Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

Google ships Gemini 4 Neon, targets coding-agent and science benchmarks

Gemini 4 Neon's threefold SWE-sweep score swing between harnesses exposes how much coding-agent performance depends on tooling rather than the model itself.

1Introducing Gemini 4 Neonblog.googleHigh11·18 21:00Z

Google DeepMind released Gemini 4 Neon on November 18, a model positioned between Carbon and Hydrogen that brings frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work, and science workflows1. At its default high thinking level, Google reports Gemini 4 Neon costs 32% less per task than Carbon11,12. Four thinking levels let teams trade intelligence for token spend. Gemini 4 Neon replaces Argon in the Gemini app's model picker, where it now occupies the "Deeper analysis" slot9. The model is described as three times faster than Carbon3,7.

The model is priced at $3 per million input tokens — with cached input tokens at $0.15 per million, a 95% discount — and $15 per million output tokens, and carries a 1 million-token context window. Those rates are twice Hydrogen's input and output prices and 25% below Carbon's standard rates.

Benchmark positioning

Gemini 4 Neon scored 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, ranking third overall and ahead of comparable models, landing behind larger, slower models. On the Vals Index, the model ranked second and landed on the Pareto frontier. It ranked second on steerability — adapting to mid-task user corrections — on Arena's new Agent Arena Index.

Across coding evaluations, Gemini 4 Neon ranked first overall on the Artificial Analysis Coding Agent Index with a score of 58.66. It scored 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep, the model reached 39.0% inside Google Antigravity and led on the mini-SWE harness at 12.8%2.

Google's benchmark table shows Gemini 4 Neon scoring 56.3% on AutomationBench and 75.8% on Agents' Last Exam, ahead of Claude Opus 5.5 and GPT-6.2 Sol in both tests. Opus 5.5 came out slightly ahead on Terminal-Bench 4.1 and FrontierCode 2.0, while Sol scored higher on Chartography.

Harness gap on SWE-sweep

The SWE-sweep results exposed a wide gap between harnesses. Gemini 4 Neon resolved 12.8% of SWE-sweep using the mini-SWE harness but jumped to 39.0% inside Google Antigravity, a more-than-threefold improvement. Claude Opus 5.5 showed a similar pattern, climbing from 3.6% with mini-SWE to 10.1% with Claude Code. GPT-6.2 Sol went the other direction, falling from 11.6% with mini-SWE to 8.3% with Codex. A separate FrontierHarness benchmark found that running the same open model could cost up to five times more depending on the harness.

In science and knowledge benchmarks, Gemini 4 Neon scored 71.3% on HLE-Diamond, 84.0% on RiemannBench, and 70.5% on Terminal-Bench Science.

Agentic tasks and rollout

On Arena's Agent Arena Index, the model debuted second globally with a score of 1336 across more than 4,200 real-world agent sessions. It took first place on Andon Labs' Vending-Bench 2 with a score of $16,843, on Zapier's AutomationBench at 56.3%, and on both the Legal Research Bench and Tax Agent Bench. It scored 73.5% on Chartography. In the Gemini app, Gemini 4 Neon can run multi-step Tasks and draw on connected services such as Gmail and Docs.

Google reported that Gemini 4 Neon records its lowest fabrication rate yet in internal Google Antigravity testing. The model achieved a 99.2% pass rate on its agent honesty evaluation and a 0.09% misaligned action rate on its agentic safety suite. On Gray Swan's IPI benchmark, Gemini 4 Neon resisted every attack at its highest thinking level. It matched Carbon with 0.0% cheating on Terminal-Bench 4.0 and took shortcuts on 2.45% of tasks on its coding cheating evaluation.

On cybersecurity benchmarks, Gemini 4 Neon reached 70.7% on Collinear's CWE-bench v2, a 3.4% uplift over Carbon. It achieved 90.5% on SEC-bench Pro, 51.26% on WebExploitBench v1.2, and 76% on the hard tier of AutoBaxBench. A cyber-specific variant is rolling out to trusted defenders through the Fairwind Program.

Gemini 4 Neon is available through the Gemini API in Google AI Studio, Android Studio, Gemini for business, and for paid Google AI Pro and Google AI Ultra subscribers in Google Antigravity and the Gemini app. Vals reports a maximum output limit of 1 million tokens5. The model supports spreadsheets, PDFs, charts, audio, and video in the same request. The announcement was authored by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google.

ANALYSIS The release arrived the same day Google DeepMind rolled out Genie 4 to Google AI Ultra subscribers[4], broadening the batch of model updates shipping under the Gemini 4 generation. Gemini 4 Neon's cached-input pricing at $0.15 per million tokens and its 32% per-task cost advantage over Carbon target cost-sensitive agentic workloads where long contexts are reused across calls. The harness-driven threefold swing on SWE-sweep — and the opposite direction of GPT-6.2 Sol's scores — puts pressure on the industry to standardize how coding-agent benchmarks are reported.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.