deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

Google launches Gemini 4 Neon, a frontier model for end-to-end engineering and science tasks

Google has launched Gemini 4 Neon, a frontier model aimed at completing complex tasks end to end in engineering, knowledge work and science, priced at $3 per million input tokens.

Google launches Gemini 4 Neon, a frontier model for end-to-end engineering and science tasks

Google has launched Gemini 4 Neon, a frontier-class AI model built to carry complex tasks through from start to finish across software engineering, knowledge work and science. According to Google's announcement, which surfaced on the Hacker News front page, the model is rolling out now in the Gemini API in Google AI Studio, in Gemini for business via the API and model dropdown, and for paid Google AI Pro and Ultra subscribers in both Google Antigravity and the Gemini app.

Pricing sits at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — a 95% discount — alongside a one-million-token context window.

Where it lands on the leaderboards

At its xhigh thinking level, Neon scores 61 on the Artificial Analysis Intelligence Index, which Google says puts it third overall, just behind larger and slower models. It ranks second on the Vals Index at 70.9% and sits on that index's Pareto frontier. On the new Agent Arena Index, drawn from more than 4,200 real-world agent sessions, Neon debuts second globally with a score of 1336, and Google also claims the number-two spot there for steerability — how cleanly the model adapts when a user corrects it mid-task. The company additionally reports its lowest fabrication rate to date in internal Antigravity testing.

Enterprise, analytical and scientific work

Google pitches Neon as a synthesizer of messy enterprise inputs, combining spreadsheets, PDFs, charts, audio and video inside a single context window and turning them into structured outputs. It claims first place on the Legal Research Bench and the Tax Agent Bench, top results on most Surge AI benchmarks spanning finance, accounting and spreadsheet modeling, and number-one finishes on both Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843), a long-running business simulation. For scientific and mathematical reasoning, Google reports best-in-class scores on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).

Built for long-running coding agents

Neon is engineered to keep its footing through long, tool-heavy agent sessions, Google says, trimming redundant file reads, sharing progress as it works and handling context compaction reliably. It can run standalone or act as an orchestrator coordinating parallel Hydrogen subagents inside Antigravity. On developer benchmarks, Google claims first place overall on the Artificial Analysis Coding Agent Index (58.6), plus 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting. The model handles multi-file debugging and repository-wide refactoring, and on SWE-sweep it scores 12.8% on the mini-SWE harness — a figure Google says more than triples to a state-of-the-art 39.0% when the model runs inside Antigravity.

Cost control and a security-focused variant

The model ships with four thinking levels that scale steadily, letting teams trade intelligence against token spend. At the default high level, Google says Neon pairs a strong intelligence score with a 32% lower cost per task than Carbon, and roughly 50% to 80% lower cost per task than leading competing frontier models. Even at xhigh, its deepest reasoning setting, it reportedly costs over 36% less per task than rivals at their highest settings. A "Neon for cyber" variant targets security work: Google reports state-of-the-art secure code generation on the hard tier of AutoBaxBench (76%), state-of-the-art penetration-testing results on WebExploitBench v1.2 (51.26%) at half the cost of comparable large models, and 90.5% on SEC-bench Pro for vulnerability finding, again at about half the per-challenge cost.

Why it matters

Neon's debut shows where frontier competition has shifted: away from raw benchmark peaks toward cost per completed task, mid-run steerability and reliability over long agent sessions. The threefold swing on SWE-sweep depending on the harness is a reminder that the environment around a model can matter as much as the model itself. The aggressive pricing, especially the steep cache discount, will pressure rivals on unit economics for high-volume agent workloads. One caveat: virtually every number here comes from Google's own announcement, and independent verification from third-party evaluators is still to come.

  • #gemini
  • #google
  • #large-language-models
  • #agentic-ai
  • #ai-benchmarks

Related posts