AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 14, 2026

Google released Gemini 4 Argon globally on October 12, 2026, targeting coding, knowledge work, enterprise workflows and cyber defense. Access spans the paid Gemini API and Gemini platform, Google AI Pro and Ultra in the Gemini app, AI Mode in Search, and Ultra tier in Antigravity. The general rollout supersedes an earlier claim that access was limited to cybersecurity defenders. 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, which matters for any cost comparison built on launch pricing.

Google is also offering 10,000 scientists free access to Antigravity's highest usage tier for six months through an application program, and cybersecurity defenders can apply through Fairwind. Google's internal results claim Argon helped separate core webpage handling from a six-million-line legacy Chromium folder, compressing months of work into weeks and eliminating a category of browser crashes. Google further reports 35% more agentic code submissions with rollbacks reduced by nearly a third. These are internal deployment claims rather than controlled productivity measurements.

Distribution reaches beyond Google's own surfaces: Argon is available in Cursor, Devin Desktop/CLI and GitHub Copilot. Cursor reports 52.8% on CursorBench at default effort, on par with Opus 5.5, while Cognition reports 60.5% on FrontierCode 1.1 alongside a 75% Devin discount through October 27. Antigravity's coding interface is entering the Gemini macOS app, and Gemini can connect to more than 40 apps including Instacart, OpenTable, Adobe and Shopify. A DeepMind researcher account describes Argon deriving a cornhole simulation from human video, performing system identification and refining throws on Apptronik's Apollo robot before achieving seven consecutive successful throws; despite zero-shot framing in one post, the described workflow included real-robot attempts and refinement.

Meta's Muse launch emphasizes personal assistants alongside coding. Active Focus targets persistence, recovery and alternative approaches, while Active Memory targets relevant, updated personal context, and Alexandr Wang identifies consumer use as the primary development use case. A caveat worth noting: the cited MemoryBench evaluates updated facts within a long conversation and does not establish durable cross-session memory. Rollout economics drew complaints, with one user exhausting a $200 plan's five-hour allowance after several upgrades and another reporting consuming the weekly $200 allowance in a day — workload-specific anecdotes rather than a settled quota comparison. Wang acknowledged deployment problems and later said Meta had been fixing rollout issues.

On independent evaluation, Artificial Analysis places Argon 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, with substantial agentic gains and a 15% AA-Omniscience hallucination rate — the lowest among models scoring at least 45. That hallucination metric is not an overall error rate, since abstention changes how it reads. Arena places Argon High fifth behind Opus 5.5, GPT-6.1 Sol, Fable 5.1 and GPT-6 Astra at Max effort, and its revised Bash Reliability signal credits correct first-time command use while isolating agent-caused failures. Roboflow reports Argon High at 84.7% text localization and recognition versus Astra's 76.8%, and 85.1% object detection versus 83.6%, at roughly fourfold lower cost with latency of 19.6 versus 29.7 seconds per image. LlamaIndex finds Argon, Muse Spark and Wonder strong on hard document OCR, citing table understanding at their respective price points plus competitive chart, reading-order and formatting recognition.

Reliability is more mixed. Ethan Mollick describes major improvement 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 greater-than-50% false-claim rate on his codebase-analysis benchmark, and judged Wonder poor at UI generation even as Design Arena projected fifth overall frontend and a top-three website placement, with Arena later placing Wonder Max 15th in WebDev but second in Text Arena. His real-work comparison praised Wonder Full for finding an existing corrective PR instead of writing unnecessary code. A separate 105-bug test reports Wonder at 44 versus Astra's 45 and Opus's 41.7. My reading is that these numbers measure different workloads and preferences, so they should not be collapsed into a single quality ranking; the practical distinction is between strong construction ability and trustworthy analysis.

Elsewhere, Artificial Analysis launched a Decision Index covering knowledge, reasoning, tool calling and safety classification, with Jev 1.13 leading published results 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. 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 — vendor claims. Bespoke announced Nimble-v4 and v4-max, claiming Decision Index gains of six and ten points over Jev 1.13, with Nimble-v4 at 9B parameters and v4-max's size unspecified. Fish Audio released Drama 3, a controllable open-weight text-to-speech model supporting word-level delivery changes and mid-sentence tone shifts; Artificial Analysis identifies it as a 35B MoE with 3B active parameters across 83 languages under the Fish Audio Community License. Drama 3 Preview leads pronunciation robustness at 93.2% versus Eleven v4's 91.7% and ranks first among open-weight models on all three AA speech leaderboards, though tenth overall in Provider Voice and fourteenth in Controlled Voice, at $15 per million characters.

On infrastructure, Anthropic reports growing from roughly 2,500 daily merges in May to more than 30,000 in September, then building a new version-control system and migrating its monorepo away from Git in five months. DeepMind's Evo2 co-evolves harness edits and model weights through proposer and reflector roles, SFT and interleaved GRPO, reaching 48.3% SWE-bench Verified and 82.1% Polyglot versus 41.0% and 62.6% for the strongest baselines. Meta and MIT's ReHarness rewrites retrieved reflections using downstream task outcomes rather than only adding or deleting memories, and LongHarness separates evidence and conclusions into explicit state, with authors reporting GLM-5.3-Flash improving from 4.4% to 61.1% exact match on OfficeQA Pro V2 without changing the model. Kylon launched a workspace for humans and existing agents such as Claude Code and Codex, aimed at coordination and review overhead. ThunderSync computes gradients as rewards or teacher scores arrive, overlapping training with ongoing execution.

Why it matters: teams choosing a frontier coding model now face a genuine price-performance decision, since Argon's introductory rates undercut Astra per task but double after November 30, so cost models built on launch pricing will be wrong within weeks. Developers standardizing on Argon across Cursor, Copilot and Devin should plan evaluation around the weaker reported areas — multi-agent execution and confident codebase analysis — rather than headline benchmark scores. The Anthropic merge data and the harness papers together suggest inference is that agent-driven load is now straining source control and evaluation tooling, not just model serving.

Read at Latent Space

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.