Google's Gemini 4 Neon Targets Long-Running AI Agents at 50% Lower Cost
Google's Gemini 4 Neon brings frontier reasoning to agent workflows at roughly half the cost of larger rivals, headlining a broader shipping spree.
Takeaways
- Gemini 4 Neon launches as the middle tier between Carbon and Hydrogen.
- Priced at $3 per million input tokens and $15 per million output, with a 1M context window.
- Scores 61 on Artificial Analysis Intelligence Index, #1 on the Coding Agent Index.
- Hits 39.0% state-of-the-art on SWE-sweep inside Google Antigravity harness.
- Roughly 50-80% lower cost per task than top competing frontier models.
- Launches with Antigravity, Genie 4, Robotics 2.5, AlphaGenome 2, and Stitch Loop.
Google has introduced Gemini 4 Neon, a model positioned between the larger Gemini 4 Carbon and lighter Gemini 4 Hydrogen. Neon targets long-running agent workflows where reasoning quality, tool use, and cost per completed task all shape production viability.
Neon’s place in the Gemini stack
| Model | Role | Best fit |
|---|---|---|
| Gemini 4 Carbon | Highest-capability tier | Difficult research and long-horizon projects |
| Gemini 4 Neon | Balanced tier | Complex, repetitive agent workflows |
| Gemini 4 Hydrogen | Lightweight tier | Fast, inexpensive everyday execution |
Google describes Neon as the default choice for work that requires sustained reasoning but cannot justify Carbon’s higher cost. That positioning matters for coding agents, research systems, and other workloads that may call a model repeatedly, consume large contexts, and run for minutes or hours.
Benchmarks favor reasoning and steering
At its highest “xhigh” thinking level, Neon scores 61 and ranks third on the Artificial Analysis Intelligence Index. It also ranks second on the Vals Index and sits on its Pareto frontier, meaning no evaluated model delivers both a higher score and a lower cost per task. On the Agent Arena Index, Neon ranks second for steerability, which measures how well a model adjusts when a user changes or corrects instructions during a task.
The published results use several thinking levels and execution environments. The Intelligence Index score uses xhigh, FrontierCode uses Neon’s default high setting, and the Gray Swan security result uses the highest setting. Teams evaluating the model should reproduce their own workloads with the intended thinking level, tools, and token limits.
The production math
| Resource | Price |
|---|---|
| Input tokens | $3 per million |
| Cached input tokens | $0.15 per million |
| Output tokens | $15 per million |
| Context window | 1 million tokens |
Cached input receives a 95% discount, which can reduce costs for agents that repeatedly reuse repositories, specifications, or conversation history. A request containing one million uncached input tokens costs $3 before output charges; the same quantity served from cache costs $0.15.
At the default high thinking level, Google reports that Neon costs 32% less per task than Carbon while retaining a high Artificial Analysis Intelligence Index score. On the Vals and Agent Arena evaluations, it reportedly costs about 50% to 80% less per task than leading frontier competitors. Those figures depend on each benchmark’s token use, thinking level, and harness, so production costs will vary by workload.
Google is rolling out Neon through the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers.
Long agent loops shape the design
Neon reduces redundant file reads, reports progress while working, and supports context compaction during extended sessions. Context compaction summarizes earlier interactions to free space inside the context window, helping an agent preserve relevant decisions without resending every prior message or tool result.
Google’s coding results indicate that Neon performs best when paired with an effective execution harness:
| Evaluation | Score | Environment or setting |
|---|---|---|
| Artificial Analysis Coding Agent Index | 58.6, ranked first | Overall index |
| Terminal-Bench 4.1 | 65.2% | Terminal tasks |
| FrontierCode 2.0 | 46.2% | Default high thinking |
| SWE-sweep | 12.8% | mini-SWE harness |
| SWE-sweep | 39.0% | Google Antigravity |
The 26.2-point SWE-sweep increase between mini-SWE and Antigravity shows how strongly orchestration affects coding-agent performance. A harness controls how the model explores files, invokes tools, verifies changes, recovers from errors, and manages context. Model comparisons therefore need to account for the surrounding agent system.
Google also demonstrated Neon using Antigravity and its /swarm primitive to build a first-person WebGL2 ski simulator from scratch. The project uses real Mont Blanc and Chamonix elevation data, 10,200 lines of custom shaders, 120 Hz physics across six routes, dynamic tracks, slab avalanches, and spatial Web Audio, with no premade assets.
Security claims come with exact rates
On security-focused evaluations, Google reports a 90.5% score on SEC-bench Pro at roughly half the cost per challenge of its comparison baseline. Neon reaches what Google describes as frontier-level ExploitGym performance while using half as many tokens as Carbon. It also scores 51.26% on WebExploitBench v1.2 at half the cost of the competing large models in that test.
Google published additional behavior and alignment measurements:
- Coding integrity: 0.0% cheating on Terminal-Bench 4.0 and a 2.45% shortcut rate on Google’s coding-cheating evaluation.
- Agent safety: a 0.09% misaligned-action rate on the agentic safety suite.
- Honesty: a 99.2% pass rate on the agent honesty evaluation.
- Scope control: under-service, or doing less than requested, fell by 43% to 56% compared with Carbon. Overreach, or acting beyond the request, fell by 15% to 35%.
- Prompt injection: a 0.0% attack-success rate on Gray Swan’s indirect prompt-injection benchmark at the highest thinking level.
These rates describe performance on named test suites under specific settings. Production deployments still require sandboxed tools, narrow permissions, audit logs, approval gates for consequential actions, and application-specific adversarial testing.
Google’s release queue keeps moving
- Gemini 4 Hydrogen targets inexpensive everyday coding and execution.
- Gemini 4 Carbon handles the most demanding research and long-horizon work.
- Genie 4, DeepMind’s interactive world model, is available through the Gemini API and Google AI Studio for the first time.
- Gemini 3.5 Transcribe adds native voice editing for in-place rewrites and translations.
- Gemini Robotics 2.5 updates DeepMind’s vision-language-action model for robotics.
- AlphaGenome 2 adds cell-level analysis for disease research.
- Stitch Loop runs continuous optimization cycles against production data until specified design goals are met.
A practical evaluation plan
- Benchmark Neon with the same repositories, tools, prompts, and approval steps used in production.
- Test high and xhigh thinking levels separately, recording latency, token use, success rate, and retries.
- Measure total cost per completed task alongside input and output token prices.
- Compare the default harness with custom orchestration, especially for file discovery, testing, and error recovery.
- Run long sessions to verify context compaction, instruction retention, and response to mid-task corrections.
- Repeat security tests with the application’s real tools, permissions, and untrusted inputs.
Google’s Gemini 4 lineup organizes model choice around workload economics. Neon’s published results, particularly the SWE-sweep increase from 12.8% to 39.0% under Antigravity, show that the model, thinking level, and harness jointly determine agent performance. Developers considering Neon will need to evaluate that full stack against their own task completion rate, latency, and budget.