Google introduced Gemini 4 Neon on Tuesday, a new addition to its element-named Gemini 4 family that the company bills as delivering frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and scientific workflows. The announcement came from Koray Kavukcuoglu, SVP at Google DeepMind and the company's Chief AI Architect, in an official blog post.
The launch continues Google's rapid cadence of Gemini 4 releases, which already includes the Carbon, Hydrogen and Argon variants — the latter of which made headlines for autonomously playing through Minecraft's end credits. Neon positions itself as the fast, efficient tier of the family: strong enough to sit near the top of independent leaderboards while costing far less than the larger, slower frontier models it trails. Follow every major model release in our AI models coverage.
Where Neon Lands on the Benchmarks
Google says Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking third overall and ahead of comparable models while landing just behind larger, slower competitors. On the Vals Index it ranks second at 70.9%, placing it on the Pareto frontier of cost versus capability, and it takes first place on both the Legal Research Bench and the Tax Agent Bench.
The enterprise story is where Google is concentrating its claims. Neon synthesizes spreadsheets, PDFs, charts, audio and video within a single 1 million-token context window and turns those inputs into structured deliverables. Google says the model ranks first on most Surge AI benchmarks across finance, accounting, spreadsheet modeling and complex constraint following. It posts 73.5% on Chartography for reading dense professional charts and performs on par with leading models on Agents' Last Exam v1.1 (75.8%). It also debuts first on Zapier's AutomationBench at 56.3% and tops Andon Labs' Vending-Bench 2 — a long-horizon simulation of running a business — with $16,843 in simulated revenue. On Arena's new Agent Arena Index, built from more than 4,200 real-world agent sessions, Neon ranks second globally with a score of 1336 and ranks second on steerability, meaning it adapts cleanly to mid-task user corrections.
Scientific and mathematical reasoning gets its own showcase: Google claims best-in-class results on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%).
Built for Long-Running Coding Agents
Coding is arguably the marquee use case. Google says Neon ranks first overall on the Artificial Analysis Coding Agent Index with a score of 58.6, posts 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep, Neon leads with 12.8% on the mini-SWE harness and more than triples that to a claimed state-of-the-art 39.0% when running inside Google's Antigravity agent harness.
The model is engineered for long-running agent workflows: it sustains performance through complex, tool-heavy tasks, cuts redundant file reads, shares progress as it works and handles context compaction reliably across long sessions. It can also act as a fast, high-judgment orchestrator directing parallel Hydrogen subagents in Antigravity. Google's demos were characteristically showy: driving Blender MCP in Antigravity, Neon built a photorealistic 3D lab bench visualizing the Gemini 4 family with elemental accuracy — Bunsen flames for Carbon and Hydrogen, gas discharge tubes for Argon and Neon. Using the /swarm command, it recreated Alan Turing's 1936 study at King's College Cambridge in interactive 3D, complete with a working Universal Turing Machine on the desk. In another demo, an agent swarm turned a single image into a 100,000-cell Google Sheet running edge detection and 3D depth mapping through more than 500,000 live spreadsheet formulas, with no embedded images or Apps Script.
Tuned Pricing With Four Thinking Levels
Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.15 per million tokens — a 95% discount designed to make long-context agent loops economically viable. The 1 million-token context window spans the model's enterprise, analytical and scientific workloads.
Google is also refining its thinking-level ladder with this release. Neon ships four levels that scale steadily, letting teams dial intelligence and token spend to each workload. At its default high level, the company says Neon pairs a strong Artificial Analysis score with a 32% lower cost per task than Carbon, and sits on the Pareto frontier of both the Vals Index and Agent Arena with roughly 50% to 80% lower cost per task than top competing frontier models. The xhigh setting adds deeper multi-step reasoning for the hardest engineering and scientific workloads while still costing 36%+ less per task than competing models at their highest settings.
A Cyber Variant for Trusted Defenders
Google also positioned Neon as a security model. Gemini 4 Neon for cyber produces more secure code than Carbon across all easy, medium and hard challenges on AutoBaxBench, reaching state-of-the-art on the hard tier at 76%. On WebExploitBench v1.2 the model posts a claimed state-of-the-art 51.26% at half the cost of competing large models, and across the code-hardening lifecycle it scores 90.5% on SEC-bench Pro (vulnerability finding, at roughly half the usual cost per challenge), achieves frontier-level exploitation validation on ExploitGym using half the tokens of Carbon, and reaches 70.7% on Collinear's CWE-bench v2 for patching — a 3.4-point uplift over Carbon. The cyber variant rolls out to trusted defenders in the coming days through Google's Fairwind Program.
Safety Metrics for Production Work
Because Neon brings frontier-level capabilities to a faster, cheaper model, Google says it paired the release with real-time safeguards against cyber and biosecurity misuse comparable to Carbon's protections — while declining far less legitimate work, refusing just 1.98% of tasks on Terminal-Bench 4.0 and none on Terminal-Bench-Science. Internal alignment monitors showed alert rates below Carbon's baselines, and the model matched Carbon with 0.0% cheating on Terminal-Bench 4.0, took shortcuts on just 2.45% of tasks in Google's coding cheating evaluation — the lowest rate of any Gemini 4 model — and posted a 0.09% misaligned action rate on its agentic safety suite alongside a 99.2% pass rate on the agent honesty evaluation.
Availability
Gemini 4 Neon is rolling out today in the Gemini API in Google AI Studio, in Gemini for business via API and the model dropdown, and for paid Google AI Pro and Ultra subscribers in both Google Antigravity and the Gemini app.
Why It Matters
Neon's pricing puts frontier-adjacent agent performance within reach of teams that could not justify top-tier model costs for high-volume workloads. If the benchmark claims hold up in independent testing, the competitive picture for agentic coding, enterprise knowledge work and defensive security tightens further — and the gap between "fast enough" and "frontier" keeps narrowing. For more launches like this one, check the latest AI developments on AI Buzz Wire.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →