginning with API and Google Ultra subscribers. Google also said it is already using the model for internal work, including large-scale code migrations.

How much Argon’s scores predict practical performance remains disputed. Angel and Ishu Agrawal urged skepticism, pointing to limited public availability and past concerns about Google’s benchmark claims. Logan Kilpatrick said thousands of software engineers now test Google releases for weeks, while Bloomberg sources reported that some employees struggled with coding tasks; Google denied the model was underperforming. Peter Yang argued that Google still needs stronger coding harnesses, including anti-gravity and its Spark personal agent. The competing assessments underline that benchmark leadership does not by itself establish how well a model performs in ordinary coding workflows.

Anthropic said Sonnet 5.5 is 30% faster and 30% cheaper than Sonnet 5. It scored 70.6% on Terminal Bench 4.0, up from Sonnet 5’s 10.3% and above Opus 5.5’s 66.4%. Its Artificial Analysis Intelligence Index score of 56 placed it second behind Opus 5.5. Sonnet 5.5 remained slightly behind Opus on Frontier Code and Cursor Bench, but narrowed the gap on GDPVALAA and NAA Briefcase. Its price advantage is not automatic: a maximum-inference benchmark run cost $7.60 per task, more than Opus 5.5 and GPT-6 Astra, though extra-high settings cut its cost by two thirds. Theo said maximum reasoning can make Sonnet overthink and its token use can reduce its value as a standalone model, while it can work well as a sub-agent. Matthew Berman described it as comparable to Opus 5.5 on game-clone visual tests while 50% cheaper and much faster.

The discussion’s broader point was that model quality is only one part of competition: tools, agents, and user experience also shape adoption. Muse reached 3 million weekly and 1 million daily active users a little over three weeks after launch, up from 500,000 weekly and 250,000 daily users at the end of its first week. Sina said Muse’s underlying model still gets things wrong and argued that consumer loyalty follows intelligence. The host said it was too early to know whether a model-first or user-experience-first strategy would win, leaving open how much rapid adoption says about durable product preference.

At a White House meeting, frontier lab leaders signed an accord committing to internal controls for frontier-model testing and deployment, with external auditors meant to verify them. Mark Zuckerberg called it a starting point, while Chuck Todd argued that companies policing themselves without public input was inadequate; the host said the audit process remained unspecified. Trump called the accord morally binding and floated a 10-member industry oversight committee, but no committee was established. Separately, an executive order gives government departments 90 days to integrate services into the America.gov AI portal, which currently searches 29,000 government websites for official answers. Marco Rubio predicted passport applications and Medicare enrollment would be available next year, but their effectiveness and the portal’s security and privacy remain open questions. The narrator also said the FTC had opened investigations into OpenAI and Anthropic over rogue-agent incidents and was drafting civil investigation demands, while cautioning that the political context warranted skepticism.