Gemini 4 Neon vs. Claude and GPT-6.2: What Does More Thinking Buy?
Gemini 4 Neon goes from $0.75 to $6.44 per benchmark task. Compare all four thinking levels with Claude Opus 5.5, Fable 5.2 and GPT-6.2 Astra.
Gemini 4 Neon is Google's new reasoning model for jobs like updating a codebase, reading a stack of documents or carrying out a workflow across apps. Google released it on November 18, 2026. It takes text, images, audio and video as input and returns text. Google puts it between Hydrogen for quick everyday work and Carbon for its hardest long-running projects. Google's launch announcement
Is Gemini 4 Neon worth using? For automating work across apps, I'd start with Neon on medium. It completed 350 of Zapier's 657 tasks at an estimated $0.58 per attempt, ahead of GPT-6.2 Astra on max. For a mix of reasoning, coding and knowledge work, I'd try Astra first. On the broader Artificial Analysis Intelligence Index, Astra on high scored 60.1 for $2.07 per task. Neon on high scored 59.6 for $3.74. The job you hand over changes which model earns its price. Zapier's workflow results · Artificial Analysis model data
What does Gemini 4 Neon cost, and where can you use it?#
The standard Gemini API price is $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. For example, reading 100,000 fresh tokens and producing 2,000 billable tokens costs $0.33. That example assumes the 2,000 include all reasoning and answer tokens. A short visible answer can still come with a long reasoning bill. Google's rates · Artificial Analysis cost breakdown
You can select Neon in the Gemini app with Google AI Pro or Ultra. Developers can use the paid Gemini API through AI Studio under the model ID gemini-4-neon. The API supports a 1 million-token context window. Google lists up to 512,000 tokens in Antigravity, where the optional Orchestrator lets Neon plan and review work that Hydrogen carries out. That distinction matters if you're choosing where to open a large project. Google's launch and access details · Antigravity team's rollout notes
Gemini 4 Neon thinking levels and cost per task#
Artificial Analysis runs the same Intelligence Index across model settings. A point is a point on its composite score. A dollar figure is the weighted average API cost of attempting one Index task, including input, caching, reasoning and answer tokens. The figures below are a November 19 snapshot. Artificial Analysis results and methodology
| Neon effort | Intelligence Index | Cost per task | Minutes per task |
|---|---|---|---|
| Low | 44.1 | $0.75 | 3.7 |
| Medium | 54.9 | $1.90 | 8.3 |
| High (default) | 59.6 | $3.74 | 14.8 |
| Xhigh | 61.4 | $6.44 | 21.5 |
The biggest jump comes early. Moving from low to medium buys about 10.8 Index points for an extra $1.16 per task. Moving from high to xhigh buys 1.8 points for an extra $2.70 and takes about seven minutes longer. Those differences use the underlying figures before rounding. I'd use high for a difficult first draft and save xhigh for a specific failure you can check, such as a change that still fails its acceptance tests. Neon's per-effort measurements
Gemini 4 Neon vs. Claude Opus 5.5, GPT-6.2 Astra and Claude Fable 5.2#
At each model's highest tested setting, Claude Fable 5.2 leads on intelligence, GPT-6.2 Astra comes closest for $4, and Neon beats Claude Opus 5.5 on both score and cost. GPT-6.2 Sol is worth including if you run a job hundreds of times. It scored 57.3 for about a dollar per attempt.
| Model at top effort | Input / output per 1M tokens | Index score | Cost per Index task |
|---|---|---|---|
| Claude Fable 5.2, max | $8 / $40 | 64.6 | $26.54 |
| GPT-6.2 Astra, max | $8 / $40 | 62.1 | $4.00 |
| Gemini 4 Neon, xhigh | $3 / $15 | 61.4 | $6.44 |
| Claude Opus 5.5, max | $4 / $20 | 59.1 | $10.04 |
| GPT-6.2 Sol, max | $2 / $10 | 57.3 | $1.01 |
Source figures are from Artificial Analysis. The Claude runs use Anthropic's default fallback configuration. Prices are the per-token rates used in this comparison. Fable 5.2 · Astra · Neon · Opus 5.5 · Sol
Why can Astra cost less when each output token costs more? Neon xhigh generated 283 million output tokens across the Index. Astra max generated 65.8 million. Input and cache charges also count, but that difference in output helps explain how Astra can charge $40 per million and still finish the average task for $2.44 less. Neon token use · Astra token use
- GPT-6.2 Astra62.1 points · max · $4.00
- Gemini 4 Neon59.6 points · high · $3.74
- GPT-6.2 Sol57.3 points · max · $1.01
- Claude Fable 5.256.4 points · medium · $3.80
- Claude Opus 5.551.8 points · high · $3.46
The budget selector takes each model's highest measured score under the selected average cost. At $1, Sol on xhigh reaches 55.7. At $4, Astra on max reaches 62.1. Neon on high also fits under $4 and edges out Opus on max, which costs $10.04. You can inspect every setting in the downloadable data for all 24 model and effort combinations.
