Skip to content
Main Site News Console

[AINews] AI Engineer Summit NYC: Gemini 4 Argon and Vals AI Benchmarks

· Latent Space Translated
播客深度访谈

Gemini 4 Argon is now generally available (GA) today, but we already covered it in detail two weeks ago as our lead story, while most benchmarks like CursorBench have merely independently confirmed the official claims.

Around 20,000 people attended AIE NYC today to hear keynotes from the frontiers of AI in finance: Anthropic showcased how they build modern AI financial services stacks, the Vals AI CEO discussed first principles behind benchmarking models like Argon, and Coatue’s Head of AI demonstrated how they turn news into investment decisions.

Transcripts for all keynotes are compiled here, with full recordings of all talks and workshops coming soon. Tune in tomorrow for Day 3:


AI Twitter Roundup

Frontier Model Releases: Gemini 4 Argon and Muse Wonder

  • Gemini 4 Argon: Google officially rolled out Argon globally for coding, knowledge work, enterprise workflows, and cybersecurity defense. Access is available across the paid Gemini API and Gemini platform; Google AI Pro and Ultra in the Gemini app; AI Mode in Google Search; and Ultra in Antigravity. This official launch refutes earlier newsletter claims that access would be restricted solely to cybersecurity defenders (official announcement).

    • Pricing: Introductory API pricing is $2 / $10 per million input / output tokens, with prompt caching (cached inputs) discounted by 95%. According to Artificial Analysis, rates will double to $4 / $20 after November 30—a crucial consideration when evaluating costs (Google pricing, discount schedule).
    • Research Access: Through an application program, Google is granting 10,000 researchers six months of free access to Antigravity’s highest tier; cybersecurity defenders can apply via Fairwind (scientist initiative, application portal).
  • Google Internal Results: Google stated that Argon helped untangle core web-processing logic from a 6-million-line legacy Chromium directory, completing in weeks what would have taken months and eliminating an entire class of browser crashes. Google also reported a 35% increase in agentic code submissions and a nearly one-third reduction in rollbacks. These represent internal engineering deployment claims rather than controlled productivity benchmarks (engineering report).

    • Physical Task Demo: DeepMind researchers demonstrated Argon inferring a cornhole simulation from human demonstration videos, performing system identification, and incrementally tuning throws on Apptronik’s Apollo robot—culminating in 7 consecutive successful tosses. Although one tweet mentioned “zero-shot,” the actual workflow described involved multiple physical robot attempts and refinements (researcher thread).
  • Developer Ecosystem Distribution: Argon is now integrated into Cursor, Devin Desktop/CLI, and GitHub Copilot. Cursor reported a 52.8% score on CursorBench at default effort, matching Opus 5.5; Cognition reported 60.5% on FrontierCode 1.1, offering a 75% Devin discount through October 27 (Cursor, Devin, Copilot).

    • App Integrations: Antigravity’s coding canvas is being integrated into the native Gemini macOS app, and Gemini now connects to over 40 apps, including Instacart, OpenTable, Adobe, and Shopify (coding canvas, connected apps list).
  • Muse Wonder Product Positioning: Meta’s release highlights personal assistant capabilities alongside coding. “Active Focus” targets long-horizon task persistence, failure recovery, and exploring alternatives; “Active Memory” focuses on capturing relevant, continuously updated personal context. Alexandr Wang explicitly highlighted consumer applications as a primary use case for Muse (capabilities discussion, Meta positioning).

    • Note on Memory: MemoryBench, cited in the announcement, primarily tests in-context fact updating within a single long conversation and does not establish persistent cross-session memory. A detailed breakdown also noted that certain headline win rates relied on internal private evaluations or omitted relevant competing models (detailed analysis).
  • Muse Rollout & Monetization: Strong coding performance has been met with complaints about steep pricing and restrictive limits. One user reported burning through their $200 plan’s 5-hour allowance quickly after several upgrades; another initially praised the generous limits, only to exhaust their entire $200 weekly quota in a single day. These represent anecdotal user experiences across specific workloads rather than definitive quota benchmarks (pricing complaints, weekly quota discussion).

    • Deployment Fixes: Wang acknowledged rollout issues and later stated that Meta has been actively patching deployment failures. Consequently, early impressions conflate genuine model capabilities with reactions to erratic deployment conditions.