AI Agents, Standard

Gemini 4 Neon brings complex agent workflows to Copilot and Google’s AI tools

Symbolic diagram: Gemini 4 Neon brings complex agent workflows to Copilot and Google’s AI tools

Introduction to Gemini 4 Neon

Gemini 4 Neon is Google’s new model for complex coding, analytical work, and enterprise workflows, introduced on November 18, 2026. It is also generally available in GitHub Copilot, with access rolling out gradually. The central proposition is practical: Google says Neon can complete demanding tasks with intelligence approaching Gemini 4 Carbon while reducing the time and token cost involved.

That positioning matters for work that extends beyond a single answer. Google’s announcement describes a model built for long-running agents, multistep tool use, and following user corrections during a task. In software engineering, the intended result is more focused changes across repositories. In knowledge work, it is the ability to turn varied source material into structured deliverables. These are Google’s descriptions of the model’s capabilities; the launch material does not establish a measured productivity gain for every organization adopting it.

Google places Neon between Carbon, which it positions for its hardest long-horizon projects, and Hydrogen, which handles fast everyday execution. Neon is intended to balance reasoning depth and efficiency for complex, frequently repeated work. Sundar Pichai reinforced that positioning in his launch post, highlighting software engineering, knowledge work, and science alongside the rollout across Google’s products.

The evidence supports examining Neon as a combination of model capability, operating cost, and agent behavior. Its benchmark results are substantial, but they need to be read with their settings and execution environments intact. A score from an agent harness, for example, cannot by itself explain how the model will perform inside a different development tool.

Capabilities and Features

Symbolic diagram illustrating: Capabilities and Features

Reasoning, coding, and following corrections

In its Neon announcement, Google reports a score of 61 and third place overall on the Artificial Analysis Intelligence Index at the model’s xhigh thinking level. It also reports second place on the Vals Index, with 70.9%, and first place on both the Legal Research Bench and Tax Agent Bench. Although Google describes several of these evaluations as independent, the supplied results are reported through Google’s announcement rather than separate evaluation reports.

For coding, Google reports first place on the Artificial Analysis Coding Agent Index with 58.6, alongside 65.2% on Terminal-Bench 4.1. Its FrontierCode 2.0 result is 46.2% at the default high setting. Keeping that distinction visible matters: the headline intelligence result uses xhigh, while the cited FrontierCode result uses high. They are not all measurements of one fixed operating configuration.

Google also reports second place for steerability on the new Agent Arena Index, defining that capability as adapting to user corrections during a task. Separately, it reports second place overall on that index with a score of 1336 across more than 4,200 real-world agent sessions. The steerability ranking and overall ranking describe different aspects of the evaluation, even though both are second place.

For an agent handling a changing assignment, steerability is relevant because completion alone is insufficient if the model continues pursuing instructions the user has revised. Google’s claim suggests Neon is designed to accommodate that interaction. The supplied material does not give the detailed steerability methodology, so the ranking cannot establish how reliably it will follow every kind of correction.

Long sessions and mixed inputs

Neon has a 1 million-token context window. Google says it can synthesize spreadsheets, PDFs, charts, audio, and video within that window and produce structured outputs. The announcement reports 73.5% on Chartography for reading dense professional charts and 75.8% on Agents’ Last Exam v1.1. Those results address particular evaluations; they do not establish uniformly accurate interpretation of every uploaded document.

For development work, Google describes sustained performance through complex, tool-heavy tasks, fewer redundant file reads, progress updates, and reliable context compaction during long sessions. In Google Antigravity, Neon can also act as an Orchestrator directing parallel Hydrogen subagents. That is a documented Antigravity workflow in the announcement, rather than evidence that every product offering Neon exposes the same orchestration controls.

The execution environment makes a visible difference in Google’s own coding results. On SWE-sweep, the announcement reports 12.8% using the mini-SWE harness and 39.0% in Google Antigravity. Google attributes the increase to what the Antigravity harness enables. The comparison supports treating the surrounding agent system as consequential; it does not support transferring the higher result automatically to Copilot or a custom API integration.

Thinking levels, prices, and cost claims

Google says Neon offers four thinking levels, with the default high setting aimed at balancing capability and spend, and xhigh intended for deeper multistep reasoning. Its published token prices are $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens. Cached input is listed at a 95% discount, but that rate should not be treated as the price for all input.

At high, Google reports a 32% lower cost per task than Carbon alongside a high Artificial Analysis Intelligence Index score. It also claims roughly 50% to 80% lower cost per task than top competing frontier models on the Vals Index and Agent Arena, and 36%+ lower cost per task than competing models at their highest setting when Neon uses xhigh.

These are benchmark-specific vendor claims. The supplied material does not provide enough underlying task and pricing detail to turn those percentages into a forecast for a particular team’s bill. Token prices and cost per completed task answer different questions: the former gives a unit rate, while the latter depends on how much work a task consumes and whether its output is usable.

Availability and Rollout

Symbolic diagram illustrating: Availability and Rollout

The GitHub Copilot changelog confirms general availability for Copilot Pro+, Max, Business, and Enterprise users. GitHub says the rollout is gradual, so general availability does not mean every eligible account will immediately see the model.

