Google on Wednesday introduced Gemini 4 Neon, a model it describes as delivering frontier-level intelligence at a lower cost per task, aimed at complex software engineering, analytical and scientific work.

The company said Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, placing third overall behind larger and slower models, and ranks second on the Vals Index while sitting on the Pareto frontier. Pricing is set at $3 per million input tokens and $15 per million output tokens, with cached input billed at $0.15 per million, a 95% discount. The model carries a 1 million-token context window.

Neon is rolling out immediately through the Gemini API in Google AI Studio, in Gemini for business via API and model dropdown, in Google Antigravity for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers.

MetricGemini 4 Neon resultArtificial Analysis Intelligence Index (xhigh)61, #3 overallVals Index#2, 70.9%Agent Arena Index#2, 1336 across 4,200+ sessionsTerminal-Bench 4.165.2%Artificial Analysis Coding Agent Index#1, 58.6HLE-Diamond71.3%

Note: Scores as published by Google at launch; benchmarks are third-party unless described as internal.

Google's own model card benchmarks Neon against Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol. Neon leads most rows and trails on three:

BenchmarkGemini 4 NeonClaude Opus 5.5GPT-6.2 SolAA Intelligence Index v5.0.1615957Vals Index v2.270.9%70.1%66.1%Terminal-Bench 4.165.2%65.8%61.5%FrontierCode 2.046.2%47.3%43.8%CursorBench 4.165.4%58.1%-SWE-sweep, provider harness39.0%10.1%8.3%AutomationBench56.3%42.5%41.3%Agents' Last Exam v1.175.8%71.6%73.0%Chartography73.5%66.3%75.1%GraphWalks, 256K to 1M95.2%66.8%80.1%

The three losses are slim, between 0.6 and 1.6 points, and two of them fall on coding evaluations where all three models sit within a few points of one another. The gaps that open up run the other way: 28.9 points on SWE-sweep under the provider harness and 28.4 points on GraphWalks, the long-context test, both against Opus 5.5, and 13.8 points on AutomationBench. The model card puts Neon's knowledge cutoff at March 2026 and its synchronous output at up to 64,000 tokens, rising to the full 1 million through a background long-decode mode.

Google positioned the release as much around cost and control as raw capability. At its default high thinking level, Neon pairs a high intelligence score with a 32% lower cost per task than Carbon, the company's top-tier model, and lands on the Pareto frontier of both the Vals index and Agent Arena at roughly 50% to 80% below top competing frontier models on cost per task. The xhigh setting adds deeper multi-step reasoning for the hardest engineering, analytical and scientific work, and still costs at least 36% less per task than rival models run at their highest setting, according to the company.

The list price tells a different story from the per-task cost until Nov. 30. Neon is $3 per million input tokens and $15 per million output tokens, while Carbon, the larger model it beats on cost per task, is $2 and $10 under an introductory rate that runs through Nov. 30. Those Carbon prices double to $4 and $20 on Dec. 1, from which point Neon is the cheaper of the two. Hydrogen, the smallest of the three, costs $1.50 and $7.50 and is the only one with a free API tier.

ModelInput, per 1M tokensOutput, per 1M tokensFree API tierGemini 4 Hydrogen$1.50$7.50YesGemini 4 Neon$3$15NoGemini 4 Carbon$2 through Nov. 30, then $4$10 through Nov. 30, then $20No

The 32% saving Google cites is measured per task at Neon's default thinking level, not per token, so the two comparisons measure different things. In the Gemini app, Neon comes with Google AI Pro at $19.99 a month in the US; Carbon requires Google AI Ultra, which starts at $99.99.

The model ships with four thinking levels that scale steadily, letting teams tune intelligence and token spend to each workload. Google said Neon also ranks second on steerability, or adapting cleanly to mid-task corrections, on the new Agent Arena Index, and posted its lowest fabrication rate yet in internal Antigravity testing.

On enterprise tasks, the company said Neon can synthesize spreadsheets, PDFs, charts, audio and video within a single context window and turn them into structured deliverables. It ranked first on the Legal Research Bench and Tax Agent Bench, first on most benchmarks across finance, accounting and spreadsheet modeling on Surge AI, and first overall on Zapier's AutomationBench at 56.3% and Andon Labs' Vending-Bench 2 with $16,843. It scored 73.5% on Chartography and 75.8% on Agents' Last Exam v1.1. On scientific and mathematical reasoning, Google cited best-in-class results on Terminal-Bench Science at 70.5%, RiemannBench at 84.0% and HLE-Diamond at 71.3%.

The coding story is central to the launch. Google said Neon is built for long-running agent workflows, whether running standalone or acting as an orchestrator directing parallel Hydrogen subagents inside Antigravity. It ranked first on the Artificial Analysis Coding Agent Index at 58.6, scored 46.2% on FrontierCode 2.0 at its default high setting, and leads the mini-SWE harness on SWE-sweep at 12.8%, a figure that more than triples to 39.0% inside Antigravity. The company credited the harness for that gap.

A separate cyber-focused variant, Gemini 4 Neon for cyber, is being released to trusted defenders in the coming days through the Fairwind Program. Google said it produces more secure code than Carbon across easy, medium and hard challenges on AutoBaxBench, reaching 76% on the hard tier, and is state of the art on WebExploitBench v1.2 at 51.26% at roughly half the cost of competing large models. It also cited 90.5% on SEC-bench Pro and 70.7% on Collinear's CWE-bench v2, a 3.4% improvement over Carbon.

On safety, Google said Neon maintains Carbon-level protections against severe misuse and prompt injection while improving alignment. It refused just 1.98% of tasks on Terminal-Bench 4.0 and none on Terminal-Bench-Science, matched Carbon with a 0.0% cheating rate on Terminal-Bench 4.0, and recorded a 0.09% misaligned action rate on its agentic safety suite alongside a 99.2% pass rate on its agent honesty evaluation. On Gray Swan's indirect prompt injection benchmark, the company said the model with real-time safeguards at its highest thinking level resisted every attack, a 0.0% success rate.

Neon slots into a three-model family. Carbon handles the hardest, longest-horizon projects, Hydrogen covers fast everyday execution, and Neon is pitched as the balance between depth and efficiency for high-frequency production work. The launch extends Google's push to convert benchmark leadership into cheaper tokens for enterprise buyers, a pricing lever that matters as agentic coding workloads multiply token consumption.