Overview
Google DeepMind introduced Gemini 4 Neon as a model for complex software engineering, knowledge work, and scientific workflows. Google characterizes it as delivering “frontier-level intelligence” with fast end-to-end task completion, and says it is available through the Gemini API in Google AI Studio, Gemini for business, paid Google AI Pro and Ultra subscribers in Google Antigravity, and Pro and Ultra subscribers in the Gemini app [1].
The announcement positions Neon in two roles: a model that can execute tool-heavy tasks directly, and an “Orchestrator” that directs parallel Hydrogen subagents inside Google Antigravity. Google says the Orchestrator reduces redundant file reads, shares progress, and manages context compaction during long sessions [1]. All capability descriptions and rankings in this page are Google’s claims unless stated otherwise.
Context
Neon arrives as an agent model rather than a standalone chat assistant. Google says it can handle multi-file debugging and refactoring across repositories, synthesize inputs including spreadsheets, PDFs, charts, audio, and video within one context window, and generate structured deliverables [1].
Google’s announcement gives the following figures:
| Measure | Google’s reported result |
|---|---|
| Artificial Analysis Intelligence Index | Third overall, score 61 at xhigh thinking |
| Vals Index | Second, 70.9% |
| Artificial Analysis Coding Agent Index | First overall, 58.6 |
| Zapier’s AutomationBench | First, 56.3% |
| Andon Labs’ Vending-Bench 2 | First, $16,843 |
| Arena’s Agent Arena Index | Second globally, score 1336 across more than 4,200 agent sessions |
| Agent Arena steerability | Second |
| Terminal-Bench 4.1 | 65.2% |
| FrontierCode 2.0 | 46.2% at the default high thinking setting |
| Terminal-Bench Science | 70.5% |
| RiemannBench | 84.0% |
| HLE-Diamond | 71.3% |
| Legal Research Bench | First |
| Tax Agent Bench | First |
| SWE-sweep | 12.8% on the mini-SWE harness; 39.0% in Google Antigravity |
Google also says Neon has a one-million-token context window, four thinking levels that adjust intelligence and token spending, and API pricing of $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens [1].
Why it matters
The clearest practical point is the harness gap. Google reports that Neon leads SWE-sweep on the mini-SWE harness at 12.8%, then more than triples that result to a state-of-the-art 39.0% in Google Antigravity [1]. The source does not say the two setups are directly equivalent, so the figures are not interchangeable. The comparison nevertheless indicates that the execution environment can materially affect a model’s result.
For teams, the practical implication is to compare models inside their own tooling and workflow rather than treat a cross-harness leaderboard position as a prediction of production performance [1].
Google also says that, at its default high thinking level, Neon combines a high Artificial Analysis Intelligence Index score with a 32% lower cost per task than Carbon [1]. The announcement does not state the full benchmark conditions, workload, or calculation behind that comparison, so it does not establish total cost for a particular team.
Opportunities
- Repository-level engineering workflows: Google says Neon handles multi-file debugging and refactoring across repositories with targeted edits. Teams could evaluate workflows that ask an agent to locate a defect, make a constrained patch, and explain the change [1].
- Agent orchestration: Google describes Neon directing parallel Hydrogen subagents in Google Antigravity, with progress sharing and context compaction. This is a vendor-described workflow rather than independently validated performance [1].
- Document-heavy business workflows: The one-million-token window is a design possibility for processing spreadsheets, PDFs, charts, audio, and video together. Google claims Neon can do this, but no workload-specific success rate, latency figure, or total-cost result is provided [1].
- Cost-aware routing: Four thinking levels offer a possible way to send routine work to lower-spend settings and reserve higher settings for difficult analysis. The announcement does not provide the data needed to determine the appropriate split [1].
- Spreadsheet automation: Google’s demonstration produced a 100,000-cell sheet with more than 500,000 live formulas and no embedded images or Apps Script. This suggests a route toward automated financial models and image-processing spreadsheets, but the output would need correctness and formula-maintenance checks [1].
Risks
General operational cautions: The following apply to tool-using agents generally and are not specific findings from Google’s announcement.
