Skip to main content

Gemini 4 Neon: cache cost and task-based routing

Camila Duarte
Camila DuarteNovember 20, 202619 min. read
Gemini 4 Neon: cache cost and task-based routing

What changes with the arrival of Gemini 4 Neon?

Google announced Gemini 4 Neon on November 18, 2026, with cached input at US$ 0,15 per million tokens, compared with US$ 3 without cache, on the paid Standard tier.[1][3] The vendor announcement changes the calculation for gateway operators: they need to assess task class and actual cache use together.

Key facts

  • Neon Standard input: US$ 3 per million tokens; cached input: US$ 0,15.[1][3]
  • The 95% discount applies to input tokens that hit the cache, not to output or the entire bill.[1]
  • The choice requires measuring quality, latency, and cost per accepted task in the reader's operation.

Gemini 4 Neon changes the economic comparison between AI routes because Google charges different prices for regular input and cached input on its direct API. In the paid Standard pricing table consulted on November 20, 2026, the difference between US$ 3 and US$ 0,15 per million tokens represents 95% only for the portion of input billed as cached.[1][3] To estimate operational spending, the team must add output, unsuccessful attempts, and other charges from the contracted channel. The route decision also depends on accepted tasks and the response time observed for each class of work. The announcement does not prove that a particular integration receives the discount or that Neon is available in Nexforce Router. A team can attribute savings to its operation only after reconciling usage records with actual charges and checking acceptance criteria.

The discount does not cover the whole bill. A team that switches models based only on input price risks saving on one invoice line while spending more on the attempts needed to deliver an accepted result. That is what needs testing, not an automatic route change.

What did Google announce for Gemini 4 Neon?

In the official November 18 announcement, signed by Koray Kavukcuoglu, Google places Neon between Hydrogen, aimed at everyday work, and Carbon, intended for harder, longer tasks.[1] That is the vendor's positioning, not a universal ranking of quality for every task.

According to Google, Gemini 4 Neon occupies the middle position in the family between Hydrogen and Carbon, but that position does not determine which route wins on a specific task.[1] The vendor reports a context window of one million tokens and, in the paid Standard pricing table for the direct API consulted on November 20, 2026, charges US$ 3 per million input tokens, US$ 0,15 for cached input, and US$ 15 for output, including reasoning.[1][3] In the decimal-comma notation used here, those rates are US$ 3,00, US$ 0,15, and US$ 15,00 respectively. Google also reports 61 points on the Artificial Analysis Intelligence Index with xhigh and 1.336 points on the Agent Arena Index. Those results belong to the conditions of the cited tests, not to Nexforce's operation. The enterprise search company cited by Google reported, in early evaluations, three times the speed and a cost equivalent to 60% of Carbon, without disclosing all the conditions.[1] Teams should check quality and cost on their own tasks before switching routes.

Throughout this article, commas mark decimals and periods group thousands. For example, 1.336 means one thousand three hundred thirty-six, and 4.096 means four thousand ninety-six.

An available window does not mean every request needs to use all of it.

In the cited indexes, Google reports third place overall on the Artificial Analysis Intelligence Index and second place on the Agent Arena Index. The ranking for following corrections is a separate measure, distinct from the overall ranking.[1] Reasoning level, input composition, and acceptance criteria can change the outcome of a local comparison.

The enterprise search company cited by Google also reports quality equal to or better than Carbon's.[1] Those early evaluations are not a promise that applies to every operation.

The comparison with Gemini 4 Hydrogen pricing and cost per request[4] helps separate each model's role from the actual cost of a call. The new input to the decision is how much context repeats and is actually billed as cached.

Why does caching change input cost?

In Google's direct pricing table consulted on November 20, one million uncached input tokens cost US$ 3 on Neon; the same quantity of input billed as cached costs US$ 0,15.[1][3] The difference is 95%, but it applies only to the input that receives that rate.

Gemini 4 Neon's cache reduces the price of input tokens the vendor actually classifies as cached, not the cost of the entire request. Google's paid Standard pricing table consulted on November 20, 2026 lists US$ 3 per million regular input tokens and US$ 0,15 per million cached input tokens.[1][3] The difference between the rates is 95%. If only part of the input receives the discount, input billing combines the two rates; output and any tools are calculated separately. The documentation describes implicit caching, a minimum of 4.096 input tokens for Neon, and recommends common prefixes in requests close together in time.[2] These recommendations do not guarantee a hit. In the Interactions API, usage.total_cached_tokens reports cached tokens, but the operator must check whether the integration passes through that field and whether the discount appears in the contracted channel's billing.

