Google made Gemini 4 Neon generally available on November 18, 2026, adding a paid API option for teams building coding and document assistants. Its Standard rates are US$3 per million uncached input tokens, $0.15 per million cached input tokens and $15 per million generated tokens—including thinking.
Tokens are the units used to count model usage. Thinking is internal reasoning: Google charges for the full amount generated, not just the shorter reasoning summary or final answer you see. A budget based only on the visible answer will therefore miss part of the bill.
Neon has no free API tier. Google AI Pro and Ultra include access in the AI Studio interface, but API-key and external-app usage are billed separately. If free API access is the deciding factor, compare Gemini Hydrogen’s access and pricing; that is a separate model choice, not evidence of equivalent results.
What would 1,000 requests cost?
Illustrative calculation, not a measured workload: assume 1,000 independent, single-hop requests in a billing period. Each processes 20,000 input tokens, of which 15,000 are actually billed as cached, and generates 6,000 thinking tokens plus a 2,000-token answer.
Using Google’s Neon Standard list prices as of November 18, the token charges would be:
Usage category | Tokens per request | US$ per million tokens | US$ for 1,000 requests |
|---|---|---|---|
Uncached input | 5,000 | $3.00 | $15.00 |
Cached input | 15,000 | $0.15 | $2.25 |
Thinking | 6,000 | $15.00 | $90.00 |
Visible answer | 2,000 | $15.00 | $30.00 |
Total | 28,000 | — | $137.25 |
Each row is tokens per request × 1,000 requests ÷ 1,000,000 × the applicable rate. The total is $0.13725 per request, or $137.25 for 1,000. Omitting thinking would understate this example by $90.
With no cache hits, the same workload would cost $180. Caching saves $42.75, or 23.75% of the total—not 95%. The 95% discount applies only to eligible cached input; thinking and answer charges remain.
The assumed cache share is not a promised hit rate, and cold requests can cost more. These totals exclude tools, extra model calls, retries, continuation hops, hosting, managed-agent charges, subscriptions, tax and currency conversion. They also exclude credits or negotiated discounts. Actual reasoning volume and answer length depend on the task.
When does the cache discount apply?
Neon uses implicit caching: Google automatically discounts input when it finds reusable content. Explicit cache objects are not supported. The minimum eligible input is 4,096 tokens, but crossing that threshold does not guarantee a hit.
Repeated conversations have another wrinkle. Google says that when a Gemini 4 turn adds fewer than 2,050 tokens, the previous request’s last 2,050 tokens are reprocessed at ordinary input rates. Added model output, thinking and tool results count toward that threshold. Use the returned cache count rather than assuming every repeated token gets the discount.
Teams comparing providers can also read when Bedrock prompt caching saves money. Its cache-write economics are provider-specific; do not transfer those rules to Neon.
Should you choose Standard, Batch or Flex?
Neon’s Batch and Flex token rates are $1.50 for uncached input, $0.075 for cached input and $7.50 for generation, per million tokens. With exactly the same usage as the example, either tariff gives $68.63 after rounding. That is a price calculation, not proof that the same job runs identically on both services.
Standard: a starting point for interactive work where the user is waiting for the answer.
Batch: consider it for independent, non-urgent tasks. The Batch API processes GenerateContent requests asynchronously, with a target turnaround of 24 hours—not a guaranteed deadline.
Flex: consider it when delays and retries are acceptable. This preview tier has variable latency and best-effort availability. Capacity shortages can return 503 or 429 errors; your application must handle retries. Google does not automatically upgrade an unavailable Flex request to Standard.
Priority costs $5.40 for uncached input, $0.27 for cached input and $27 for generation per million tokens: $247.05 for the illustrative workload. Check feature compatibility before moving a workflow between tiers. Running an interaction in the background does not, by itself, qualify it for Batch pricing.
How much output fits the shared token budget?
Input, thinking and the answer share the same 1,048,576-token window. That is a combined capacity limit, not a million-token input allowance plus a separate million-token answer allowance.
For example, if the complete input—including instructions and tool context—uses 900,000 tokens, only 148,576 remain for thinking and the answer: 1,048,576 − 900,000. A requested generation allowance of 262,144 would not fit. Caching changes the price of eligible input, not the space it occupies.
The default generation ceiling is 65,536 tokens, counting thinking and answer together. For longer generation, Google’s long-decode path requires a higher max_output_tokens allowance and background execution or client-side continuation. Merely enabling background execution leaves the default cap unchanged; raising the cap does not guarantee a longer answer.
Long decode divides generation into successive steps, called hops. Each hop processes the prompt and prior generated content again, with implicit-cache discounts on repeated input. Consequently, summed input usage across hops can exceed the live context size. Price the final aggregated usage, not just the original prompt, and do not add together cumulative progress reports.
Background execution requires stored interactions; client-side continuation can use unstored requests. Check completion status before treating the answer as finished. Cancelled and failed interactions still incur charges for work already done.
Which controls help keep spending predictable?
Start by choosing reasoning effort deliberately. Neon offers low, medium, high and xhigh; high is the default. Google says xhigh increases token use and latency. Evaluate the lowest setting that meets your task’s quality requirements rather than assuming the most intensive option is necessary.
Keep enough generation allowance for an answer. The thinking documentation explains that max_output_tokens covers reasoning as well as visible output. An overly small cap can be exhausted during thinking, leaving an incomplete response with no useful answer.
Budget tools separately. Google Search grounding includes 5,000 monthly search queries shared across Gemini 3 and newer models, then charges $14 per 1,000. One model request can trigger several queries. Google also counts grounding instructions and function declarations as tokens, so the user’s prompt alone is not the full input.
Before increasing traffic, check your active rate limits in AI Studio and set application-level limits on simultaneous jobs and retries. Google’s experimental project spend caps can overshoot because billing data may lag by around ten minutes and long-running jobs can continue accruing charges. These caps are unavailable to invoiced or offline accounts.
For the first budget, keep separate totals for uncached input, cache hits, thinking, answers and tools—including unsuccessful work. Then compare spending with the number of tasks that produced an acceptable result. Token rates alone cannot tell you which model or setting is cheapest per completed job.