- Tool-using agents that read repositories, write files, operate spreadsheets, or call external systems can introduce unauthorized changes and prompt-injection exposure.
- Code refactoring, spreadsheet formulas, and business simulations can produce plausible but incorrect output, so teams need review, testing, and rollback controls.
- Orchestration and demonstrations tied to one platform can make it harder to switch agents, harnesses, or providers.
- Organizations handling legal, tax, financial, or other sensitive material need to assess data retention, regional processing, quotas, contractual protections, and liability on their own.
Announcement-specific gaps: Google’s post does not provide a Gemini API model identifier, quota limits, regional availability details, service-level terms, data-handling terms, or production reliability figures [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Vendor statementGoogle says Gemini 4 Neon reached 39.0% on SWE-sweep in Google Antigravity versus 12.8% on the mini-SWE harness, indicating that results vary with the execution harness.[1]
- Vendor statementGoogle says Neon ranks first overall on the Artificial Analysis Coding Agent Index with a score of 58.6.[1]
- Vendor statementGoogle says Neon has a one-million-token context window and API pricing of $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens.[1]
- Vendor statementGoogle says Neon offers four thinking levels and serves as an Orchestrator directing parallel Hydrogen subagents in Google Antigravity, with progress sharing and context compaction.[1]
- Vendor statementGoogle says Neon places second globally on Arena’s Agent Arena Index with a score of 1336 across more than 4,200 agent sessions, and separately ranks second on steerability.[1]
- Vendor statementGoogle says Neon at its default high thinking level combines a high Artificial Analysis Intelligence Index score with a 32% lower cost per task than Carbon; the announcement does not state the full comparison conditions.[1]
- HypothesisThe SWE-sweep difference suggests teams should evaluate models within their own tooling and workflow rather than rely on cross-harness leaderboard positions.[1]
- Vendor statementGoogle says Neon handles multi-file debugging and refactoring across repositories and reports first place on Legal Research Bench, Tax Agent Bench, Zapier’s AutomationBench at 56.3%, and Andon Labs’ Vending-Bench 2 at $16,843.[1]
- [1]Introducing Gemini 4 Neon●deepmind.googleretrieved Nov 18, 2026
Signals
- Nov 18Introducing Gemini 4 Neon●deepmind.google
Timeline
Analysis
Google reports first place on Legal Research Bench, Tax Agent Bench, Zapier’s AutomationBench (56.3%), Andon Labs’ Vending-Bench 2 ($16,843), and the Artificial Analysis Coding Agent Index (58.6). It reports second place on the Vals Index at 70.9% [1]. On Arena’s Agent Arena Index, Neon places second globally with a score of 1336 across more than 4,200 agent sessions; separately, it ranks second on the index’s steerability measure [1].
For scientific and mathematical reasoning, Google reports 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, and 71.3% on HLE-Diamond. For developer and research workflows, it reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high thinking setting [1].
A reasonable hypothesis is that Neon’s significance lies in combining long context, adjustable reasoning, agent orchestration, and access across several Google products rather than in one leaderboard position [1]. The SWE-sweep contrast supports a second hypothesis: tool design, context management, and the surrounding harness remain central engineering responsibilities even when model capabilities are fixed [1].
Limitations & uncertainties
- The available material does not include the underlying reports for the cited third-party evaluations, not only Google’s own measurements. The rankings and scores therefore cannot be independently checked from it [1].
- The “frontier-level” and “best in class” labels are Google’s characterizations, and the post does not identify their basis [1].
- The source is truncated at “Agent Are.” This cuts off Google’s Pareto-frontier statement about the Agent Arena Index, but it does not affect the separately reported second-place global ranking and score of 1336 [1].
- The announcement does not give the full benchmark conditions, workload, or calculation behind the 32% lower cost-per-task comparison with Carbon, nor does it fully describe the differing SWE-sweep harnesses [1].
- No independent user testimonials, production case studies, latency measurements, production reliability figures, or total-cost results across a common workload are provided [1].
- The creative, spreadsheet, and WebGL examples are Google-produced demonstrations rather than independent user reports [1].
- No Gemini API model identifier, quota limits, regional availability details, service-level terms, or data-handling terms are given [1].
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.