The calculation is 1 - 0,15 / 3 = 0,95. If a request reuses only some tokens, the rest follow the applicable input rate. Output remains separate. Tools and any other applicable charges also enter the cost. The percentage does not describe savings on the full bill.

inline-01.png The 60% in the testimonial has a different denominator: it is Neon's cost relative to Carbon's cost in the initial evaluation reported by the company cited by Google.[1] There is no legitimate calculation that adds 95% to 60% or treats one measure as confirmation of the other. The first compares two Neon input rates; the second compares cost between models in a specific test.

In the Google API cache documentation, caching for Gemini 4 models is implicit. For Neon, the page lists a minimum of 4.096 input tokens and recommends putting common content at the start of the prompt and sending requests with similar prefixes close together in time.[2] These are conditions and recommendations intended to improve the chance of a hit, not a guarantee that each request will be billed at a discount.

A prefix can repeat in the text without producing a cache hit billed by the vendor. In the Interactions API, the usage.total_cached_tokens field reports tokens that hit the cache.[2] Before using it in a gateway dashboard, the team must verify that the endpoint and integration path preserve that data; when metadata is absent, the result is unknown, not necessarily a miss.

The documentation also describes a condition for multi-turn conversations in Gemini 4: to reuse the entire previous request, the new request must add at least 2.050 tokens. Below that threshold, the last 2.050 tokens of the previous request are reprocessed at the standard rate.[2] This is a rule in the API documentation and must be checked against the endpoint used in the operation.

What changes in task-based route selection?

A policy based only on the model's nominal tier does not show the cost of a completed task. The alternative is to compare routes by work class, minimum quality, latency, and observed cache use, using effective prices from the contracted channel. Neon adds an option to the analysis, not a route that wins by definition.

Task-based routing compares routes that meet the quality, security, and latency limits defined for each class of work. Among eligible routes, the team chooses the one with the lowest cost per accepted task, adding attempts and charges from the contracted channel. The cache rate helps explain the cost difference, but it does not replace quality measurement. Gemini 4 Neon enters that comparison as a candidate, not an automatic winner. The test must separate tasks with stable prefixes from tasks with variable inputs, because the cached-token rate for one class does not necessarily represent another. If no route meets the limits, the team responsible suspends automation and sends the decision for human review or uses a fallback previously approved for the task. The current route stays in place only if it meets the applicable limits. A cheap route that fails does not become eligible because of its price.

CriterionChoice guided only by model tierTask-based choice
InputPrioritizes advertised unit priceSeparates uncached tokens from tokens billed as cached
QualityAssumes the chosen tier is sufficientMeasures the proportion of accepted tasks in the class
LatencyUses a generic model expectationCompares the observed distribution with the service limit
CostCompares the rate per tokenAdds attempts and charges per accepted task
DecisionKeeps a broad route for different usesSelects among eligible routes and reviews the choice when data changes

The table compares approaches; it does not claim that every company used the first one. A single organization can have processes with a stable prefix and others in which every input is new. In that case, the proportion of reused tokens cannot be transferred from one class to another.

inline-02.png The figure summarizes a conditional sequence. First, a route must meet the minimum criteria. Then the team compares cost per accepted task among the candidates that passed. If none passes, the person responsible for the operation suspends automation, sends the task for human review, or uses a previously approved fallback; the current route stays in place only if it meets the applicable limits.

Routing when token prices change[5] already requires teams to review rates and actual call behavior. With Neon, prefix reuse adds a variable that can differ even between requests to the same service. A fixed rule for the entire application would hide that difference.

There is also a limit to the reverse argument: a simple task does not need to be sent to a middle-tier model just because it offers caching. If an eligible route meets acceptance criteria at a lower cost and sufficient latency, the use of a model with a higher price per token needs to be justified by measured results.

How can teams measure cost per accepted task?

Cost per accepted task is the sum of charges attributable to a cohort divided by the number of completed and accepted tasks in that same cohort. The formula includes unsuccessful attempts and actual costs, not just the call that produced the final response. If there are no accepted tasks, the metric is undefined.

Cost per accepted task compares spending for a class of work with the number of tasks in that class that passed acceptance. To calculate the metric, the team adds uncached input tokens, cached input tokens, and output tokens across all attempts, applying the effective rate for each component. Tools and other applicable charges also enter the numerator. The denominator is not the number of calls: it is the number of completed and accepted tasks, judged by the same criteria for each route. Telemetry records total input, the cached subset, output, and the class success rate. When cache data is incomplete, the team reports coverage and does not treat unknown requests as misses. Before attributing savings to a route, the team reconciles recorded consumption and contracted rates with the vendor's or intermediary's charges.

The numerator includes uncached input, cached input, output, and reasoning when it is billed as output, as well as tools and other charges that actually apply. The same task may require a second attempt or automated review: excluding that expense would make the route look cheaper than it was. The calculation should use the invoice or billing data from the chosen channel, not an assumed rate.

