Gemini 4 Neon and the Cost of Extra Reasoning

The second model call may add cents to an API bill and leave an engineer waiting several minutes for a patch review. Google’s Gemini 4 Neon makes it possible to choose how much reasoning to spend on a request. For a team reviewing a patch or reconciling a messy document, the useful question is which setting earns its place in the work.
What Google Released on November 18
Google announced Gemini 4 Neon on November 18, 2026, as a Gemini 4 model for recurring work that requires substantial reasoning. Neon occupies the middle position between Hydrogen and Carbon in the Gemini 4 lineup. The launch announcement describes a rollout in the Gemini API and Google’s workplace Gemini offering. Google AI Pro and Ultra subscribers were included in the Gemini app and Antigravity rollouts. Google’s developer guide for Neon lists paid access to the Gemini API and specifies Pro or Ultra for use in AI Studio.
The model specifications confirm support for PDFs and images on the same endpoint teams use for text. Teams can submit video and audio as inputs. Neon returns its answers as text. Native Google requests share a 1,048,576-token window across input and generation. The same allowance covers the model’s internal reasoning. Synchronous requests have a 65,536-token output limit. Much longer output requires background execution or client-side continuation, subject to the space left in the same context window. Account access and output limits deserve a check on the actual route an application will use.
Engineers migrating an existing Gemini application have a specific reason to check the output limit: a small output cap can cut a response short during reasoning. Google advises setting the desired effort separately from the output ceiling. In a pilot, record incomplete responses as their own outcome and inspect the usage record. The bill may appear low for a short response, even though the operator still has to rerun the job.

High Is Google’s Default
Google’s thinking guide names High as Neon’s native default when the application leaves the effort setting unspecified. Developers can request Xhigh for work demanding the extra reasoning. Lower settings are available for jobs with tighter timing or token budgets. The Google API default describes Google’s own entry point; an application using a gateway should set the effort on purpose and confirm the selected endpoint accepts it. Provider routing determines which endpoint receives the request; reasoning effort determines how the model handles it.
Artificial Analysis reports Neon at 61 on Xhigh and 60 on High on its public Intelligence Index v5.0.1 display, rounded to whole points. OpenRouter’s benchmark table republishes the Xhigh figure as 61.4. The analysis firm’s public table reports the High result as an integer. Artificial Analysis’ reported task times on Google’s API were 21.5 minutes for Xhigh and 14.8 minutes for High on its test suite. Both timings come from the benchmark suite, with requests and scoring set by the evaluator. Any apparent advantage in a company’s workload requires testing with representative requests.
Engineers can see what the added effort buys when they inspect the finished patch and reproduce the problem it addresses. Test a case where the relevant behavior is split across services, then check if the proposal passes the existing tests and explains which service owns the fix. The next case may request a small change to a well-documented file. Extra minutes earn their place if the harder case needs less review or the proposal passes a test the cheaper run missed.
Google and OpenRouter Price the Request by Route
Google’s Gemini Developer API price table lists Neon Standard at US$3 per million input tokens and US$15 per million output tokens, with model reasoning included in billed output. The native rate for cached input is US$0.15 per million tokens on qualifying implicit cache hits. Google’s Flex tier lists US$1.50 for input and US$7.50 for output on the same basis. The native Priority rates are US$5.40 and US$27 per million. Shared reference text can lower spend when it qualifies for caching; live request data still needs its own budget.
OpenRouter’s provider table lists the Google Vertex and Google AI Studio standard routes at US$3 and US$15 per million input and output tokens on November 20. Google AI Studio Flex was listed at US$1.50 and US$7.50. The two listed Priority routes quoted US$5.40 and US$27. OpenRouter lists Flex and Priority outside Standard routing for this model. The gateway’s model identifier is google/gemini-4-neon. Eligible endpoints and permitted fallback depend on the routing settings. Quote an endpoint’s price alongside the model’s name whenever the distinction can change someone’s bill.
Consider a hypothetical request for a document review with a fixed total of 40,000 uncached input tokens and 6,000 billable output tokens. The output allowance in the example includes reasoning. It costs US$0.21 at the published Standard rates and US$0.105 at the listed Flex rates. Priority costs US$0.378 for the same token counts. Extra reasoning changes token counts in real requests, and a retry adds another billable attempt, generating chargeable usage. Tool charges and human review belong in a separate part of the calculation. Developers can compare the model charge with the matching usage record.

First Token and Finished Job Are Different Deadlines
Spreadsheet checks can wait until the audit is complete, while a live support question needs a useful first response as a customer watches the screen. Artificial Analysis recorded 56.51 seconds to the first answer token at Xhigh and 40.64 seconds at High in its Google API measurements. OpenRouter separately reports endpoint latency and output throughput for its provider fleet. The provider page showed 3.42 seconds for Vertex Standard and 3.93 seconds for AI Studio Standard in the November 20 snapshot. Endpoint readings describe OpenRouter’s own traffic across requests, so the publisher’s numbers and Artificial Analysis’ effort-specific readings belong in separate comparisons. Compare measurements from the same test setup and include the time external tools add when you assess a route for live chat.
Consider a retailer’s policy review in which a model can compare a return request against the applicable rules, but a colleague still decides whether to grant a discretionary exception. The first result needs a usable policy passage and a correct customer record. Work ends when the reviewer has a decision-ready case. Staff can verify a common return-window answer against a short passage in the approved policy. Lowering effort on the routine question can free up computing time and review capacity for exceptions.
Set a deadline for the first answer a customer can use and track a separate deadline for the completed job. Label the time when a usable reply first appears, then log when the work is complete. For a chat request, a greeting can arrive while the agent is still looking up a policy. In a background audit, the first streamed text has little value if the analyst receives the result too late to check it. Compare both durations with the current workflow’s deadline on the same cases.

What a Pilot Should Measure
One useful test for an AI operations agent starts with anonymized tickets and the same company policies reviewers use today. Include cases with a missing field or a source conflict so the model has a real opportunity to escalate. Run the same sample with High and Xhigh. Keep the available tools and source material constant in both runs. Reviewers can score answer accuracy and trace requests requiring discretion to the person who owns the decision.
For custom AI agents, the useful unit of cost is an accepted result, measured against the same policy and business deadline. Add the spend on retries and the reviewer’s time to the total cost of completing the work. Compare the settings for the same class of case after counting failures as part of the spend. If High passes the agreed quality threshold on straightforward requests, keep the extra reasoning for the failures it demonstrably corrects. The first useful decision may be as simple as assigning one exception type to Xhigh and leaving the rest of the queue as is.

Make your brand matter.

