Google introduces Gemini 4 Neon, ranked #3 overall at its xhigh level
Google presents Neon as a model for complex software, knowledge and science work, positioned between its faster Hydrogen and more powerful Carbon models. Its launch pairs benchmark claims and lower task costs than Carbon with a general release across Google products and a more restricted route for its cyber version.
At xhigh thinking, Neon scored 61 on the Artificial Analysis Intelligence Index and ranked third overall.2
Google reports a 56.3% result and first place overall for Neon on Zapier’s AutomationBench.2
At the default high thinking level, Google says Neon’s cost per task is 32% lower than Carbon’s.2
Story
What Google says Gemini 4 Neon changesGoogle has introduced Gemini 4 Neon for work spanning software engineering, knowledge tasks and scientific workflows, describing it as a model intended for complex processes rather than a single narrow task.2 Its announced position is between Hydrogen, which Google characterizes as faster, and Carbon, which it calls more powerful; the company associates Neon with demanding workloads that occur frequently.1 At the xhigh thinking setting, Neon received 61 on the Artificial Analysis Intelligence Index and placed third overall, one indication of the performance position Google presents for the model.2 Google also reports that Neon leads the Legal Research Bench and Tax Agent Bench, giving its launch case examples from legal and tax work alongside the broader claims about workflows.2
The rankings vary by evaluation: Neon is second on the Vals Index and also second for steerability on the Agent Arena Index, rather than occupying the same position across every measure.2 For automation, Google reports a 56.3% result on Zapier’s AutomationBench and first place overall, a different result and ranking from its placement on the Artificial Analysis Intelligence Index.2 On the Artificial Analysis Coding Agent Index, the company says Neon is first overall with a score of 58.6, while its Terminal-Bench 4.1 result is 65.2% and its FrontierCode 2.0 result is 46.2% at the default high setting.2 These results refer to distinct named benchmarks and, where specified, different settings, so they do not amount to a single interchangeable score for all software or knowledge work.2
The science-related results Google lists include 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond.2 For SWE-sweep, Google reports two results using different environments: 12.8% on the mini-SWE harness and 39.0% in Google Antigravity.2 Because the reported SWE-sweep figures come from separate harnesses, the figures describe results in those respective setups rather than a direct comparison under one shared setup.2 Google says Neon can coordinate parallel Hydrogen subagents inside Antigravity, assigning the model an orchestration role in which it directs other agents rather than only completing work as a standalone model.2
Neon’s context window is 1 million tokens, and Google says it can combine spreadsheets, PDFs, charts, audio and video into structured outputs.2 That capability description covers several input formats and a structured deliverable, while the stated context-window figure specifies the amount of context the model can handle.2 Google says there are four thinking levels, allowing teams to adjust the balance of intelligence and token expenditure to the task they are running.2 The levels make workload adjustment part of the product description, but the launch information does not give a separate price for each level.2
Google lists input pricing at $3 per million tokens and output pricing at $15 per million tokens; cached input is $0.15 per million tokens, which the company describes as a 95% discount.2 Those prices distinguish tokens sent to the model, tokens it generates and input tokens served from cache, rather than applying one rate to all token use.2 At the default high thinking setting, Google says Neon’s cost per task is 32% below Carbon’s, a comparison about task cost rather than a claim that every request has that same price difference.2 Google’s stated market case therefore combines high benchmark placements with a lower per-task cost than Carbon, although the cost comparison is specifically tied to the default high setting.2
Google says Neon produced its lowest fabrication rate so far in internal testing with Google Antigravity, a claim limited to that company testing context.2 The company also reports that Neon refused 1.98% of tasks on Terminal-Bench 4.0 and refused none on Terminal-Bench-Science.2 These refusal figures are attached to two named evaluations, while the fabrication statement refers to internal Antigravity testing; they are different measures and settings.2 Google says Neon retains the protections against severe misuse and prompt injection that it associates with Gemini 4 Carbon.2
The general model was being rolled out through the Gemini API in Google AI Studio, Gemini for business, Google AI Pro and Ultra in Antigravity, and Pro and Ultra subscriptions in the Gemini app.2 That distribution spans developer access, business use, Antigravity plans and Gemini app subscriptions, according to Google’s launch information.2 Neon for cyber follows a narrower initial route: Google scheduled it to reach trusted defenders in the coming days through the Fairwind Program.2 The launch thus pairs broad channels for the general model with restricted initial access for the cyber version, as Google describes the two rollouts.2
For its cyber version, Google reports a 76% score on the hard tier of AutoBaxBench, while the stated plan is to make that version available first to trusted defenders through the Fairwind Program.2 That score concerns Neon for cyber and a specified hard tier, not the general model’s complete set of benchmark results.2 Google’s launch framing consequently separates a capability result for the cyber model from its access plan, which begins with a selected defender group.2 Across the launch, Neon is presented as a mid-positioned model with several benchmark leads, flexible thinking settings, stated token prices and a mix of general and restricted distribution.12
Impact
Spreading outward, level by level- Level 1Benchmark position
Google reports Neon’s score and overall rank on the Artificial Analysis Intelligence Index, as well as first place on the Artificial Analysis Coding Agent Index.2
Fact - Level 2Workflows and orchestration
Neon is described for complex software engineering, knowledge work and science workflows, and can direct parallel Hydrogen subagents in Google Antigravity.2
Fact - Level 3Costs and use
Google lists input, output and cached-input token prices, and says Neon’s cost per task is lower than Carbon’s at the default high thinking setting.2
Fact - Level 4Access
Google describes broad rollout channels for the general model and a restricted initial channel for Neon for cyber.2
Fact
Sources
What each source supportsWritten by AI from the sources listed below: every fact was checked word for word against its source, and inference is marked apart. How we write