For the class and period being compared, the formula is custo por tarefa aceita = C_total / A, with C_total = (I_sem_cache × P_entrada + I_cache × P_cache + O × P_saída) / 1.000.000 + C_outros. I_sem_cache, I_cache, and O are the totals of uncached input, cached input, and output tokens across all attempts; P_entrada, P_cache, and P_saída are their respective effective rates per million tokens. C_outros combines other attributable charges, such as tools, and A is the number of accepted tasks. If rates vary during the period, the team calculates each portion by price band before adding them. When A = 0, the ratio is undefined.

In Neon's paid Standard pricing table consulted on November 20, for I total input tokens and H of them actually billed as cached, the input component is [3 × (I − H) + 0,15 × H] / 1.000.000, in dollars.[3] For I > 0, the savings on that input, compared with the same input without caching, are 95% × H / I. The calculation does not estimate savings on the total bill or represent an observed customer result.

The denominator needs a criterion decided before the test. A structured extraction, for example, must pass required-field and content-validation rules defined by the team; a response that arrives quickly but fails acceptance does not count as an accepted task. The criterion must be the same for the routes being compared.

Cache measurement requires two readings. The first is the fraction of requests in the class that reuse the common prefix identified by the team. The second is the fraction of input tokens reported as cached: cached tokens divided by the corresponding total input tokens. One measures the presence of reuse; the other measures its weight in consumption. They are not equivalent, and the cache counter alone does not prove a discount on the bill.[2]

For each class and route, telemetry records total requests, requests with valid cache metadata, input tokens, cached input tokens, and output tokens per attempt. It also records tasks started and accepted, to calculate the success rate as accepted tasks divided by tasks started in the cohort. The token-weighted cache hit rate is the sum of cached tokens divided by the sum of total input tokens for the same requests with valid metadata; the team first confirms the endpoint fields' definitions and compatibility. Coverage is the number of requests with valid metadata divided by the total. Requests without metadata stay outside the cache ratio and are reported as unknown, not as misses.

To reconcile billing, the team groups records by the same period, model, channel, and tier as the invoice, applies contracted prices, and compares tokens and amounts with the vendor's or intermediary's statements. Discounts, adjustments, tools, and other charges are identified separately, without counting reasoning already included in output twice. Discrepancies are recorded and investigated; if channel data does not support reliable attribution, the billed discount and cost per class remain unconfirmed.

Identifying a prefix version prevents a misleading reading when shared text changes. The team also records model and version, reasoning level, context range, pricing-table date, tier, team, a pseudonymized task identifier, and number of attempts. Sensitive content does not need to enter the record for the operation to be measured.

The Google pricing table, consulted on November 20, offers a counterpoint: Carbon had promotional Standard pricing through November 30, 2026, of US$ 2 per million input tokens, US$ 10 for output, and US$ 0,10 for cached input.[3] From December 1, the table listed US$ 4, US$ 20, and US$ 0,20, respectively. It would therefore be incorrect to claim that Neon always costs less per token than Carbon.

The earlier analysis of Carbon's initial price and cost per task[6] gains a new condition: the comparison must fix the date, endpoint, tier, and purchasing channel. These are vendor prices, not a Nexforce quote. A price change can reverse the ordering of routes even without a change in quality.

Which decisions should teams make now?

The recommendation is to test by task class before changing the gateway policy. The team defines acceptance, checks available telemetry, and compares routes under the same conditions, including cold and warm cache. A rate difference becomes proven savings only when it appears in the cost of accepted tasks without exceeding service limits.

To test Gemini 4 Neon without confusing price with results, the team must compare the same representative tasks on each candidate route and fix acceptance criteria before execution. As an initial operational minimum, the test uses at least 100 distinct tasks per class, evaluated on each route; that number does not guarantee statistical precision and should increase according to the difference the team wants to detect and the variability observed. The owner defines quality, security, and latency limits in advance, including the response percentile, and documents how errors and attempts are handled. The report separates cold-cache and warm-cache runs, records cache hit rate and its coverage, and reports success rate by class. The team compares reconciled costs and presents the uncertainty in the differences. If the sample or coverage cannot support the conclusion, the test remains inconclusive and does not authorize an automatic policy change.

  1. Separate work classes. The team groups tasks by objective, context size, and output requirements. It identifies common prefixes and their versions without combining tasks that have different acceptance criteria.
  2. Set quality and latency criteria. Before execution, the owner documents task-level acceptance criteria and numerical limits for success rate, security, and latency percentile for each class; the owner also defines when a failure requires another attempt or human review.
  3. Check billing data. The team verifies whether the contracted channel reports cached input tokens and whether the gateway passes through the metadata. It records missing data as unknown and includes the operation's actual charges.
  4. Compare with controls. The team uses at least 100 real anonymized tasks per class, the same sample for each route, and comparable configurations; it records cache hit rate, coverage, sample size, dispersion, errors, and attempts.
  5. Choose and review the route. Among routes that meet the limits, the team prefers the one with the lowest cost per accepted task. If none passes, the owner suspends automation, sends the task for human review, or uses a previously approved fallback; the current route stays in place only if it meets the applicable limits.