GitHub lists the model picker in Visual Studio Code, Visual Studio, Copilot CLI, GitHub Copilot coding agent, GitHub Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode, and Eclipse. This provides several routes into existing development environments, although the announcement does not establish identical behavior across those surfaces.

Business and Enterprise administrators can manage Neon through the model policy in Copilot settings. Under default model enablement, new models are automatically enabled unless an administrator has disabled the global default or specifically disabled the model. For an eligible organization, visibility can therefore depend on both rollout status and administrator policy.

GitHub says Neon is billed at provider list pricing under usage-based billing. Its announcement supplies no separate numerical Copilot rate table. Google’s published token prices provide launch pricing context, while GitHub’s own billing description governs the claim made about Copilot availability.

Google’s rollout includes the Gemini API paid tier in Google AI Studio and Android Studio, Gemini for business through the API or model dropdown, and Google Antigravity for paid Google AI Pro and Ultra subscribers. Google also says Pro and Ultra subscribers can try Neon in AI Studio and select it in the Gemini app. These are distinct access routes with subscription or paid-tier prerequisites, rather than a statement of unrestricted access.

A specialized Gemini 4 Neon for cyber has a separate rollout. Google says eligible Fairwind Program members will begin receiving access over the following days. The announcement directs prospective participants to their Cloud account team. Its cyber evaluation results should consequently be understood alongside that restricted access path, rather than assumed to describe a capability available to every Neon user at launch.

Impact on Software Engineering and Knowledge Work

More useful autonomy depends on scope control

Neon’s most relevant software engineering claims concern how it behaves while completing a job. GitHub, attributing the comparisons to Google, says it starts responding faster than Carbon, makes fewer unrequested changes on coding tasks, and is less likely to make unsupported claims about code, tool results, or tests. Those behaviors matter because the value of an agent’s output depends partly on how much verification and correction it requires afterward.

Google backs its behavioral positioning with internal evaluations. It reports a 99.2% pass rate on an agent honesty evaluation and a 0.09% misaligned action rate on an agentic safety suite. It also reports reductions in under-service of 43%–56% and overreach of 15%–35% against Carbon across different effort levels. These figures describe Google’s evaluations; the supplied evidence does not include their full protocols or establish equivalent rates in a customer’s repository.

That distinction leaves a concrete implication for adoption: the relevant question is whether Neon completes the intended change while preserving the requested scope. Its coding scores and reported behavioral improvements make that a reasonable focus for evaluation, but they do not remove the need to inspect the resulting changes and verify test claims. A faster initial response is also a different measure from faster completion of a correct, accepted task.

Google’s security claims require similar care. It reports that Neon, with real-time safeguards at its highest thinking level, matched a 0.0% attack success rate on Gray Swan’s indirect prompt injection benchmark. That result is conditional on the benchmark and safeguards described. It does not establish immunity to malicious instructions in every webpage, document, email, or codebase the model may encounter.

Knowledge work spans interpretation and delivery

For knowledge workers, Google’s case combines mixed-input synthesis with reasoning and structured output. In the Gemini app, the company describes more precise reading of spreadsheets, PDFs, and photos, reports and summaries from dense material, and incorporation of relevant information from Gmail and Docs. It also says Tasks in Gemini can run on Neon for complex multistep jobs.

The announcement’s examples include analyzing billing statements to investigate a utility charge and cross-checking receipts and documents for a refund negotiation. These illustrate Google’s intended uses. They are not independently verified demonstrations of savings or successful outcomes.

Broader workflow evaluations add another perspective. Google reports first place on Zapier’s AutomationBench with 56.3% and on Andon Labs’ Vending-Bench 2 with $16,843 in a long-horizon business simulation. The latter is a simulation result, not customer revenue. Neither score should be translated directly into a workplace productivity percentage.

Scientific work is part of the same positioning. Google reports 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, and 71.3% on HLE-Diamond. These results support the company’s emphasis on technical reasoning, but the package supplies no detailed scientific deployment study. The defensible interpretation is that Neon has reported strengths across several technical evaluations, while its effect on a particular research workflow remains an open question.

For both developers and knowledge workers, the strongest practical case is therefore conditional: if the reported reasoning, correction-following, and scope discipline carry over to the user’s tasks, Neon could reduce the effort needed to reach an acceptable deliverable. The launch evidence establishes a reason to test that proposition, rather than a universal productivity outcome.

Conclusion

Gemini 4 Neon arrives with a clear role: complex, repeated work that calls for substantial reasoning while keeping time and token costs in view. GitHub confirms its Copilot availability, and Google provides a broader rollout across developer, enterprise, and consumer products. The reported benchmarks and behavioral evaluations explain why it merits attention for agentic workflows.

The unresolved issue is how those results transfer into specific working environments. Google’s own SWE-sweep comparison demonstrates that the harness can materially change outcomes, while its thinking-level results show that configuration affects the capability and cost comparison. Neon’s practical value will depend on that combination of model, execution environment, and task requirements. The evidence supports a promising option for demanding work; demonstrated gains in a particular team still require evaluation in that team’s workflow.