Skip to content
AIBRAINWORX / THE INDEPENDENT AI LABRESEARCH · HARDWARE · OPEN INTELLIGENCE
Models & Weights / Enterprise & Legal / Model Watch

Google DeepMind Launches Gemini 4 Neon for Agentic Coding and Enterprise Workflows

Google DeepMind introduces Gemini 4 Neon with a 1 million-token context window, configurable thinking tiers, and multi-agent orchestration in Google Antigravity.

Listen to this articleGoogle Voice
0:00
Official company logo or product image for google-deepmind-gemini-4-neon-launch
Official image or organization logo from deepmind.google; identification only, no endorsement.
Event / context: 2026-11-18. Source check: October 28, 2026.

Google DeepMind has released Gemini 4 Neon, a frontier-tier commercial model built for fast end-to-end task completion across software engineering, quantitative research, and enterprise knowledge work. Announced by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, the model prioritizes mid-task steerability, multi-step instruction following, and reduced fabrication in long-running workflows rather than chasing raw latency-heavy scale alone.

Gemini 4 Neon Targets Fast End-to-End Agentic Workflows

Positioned alongside larger models in the Gemini 4 family, which Google DeepMind references alongside sibling variants Gemini 4 Carbon, Gemini 4 Hydrogen, and Gemini 4 Argon, Neon aims to sit on the speed-to-capability frontier for production automation. The model is rolling out immediately through the Gemini API in Google AI Studio, Gemini for business via both API and model dropdown, Google Antigravity for paid Google AI Pro and Ultra subscribers, and the consumer Gemini app for Pro and Ultra tiers.

Rather than treating prompts as single-turn completions, Google DeepMind designed Neon to ingest mixed enterprise formats within one session. The model accepts spreadsheets, PDFs, dense charts, audio, and video natively inside a unified 1 million-token context window and converts those inputs into structured deliverables.

Context Window, Pricing, and Orchestration Architecture

At the API layer, Google DeepMind prices Gemini 4 Neon at $3 per million input tokens and $15 per million output tokens, with cached input tokens discounted by 95% to $0.15 per million. To help engineering teams balance reasoning depth against cost, Neon exposes four configurable thinking levels, including a default high setting and a top-tier xhigh setting, allowing predictable scaling between lightweight tasks and deep multi-step planning.

For sustained agentic execution, Neon incorporates context compaction for long sessions, cuts redundant file reads, and streams progress updates mid-task. Inside Google Antigravity, Neon can operate either as a standalone coding agent or as a high-judgment Orchestrator directing parallel Hydrogen subagents via the platform's /swarm command. Google DeepMind did not disclose parameter counts, training compute, or underlying TPU hardware specifications in the launch announcement.

Benchmark Claims Across Coding, Science, and Enterprise Tasks

Google DeepMind reports broad benchmark gains across third-party and internal evaluations, though independent verification will determine how well these numbers hold in production environments. At its xhigh thinking level, Neon scores 61 on the Artificial Analysis Intelligence Index, placing third overall behind larger, slower models. On the Vals Index, it ranks second overall at 70.9% and places first on both the Legal Research Bench and Tax Agent Bench, alongside top rankings across Surge AI benchmarks covering finance, accounting, spreadsheet modeling, and complex constraint following.

In multimodal and long-horizon automation tests cited by the company, Neon reaches 73.5% on Chartography, 75.8% on Agents' Last Exam v1.1, 56.3% for first place on Zapier's AutomationBench, and $16,843 for first place on Andon Labs' Vending-Bench 2 business simulation. On Arena's Agent Arena Index spanning more than 4,200 real-world agent sessions, it debuts second globally with a score of 1336 and ranks second on mid-task steerability. Scientific and mathematical scores include 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, and 71.3% on HLE-Diamond.

On software engineering tasks, Neon ranks first on the Artificial Analysis Coding Agent Index at 58.6, scores 65.2% on Terminal-Bench 4.1, and reaches 46.2% on FrontierCode 2.0 at its default high setting. Notably, its performance on SWE-sweep illustrates how heavily agent results depend on scaffolding: Neon scores 12.8% on the mini-SWE harness, then climbs to 39.0% when paired with the Google Antigravity execution harness.

Developer and Enterprise Workflow Integration

For software teams, Google DeepMind emphasizes multi-file debugging, targeted repository refactoring, and interactive multi-agent builds. In Antigravity demonstrations, Gemini 4 Neon used /swarm and Blender MCP to assemble a 3D lab visualization pairing Bunsen flames and gas discharge tubes to represent the Gemini 4 lineup, and recreated Alan Turing's 1936 study at King's College, Cambridge from archival photographs with a working Universal Turing Machine.

Additional Antigravity showcases include generating a WebGL2 freeride ski simulator over Mont Blanc and Chamonix elevation data using 10,200 lines of custom shaders and 120 Hz physics with no pre-made assets, as well as converting a single image into a 100,000-cell Google Sheet powered by over 500,000 live formulas for edge detection, depth mapping, and interactive relighting without Apps Script or embedded images.

Commercial Access, Harness Dependencies, and Verification Caveats

Despite strong headline scores, practitioners face clear operational and licensing constraints. Gemini 4 Neon is a proprietary, closed-weights service gated behind Google's commercial APIs and paid subscriptions, offering no local self-hosting option or weight inspection for regulated on-premises deployments.

Furthermore, the lowest fabrication rate cited by Google DeepMind comes from internal Google Antigravity testing rather than audited external studies, and the nearly threefold jump on SWE-sweep between mini-SWE and Antigravity underscores that developer outcomes remain tightly coupled to harness design and tool orchestration. Organizations applying Neon to legal, tax, financial, or production code pipelines will still need strict human review and domain-specific regression testing.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-18.

Follow your curiosity.

Pass the signal.