For cold and warm runs, the team records times, cadence, prompt version, model, reasoning level, and channel. The design prevents one route from inadvertently warming the other's cache; results are aggregated only according to the observed proportion of cold and warm conditions in real traffic. Results for each condition, latency percentiles, and uncertainty in the difference remain visible in the report.

A sequence with many similar requests can appear representative of an operation that receives varied inputs spaced apart in time. Conversely, a cold-cache-only test can miss a reuse pattern present in the actual routine.

The team should also track the latency distribution, not just the average. A route that improves the average but leaves some tasks above the response limit requires investigation before it receives more traffic. The same applies to a route that cuts spending per call while increasing the number of attempts.

The question discussed in the analysis of token price and AI cost per task[7] remains valid: does a lower rate change the cost of the delivered task? For Neon, the answer must include how many tokens actually hit the cache and how many tasks passed acceptance.

Frequently asked questions

The cache discount applies to reused input, not the entire bill, and no request is guaranteed a hit. To choose a route, the team must observe tokens billed as cached and divide cohort spending by accepted tasks. The prices below are Google's, not a Nexforce offer.

For Gemini 4 Neon, the 95% reported by Google compares only the Standard rates for cached and uncached input, consulted on November 20, 2026.[1][3] Output and other costs do not receive that percentage by definition. Implicit caching does not guarantee a hit on every request either: the team must observe reported tokens and check billing. To compare routes, cost per accepted task adds the attempts for a class and divides spending by the number of accepted tasks; without accepted tasks, the ratio does not exist. Direct API values are not automatically Nexforce Router prices. Neon's availability, contracted pricing, and access to cache metadata in Router require confirmation. The quality, speed, and cost results cited by Google belong to the evaluations reported by the company mentioned in the announcement, not to a Nexforce measurement.

Does the 95% discount reduce the entire bill?

No. The 95% compares US$ 0,15 for cached input with US$ 3 for uncached input per million Neon tokens on Google's paid Standard tier.[1][3] Output, input that did not hit the cache, and other charges follow their own terms. The discount does not represent savings on the whole invoice.

Will every Gemini 4 Neon request use the cache?

No. Google describes implicit caching, a minimum of 4.096 input tokens for Neon, and recommends common prefixes at the beginning of requests close together in time.[2] That does not guarantee a hit on every call. The cached-token field must be checked for the endpoint and integration in use.

What does cost per accepted task measure?

It measures total cohort spending, including all attempts and applicable charges, divided by the number of completed tasks that passed the acceptance criteria. If no task is accepted, the ratio is undefined. Using the number of calls as the denominator would answer a different question about the operation.

Is Google's advertised rate the Nexforce Router price?

Not necessarily. The quoted figures are from Google's direct API on the stated tier and may change.[3] Neon's availability, effective pricing, and access to cache metadata in Nexforce Router must be confirmed at the time of contracting. This article does not claim that the model is already available in Router.

Does the account cited by Google prove savings for every team?

No. The account of quality, speed, and cost reflects the initial evaluations of the company cited in Google's announcement, whose full conditions were not disclosed.[1] A team can conclude that it saved money only after measuring its own sample, its charges, and the number of accepted tasks.

References and Further Reading

[1] Google: Gemini 4 Neon announcement [2] Google: cache documentation [3] Google: pricing table [4] Gemini 4 Hydrogen: price and cost per request [5] LLM routing and token prices [6] Gemini 4 Carbon: price and cost per task [7] Token prices and AI cost per task [8] Nexforce Router

What is the next step for the routing policy?

Nexforce Router is gateway infrastructure with routing rules and consumption observability.[8] The news makes it relevant to check how the operation obtains the data needed to decide by task. That still requires verification. The announcement does not establish Neon availability, cache metadata pass-through, or a specific automatic rule in the product.

The next routing policy review should start with the quality, security, and latency limits for each task and the effective rate of the contracted channel. Nexforce Router is gateway infrastructure with routing rules and consumption observability, but Google's announcement does not confirm that Gemini 4 Neon is available in Router or that the specific cache metadata is passed through.[8] Those points require confirmation at the time of contracting. The team should record input, cache, and output tokens, associate attempts with the same task, and reconcile amounts with billing before comparing cost per accepted task. If metadata is absent, the report identifies the gap without assuming a discount. If no route meets the limits, the person responsible for the operation suspends automation, sends the task for human review, or applies a fallback previously approved for the task. The current route remains only if it meets the limits.

Nexforce

Accelerate your company'sbusiness and operational efficiency

We design the technology of tomorrow to boost your business operational scale

Talk to a Specialist

Related articles