Google DeepMind's model card for Gemini 4 Neon reports the model ahead of Claude Opus 5.5 on 14 of 16 shared benchmark rows and behind on two coding tests.
The results are dated November 2026 and are Google's own evaluations, not independent tests. For companies weighing coding and enterprise-workflow models, the card is a vendor's side-by-side claim, and it points to a methodology page1 that this report did not review, so the evaluation conditions are not examined here.
The two exceptions are both in agentic coding, according to the model card1. Neon scored 65.2% on Terminal-bench 4.1 against 65.8% for Claude Opus 5.5, a gap of 0.6 percentage point. The card labels that row "cheating adjusted accuracy." On FrontierCode 2.0, Neon scored 46.2% against 47.3%. On CursorBench 4.1, Neon led, 65.4% to 58.1%.
The card lists SWE-sweep under two harnesses, and the scores differ sharply. Under the provider harness, Neon scored 39.0%, Claude Opus 5.5 10.1% and GPT-6.2 Sol 8.3%. Under the Mini-SWE harness, Neon scored 12.8%, Claude Opus 5.5 3.6% and GPT-6.2 Sol 11.6%, which puts GPT-6.2 Sol within 1.2 points of Neon. The same benchmark produced very different numbers for Neon (39.0% versus 12.8%) depending on the harness.
The other rows on which Neon leads Claude Opus 5.5 are:
- AA Index v5.0.1: 61 to 59
- Vals Index v2.2: 70.9% to 70.1%
- AutomationBench: 56.3% to 42.5%
- Agents' Last Exam v1.1: 75.8% to 71.6%
- RiemannBench: 84.0% to 80.0%
- Terminal-Bench-Science 0.1: 70.5% to 63.3%
- HLE-Diamond: 71.3% to 67.9%
- MKAClimb WSR: 56.5% to 53.0%
- GraphWalks (256k to 1M, BFS, F1): 95.2% to 66.8%
- CWE-Bench v2: 70.7% to 67.1%
- WebExploitBench v1.2: 51.3% to 37.1%
The margins vary widely. The Vals Index lead is 0.8 point, while the GraphWalks lead is 28.4 points. The rows cover different tasks and scoring systems, so the 14-of-16 count is not a single measure of model quality. GPT-6.2 Sol has no score on two rows, CursorBench 4.1 and WebExploitBench v1.2.
The card says Neon is adapted from Gemini 4 Carbon and is built for complex coding, analytical reasoning and production-grade enterprise workflows. Google describes it as balancing depth and efficiency, with a 1M-token context window. The card says it is available to downstream providers via API, with access varying by channel and subscription tier, including paid Google AI Pro and Ultra subscribers in Google Antigravity and Google AI Pro and Ultra subscribers in the Gemini app. It lists no pricing.
The card says the full Safety & Security Report for Neon "will be available in the coming days." Its safety section carries forward capability determinations from Gemini 4 Argon rather than reporting new critical-capability tests, which Google says were not required because Neon's capability increase fell below its threshold.
Separately, Demis Hassabis on Nov. 20 reposted a user's claim2 that "the new Google Gemini models are the best AI models out there right now." The post does not name Neon or cite tests. Automatica reported Nov. 193 on an earlier Hassabis repost about a user's claim that Neon beat Fable 5.2, and noted the benchmark claim was not verified.