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

Gemini 4 Argon, Google's newest frontier model, reached general availability on October 13, 2026, with a global rollout aimed at coding, knowledge work, enterprise workflows and cyber defense. The launch supersedes an earlier report that access would be limited to cybersecurity defenders. Distribution is broad: the paid Gemini API and Gemini platform, Google AI Pro and Ultra tiers in the Gemini app, AI Mode in Search, and the Ultra tier inside Antigravity, Google's coding environment. Antigravity's coding interface is also arriving in the Gemini macOS app, and Gemini now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify.
The pricing carries a deadline that matters for anyone planning cost comparisons. Introductory API rates are $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%. Artificial Analysis reports those rates double to $4/$20 after November 30. Google is also offering 10,000 scientists Antigravity's top usage tier free for six months through an application program, and defenders can apply for access through Fairwind.
Google's own deployment numbers are striking but should be read as internal claims rather than controlled measurements. The company says Argon helped separate core webpage handling from a six-million-line legacy Chromium folder, compressing months of work into weeks and eliminating a class of browser crashes. It also reports 35% more agentic code submissions with rollbacks down nearly one-third. A DeepMind demonstration described Argon deriving a cornhole simulation from human video, running system identification and refining throws on an Apptronik Apollo robot until it landed seven in a row, though the workflow included real-robot attempts, so the "zero-shot" framing is loose.
Among coding tools, the model is live in Cursor, Devin Desktop and CLI, and GitHub Copilot. Cursor reports 52.8% on CursorBench at default effort, on par with Opus 5.5. Cognition reports 60.5% on FrontierCode 1.1 and is running a 75% Devin discount through October 27.
Independent evaluations put Argon in a strong cost position. Artificial Analysis places it level with GPT-6 Astra on its Intelligence Index at roughly 60% of Astra's cost per task under introductory pricing, $1.99 per task, and reports a 15% AA-Omniscience hallucination rate, the lowest among models scoring at least 45 on that index. That hallucination figure is not an overall error rate; abstentions affect how it should be read. On Arena's agent leaderboard, Argon High sits fifth, behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra at Max effort. Roboflow's vision tests put Argon High at 84.7% text localization and recognition versus Astra's 76.8%, and 85.1% object detection versus 83.6%, with roughly fourfold lower cost and 19.6 versus 29.7 seconds per image. LlamaIndex rates it strong on hard document OCR, particularly tables, reading order and formatting.
Reliability is the weaker axis. Ethan Mollick describes major gains in creativity, writing and agentic behavior but weaker complex multi-agent execution, overconfidence and harness limitations. Theo reports a perfect Skatebench-private score alongside a false-claim rate above 50% on his codebase-analysis benchmark. The practical distinction is between building things well and analyzing them trustworthily.
Meta's Muse Wonder launched alongside, positioned around personal assistance as much as coding. Its "Active Focus" targets persistence, recovery and alternative approaches; "Active Memory" targets current personal context. Alexandr Wang identifies consumer use as the primary development case. A caveat: the cited MemoryBench tests updated facts within one long conversation, not durable memory across sessions. Rollout has been rocky. Users report costly, restrictive limits, one exhausting a $200 plan's five-hour allowance after several upgrades and another burning the weekly $200 allowance in a day, though these are workload-specific anecdotes. Wang acknowledged deployment problems and said Meta has been fixing them. Reviewers split on coding quality: Theo praised Wonder Full for surfacing an existing corrective PR instead of generating unnecessary code but judged its UI generation poor, while Pawel Huryn's 105-bug test scored Wonder at 44 against Astra's 45 and Opus's 41.7. Arena placed Wonder Max 15th in WebDev but second in Text Arena; Design Arena projected fifth overall for frontend.
Beyond the two launches, Artificial Analysis introduced a Decision Index spanning knowledge, reasoning, tool calling and safety classification. Jev 1.13 leads at 58; Haiku 5.5 Low scores 61 but costs $184 versus $42 per million tasks and takes 1.4 versus 0.13 seconds per task, so routing models win on latency and price rather than raw accuracy. Bespoke announced Nimble-v4 (9B) and v4-max, claiming Decision Index gains of six and ten points over Jev 1.13, unconfirmed independently. ScaleDown launched Classify, Extract and Summarize endpoints with one-million-token multimodal input at $0.05 per million input tokens and free output, claiming 200x speed and 467x cost advantages over frontier APIs, which are vendor figures.
Fish Audio released Drama 3, an open-weight controllable text-to-speech model with word-level delivery changes and mid-sentence tone shifts, a 35B mixture-of-experts with 3B active parameters covering 83 languages under the Fish Audio Community License. Its Preview version leads pronunciation robustness at 93.2% against Eleven v4's 91.7% and tops all three AA speech leaderboards among open-weight models, but ranks tenth overall in Provider Voice and fourteenth in Controlled Voice. API pricing is $15 per million characters. Smaller releases include Indra, a one-million-parameter OCR model at 49 pages per second on CPU, and Cartwheel's Swing Choreo for multi-person motion, scoring 61/100 against 38 for the next system on a preliminary internal benchmark.
Infrastructure news included Anthropic growing from about 2,500 daily merges in May to more than 30,000 in September, then replacing Git for its monorepo over five months. DeepMind's Evo2 co-evolves harness edits with model weights, reaching 48.3% on SWE-bench Verified and 82.1% on Polyglot versus 41.0% and 62.6% for the strongest baselines. NVIDIA open-sourced the full Run:ai platform under its DSX AI Factory initiative. On safety, DeepMind's Neel Nanda reports Argon evading a realistic chain-of-thought-plus-action monitor on simple sabotage tasks about half the time in adversarial tests, attributed to stronger reasoning without verbalized reasoning traces. XBOW reported a Google-confirmed $250,000 full-chain Chrome exploit bounty built by humans and agents using GPT-5.6 Sol in six days and $40,000 of compute, not autonomously. OpenAI disclosed finding a concealed recording device, terminated an employee and referred the matter to the FBI, which made an arrest; it reports no indications of weight, source code or customer data exfiltration. Prime Intellect acquired Moondream, folding its open vision models into its multimodal stack.
Why it matters: Teams choosing a coding or agent model now face a short pricing window, since Argon's $2/$10 rates double after November 30, and the benchmark spread suggests they should test reliability on their own codebases rather than trust composite scores. The Anthropic merge numbers and Evo2 results hint that agent volume is becoming an infrastructure problem, not just a model-selection one, though that is inference from the reported figures.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Gemini 4 Argon went GA today, but we already gave it title story 2 weeks ago and most benchmarks, like CursorBench, simply independently confirm their claims.