Create more. Sell more this Black Friday.Explore
Skip to article
Gemini 4 Neon illustrated by a luminous architectural arch over water
Product Launch

Gemini 4 Neon Launch Guide for Access, Features, and Migration

Google released Gemini 4 Neon on November 18. See where to try it, what benchmarks and context limits mean for your work, and how to migrate an app.

EvoLink TeamProduct Team
10 min read
Google released Gemini 4 Neon on November 18, 2026. If your agent routinely reads across files, traces bugs, writes patches, and runs tests, put Neon on the shortlist. Within Gemini 4, Google positions Hydrogen for frequent everyday work and Carbon for its hardest, longest-running jobs. Neon fits the recurring engineering work in between that still calls for deeper reasoning. Google launch announcement

If you evaluate models through EvoLink, pick a task your team already runs and an outcome you can check. A release-note summary gives you a sense of the model’s writing style. A reproducible production bug or a report with known totals tells you whether to move real traffic.

Gemini 4 Neon specifications at a glance

What to checkGemini 4 Neon
Google model IDgemini-4-neon
Developer accessPaid Gemini API and Google Cloud's Gemini Enterprise Agent Platform
App and coding-tool accessEligible Google AI Pro and Ultra users in Gemini and Antigravity
Gemini API context1,048,576 tokens shared by input, thinking, and output
Gemini API output65,536 tokens for thinking and output combined by default in a synchronous response. Long decode can extend output within the shared context budget
Inputs and outputText, images, audio, video, and PDF input. Text output
Thinking levelslow, medium, high (default), and xhigh
The one-million-token figure is a shared budget. A long input leaves less room for thinking and the response. If your app sends a large archive and expects a long report back, budget both sides and verify that your integration supports continuation. Google model specifications · Google Cloud release notes

Where Gemini 4 Neon is available

The same model ID, gemini-4-neon, can come with different context limits, built-in tools, and billing rules. Pick the access route that matches the workflow you want to test.
SurfaceWhat you can tryWhat to verify
Google AI Studio and Developer APITry the model in AI Studio. The paid Developer API exposes Neon through Interactions and GenerateContentConfirm project billing and test the exact API method your application uses
Google CloudGoogle documents the global endpoint and the us and eu multi-regionsMatch the project's location and deployment requirements
Gemini app and AntigravityEligible Pro and Ultra users can try Neon. Antigravity documents up to 512K contextCheck your plan and measure quota in that product separately from API spend
Cursor and third-party gatewaysCursor lists 200K context with a 1M maximum. OpenRouter and Vercel list google/gemini-4-neonCheck tool support, enabled context, and billing through that specific route
An IDE’s repo search, file selection, and context compaction change the input the model sees. If two teams compare an Antigravity run and a raw API request, record which files and tools each run received before interpreting the results. Try Neon in AI Studio · Google Developer API · Google Cloud locations · Antigravity limits · Cursor model reference · OpenRouter · Vercel AI Gateway

Which tasks to try first with Neon

If your team needs to…Start the evaluation hereWhat a good result looks like
Run routine edits or extract known fields at volumeHydrogenValid output with few retries
Fix a recurring bug that crosses modulesNeon highThe failing case passes, existing behavior holds, and the patch stays in scope
Reconcile a report across tables, charts, and PDFsNeon highImportant figures trace back to their source, and missing evidence is flagged
Resolve an ambiguous architecture or debugging problemCarbon as a baselineThe team can reproduce the diagnosis and validate the proposed change

Suppose the PDF summary of a monthly report disagrees with the spreadsheet. Give Neon the source files, reporting period, and fields to reconcile. Ask for a worksheet name, cell range, or page reference for every discrepancy. Include a known mismatch and a figure with no traceable source. A passing result finds the mismatch and flags the unsupported figure.

That kind of test matters even when benchmark results are strong. Plotly's data-analysis evaluation gave both Neon and Hydrogen 43 out of 43, with Hydrogen costing less on that test. If your existing model already clears the acceptance bar, the launch gives you another option for the cases it struggles with. Plotly evaluation

What the launch demos and third-party tests show

Google demonstrated a WebGL skiing simulator and an interactive spreadsheet built from a static image using formulas. Use those examples to plan checks for your own Antigravity work. Reload a web prototype and exercise its key controls. In a spreadsheet, trace a few summary cells to their inputs, then change a source value and check that the dependent formulas recalculate. Google launch demos
Google demo of Neon’s WebGL ski simulator in Antigravity
Google demo of Neon turning a static image into an interactive spreadsheet with formulas
For an agent that executes business workflows, Zapier offers a different test. Neon passed 370 of 657 tasks (56.3%) at xhigh in AutomationBench. Each case checked the final state of simulated apps. A support workflow can apply that idea by verifying the ticket's final owner and status, checking that no duplicate ticket was created, and requiring the agent to surface an unresolved account match. A plausible completion message alone leaves these failures invisible. Zapier AutomationBench report
Zapier AutomationBench scored Gemini 4 Neon at xhigh on the final state of 657 tasks. It passed 370, or 56.3%.
Zapier AutomationBench scored Gemini 4 Neon at xhigh on the final state of 657 tasks. It passed 370, or 56.3%.