Which benchmarks favor Neon?#
Automating work across apps. Zapier checks the final state of a simulated company after each task. All required changes have to be correct to pass. On the same 657 tasks, Neon medium passed 53.3%, ahead of Astra max at 52.4% and Fable max at 51.4%. Neon high passed 54.3% for $0.88 an attempt. Xhigh passed 56.3% for $1.69. That's 13 extra completions over high in 657 attempts. Using Zapier's rounded costs, repeating the whole set at xhigh would cost about $532 more, or roughly $41 per additional completion. That calculation includes failed attempts and uses Zapier's cache-price assumption. Zapier AutomationBench report
Finance, coding, legal and tax work. Vals weights these fields by their share of US GDP. Fable led at 71.31% for $17.94 per test. Neon reached 70.87% for $7.40, with Opus at 70.14% for $24.77 and Astra at 67.66% for $10.68. If your work resembles this mix, Neon closed to within half a percentage point of Fable for about 41% of its cost. Vals Index v2.2 leaderboard
Coding with an agent. Artificial Analysis measured Neon xhigh in Antigravity CLI at 59 on its Coding Agent Index for $6.86 per task. Claude Code with Opus max or Fable xhigh scored 58, at $12.61 and $15.47 respectively. Here you're choosing a model and the software that lets it read files, edit code and run checks. On individual coding tests the order can flip. Google's evaluation reports 46.2% for Neon and 47.3% for Opus on FrontierCode 2.0. Artificial Analysis coding agent results · Google's evaluation details
What early users are saying about Gemini 4 Neon#
Developer Flavio Copes gave Neon and Opus the same CSV parser task. Both passed all 27 checks. At high effort, Neon took 49 seconds and cost 9.1 cents. Opus took 12 seconds and cost 2.9 cents. Then he asked both to find two posts in the same 385 KB of Markdown. Both found the right pair. This time Neon cost 33 cents to Opus's 57 cents. Reading a lot and returning a little is a different job from thinking through parser edge cases. Copes's prompts, timings and costs
Plotly checked all 43 answers in its data analysis exam by hand after correcting a grading error. Neon earned a perfect score for $17.03. Hydrogen also got all 43 for $6.36. Plotly liked how Neon explained ambiguities, though it found Astra more concise. For an analyst who wants a number and one sentence for a slide, that difference may matter more than the tie. Plotly's results and example answers
In the Antigravity community, Reddit user phpHater0 praised the readability of Neon's work after two hours on a project. In the same thread, Duck_nine reported that Neon Orchestrator used their five-hour allowance before it finished checking a project. These are individual accounts from launch week. Subscription allowances and metered API costs answer different questions. First-hand Antigravity discussion
Which model would I start with?#
If you're automating a repeatable business process, try Neon medium on a handful of finished jobs where you know the right outcome. Count how many need fixing and what each fix costs. Zapier's results put a real price on the decision to move up to xhigh.
For mixed technical work with a firm API budget, Astra high is the first comparison I'd run. For short functions with clear tests, keep Opus 5.5 in the trial. For a difficult task where one better answer would repay a much larger bill, Fable 5.2 has the highest overall score here. And if a cheaper model already gets every answer right, as Hydrogen did on Plotly's exam, spend the difference on checking the output.
For more detail on the alternatives, see my GPT-6.2 Astra and Sol benchmark comparison, Claude Fable 5.2 vs. Opus 5.5 effort breakdown and Gemini 4 Hydrogen price and thinking-level guide.
Moe shares tool walkthroughs and lessons from real projects. Mechanical engineer, then venture capital, now building AI tools for creators and small businesses. More about Moe