When to use Orchestrator in Antigravity

In Orchestrator (Neon), Neon plans and reviews while Hydrogen implements and runs checks. If an unexpected interface or a failed check invalidates the plan, control returns to Neon. This is a useful option to test when discovery is difficult and most edits are straightforward, such as a package migration with a few compatibility decisions. Try direct Neon as the baseline when the difficult reasoning continues throughout implementation. The Orchestrator setting has a fixed configuration, while direct model selection exposes the thinking level. Antigravity model selection

How much to budget for the first test

Google's standard paid API lists Neon at $3 per million input tokens, $15 per million output tokens, and $0.15 per million cached input tokens. Output pricing includes thinking tokens. The Google Developer API uses implicit caching. Treat cache hits as a saving you can measure after the first run. Google API pricing
For example, a task with 60,000 uncached input tokens, 20,000 cached input tokens, and 8,000 billable output tokens would cost $0.303 at those rates before paid tools or retries. The billable output includes thinking. If the agent makes another attempt, count that call toward the same task.
On November 19, 2026, EvoLink’s public pricing feed listed default-group rates of $2.85 input, $14.25 output, and approximately $0.143 cached input per million tokens. Check the Gemini catalog for availability and confirm your account rate before routing a workload. If you already use EvoLink's Gemini integration, keep the same inputs and acceptance checks when you evaluate another route. EvoLink pricing feed · Gemini collection · Gemini integration guide
Read the Neon and Carbon comparison for a same-family routing decision or the Neon and Opus 5.5 comparison for a coding-agent evaluation.

How to migrate an existing Gemini app to Neon

Start with one repeatable task. For a Gemini 4 integration, Google documents the same request schema for a switch from Carbon, Argon, or Hydrogen. Pin the thinking level during comparison. Carbon and Argon default to medium, while Neon defaults to high, so leaving both at their defaults also changes the amount of reasoning the models are allowed to use. Google migration checklist

Check three areas when upgrading from Gemini 3.

  • Caching. Replace explicit cache creation and references such as cached_content with a flow that works with Neon's implicit caching.
  • Multi-turn state and thought signatures. If the client manages GenerateContent history, return the signatures Google sent without changing them. Follow the Interactions API's state contract when using that interface.
  • Tools and structured output. Conform to the API schema for function tools and structured output. Validate the follow-up turn after a tool returns its result. Google migration checklist

Make long generation resumable

If all inputs total 300,000 tokens, the shared window leaves 748,576 tokens before thinking and output. A standard synchronous response still caps thinking and output together at 65,536 tokens by default. For a longer generation, Google documents background execution through Interactions or client-managed continuation on a supported interface. Test a stopped run with a document that has an obvious last section. Confirm that the app checks final status, handles partial output, and resumes without repeating sections or creating a second deliverable. Google Long Decode guide

Pick a service tier for your deadline

Google's paid Developer API also lists Batch and Flex at $1.50 input, $7.50 output, and $0.075 cached input per million tokens. Priority lists $5.40, $27, and $0.27 on the same basis. Use the tier's documented delivery behavior to choose a route. An overnight review queue and an interactive coding request have different time budgets. Compare the whole run and confirm that the billed tier matches the request. Google Neon service-tier pricing

FAQ

When did Gemini 4 Neon launch?

Google announced it on November 18, 2026.

Is it free through the Gemini API?

Google lists Neon on the paid API tier. App and IDE access follow separate subscription rules.

What is the Google API model ID?

gemini-4-neon.

Can I send a video or PDF?

Google lists both as supported inputs. Neon returns text.

Can it output a million tokens in one request?

Thinking and output share the default synchronous limit of 65,536 tokens. Use long decode for longer generation and keep the total within the shared context budget.

Which thinking level should I try first?

Start with high for a complex task. Try medium when the task has clear checks, or test xhigh on cases that stall at high.

Should I replace Hydrogen or Carbon?

Evaluate Neon on a task class you can verify. Keep Hydrogen or Carbon wherever it already meets your quality, time, and cost requirements.

Does a Gemini subscription include API usage?

Google offers Gemini app and Antigravity subscriptions separately from paid developer API access. Google launch availability

Why does Antigravity show a smaller context window?

Google documents up to 512K in Antigravity and a shared 1,048,576-token window in the Developer API. Size your prompt for the surface running it. Antigravity model limits

Can I reuse an explicit Gemini cache?

Neon accepts implicit cache hits. Revise code that explicitly creates or references a cache object, then confirm cached usage in the billing data. Google migration checklist

Does Neon make sense for overnight jobs?

Google offers lower Batch and Flex rates for Neon. For a job with a firm deadline, test completion rates and queue times on your own data. Google pricing

Connect AI models through one API

Compare models and current pricing, then follow the API documentation to integrate your chosen model.