Google has publicly confirmed that it started pre-training Gemini 4, in its July 22, 2026 Q2 letter and the July 21 Gemini 3.6 Flash announcement. As of September 16, 2026, the public Gemini API model catalog, pricing page, and changelog contain no Gemini 4 entry: there is no callable model ID, price, specification, release date, or public access path. Use this page to track verified access and prepare a reversible EvoLink evaluation without guessing the API contract.
We will email you after Google publishes a developer API path and EvoLink verifies the route, returned model identity, usage, pricing, and production behavior.
EvoLink product alerts only. This is not a Google program and not an access guarantee.
Want faster updates? Join the EvoLink Discord
The update will include the verified request model ID, live EvoLink pricing, supported inputs and parameters, access limits, and a migration path that retains a current Gemini fallback.
The status is based on first-party Google records checked on September 16, 2026, not on release-date predictions, leaked specification tables, or community model labels.
Google's public model catalog, pricing page, and API changelog do not list Gemini 4, a request model ID, commercial terms, a preview program, or a release date. The newest listed Gemini models are Gemini 3.8 Flash and Gemini 3.8 Live.
EvoLink has no Gemini 4 route, SKU, or integration date. Use a listed current Gemini route and keep the future candidate unset until request success, returned identity, usage accounting, and billing are verified.
States that Google "started our most ambitious pre-training run yet, for Gemini 4". It names the model but gives no date, variant, specification, or API plan.
Lists current Gemini API models through Gemini 3.8 Flash and Gemini 3.8 Live, with no Gemini 4 model code, preview ID, or callable route.
Google publishes no Gemini 4 pricing row or API changelog entry. Third-party price and context claims are predictions, not documentation.
This page changes from availability watch to live access only after Google publishes the API contract and EvoLink verifies a successful authenticated request, returned model identity, usage, and billing. A Google announcement alone does not change the EvoLink status.
Google confirms the model is in development, but it has published no model card, capability table, pricing, or benchmark for Gemini 4. The items below are what developers say they need, written as tests to run later, not as claims about the model.
No Gemini 4 capability is confirmed beyond its status as the next-generation Gemini model in pre-training as of July 2026. The useful question is whether a released route creates a measurable advantage over the Gemini 3.8 Flash and Gemini 3.1 Pro routes you can evaluate today.
Teams want a next-generation model that justifies migration effort on difficult coding, reasoning, and knowledge work. A new version number alone does not prove that gain.
Agent teams care about complete traces, valid tool calls, recovery after failures, and low reviewer intervention, not only one-shot scores.
A newer route is valuable only when tokens, retries, latency, fallbacks, and review effort produce better economics for accepted work.
Production adoption needs stable identity, inputs, tools, quotas, error behavior, usage records, and rollback, not a consumer announcement alone.
Prepare evidence, configuration, and rollback paths now so launch-day work is validation instead of a provider-specific rewrite.
Do not hard-code a guessed Gemini 4 identifier. Leave the candidate unset until EvoLink verifies a successful request and returned model identity.
Collect real coding, agent, document, multimodal, tool, and structured-output tasks with objective acceptance criteria.
Record quality, latency, tool errors, token use, retries, review effort, and accepted-task cost for Gemini 3.8 Flash and Gemini 3.1 Pro.
Set thresholds for task success, p95 latency, total cost, incidents, policy behavior, human correction, and rollback.
Capture requested and returned model, parameters, inputs, tools, usage, finish reason, safety blocks, fallback chain, and billing.
Keep a measured current route serving users until the candidate proves a workload-specific improvement under controlled traffic.
Use the documented behavior of Gemini 3.8 Flash and Gemini 3.1 Pro to prepare checks. These baselines are for your evaluation harness; they are not Gemini 4 specifications or price forecasts.
Current Gemini 4 answer: model ID, pricing, limits, input contract, variants, release date, and EvoLink route are all unpublished or unverified.
| Field | Documented current Gemini reference | Verify for Gemini 4 |
|---|---|---|
| Model ID and variants | gemini-3.8-flash and gemini-3.1-pro-preview are documented endpoints | Exact Gemini 4 ID, whether Pro/Flash variants exist, release stage, snapshot behavior, and returned identity |
| Standard token price | Gemini 3.8 Flash and 3.1 Pro have separate published prices | Official Gemini 4 rates and live EvoLink route pricing; never inherit a 3.x price |
| Batch, caching, and service tiers | Support and rates vary by current Gemini endpoint and prompt size | Eligible tiers, cache TTL and rules, storage price, and route billing |
| Context and output | Gemini 3.8 Flash documents a 1M-token context window | Gemini 4 input and output limits, useful-context behavior, truncation, and long-context pricing |
| Inputs and tools | Current endpoints document multimodal inputs, thinking, tools, grounding, and structured output | Supported inputs, function schema, grounding, code execution, streaming, and error contract |
| Quotas and data terms | Limits and data-use rules vary by tier, account, and Google channel | RPS, TPM, concurrency, regions, preview terms, retention, and 429 behavior |
| EvoLink route | Keep a measured current Gemini route as the fallback | Request success, returned identity, usage, capacity, billing, and rollback |
These are the request fields teams should prepare and re-test on launch day. Current Gemini behavior is a compatibility baseline, not a Gemini 4 specification.
Gemini 4 parameter support is unknown. Keep the candidate model ID unset until a public API contract and a verified EvoLink route exist.
| Parameter or field | Current Gemini behavior | Gemini 4 launch-day check |
|---|---|---|
| model | Current endpoints use documented model codes | Exact ID, release stage, aliases, and returned identity |
| contents | Gemini content and part objects | Roles, text, image, audio, video, PDF parts, and validation |
| systemInstruction | Supported by current Gemini generation endpoints | Format, precedence, token accounting, and limits |
| maxOutputTokens | GenerationConfig output ceiling | Maximum, defaults, truncation, and billing |
| temperature / topP / topK | Sampling controls vary by model and task | Accepted ranges, defaults, interactions, and errors |
| thinkingConfig | Thinking controls are documented for current Gemini 3 models | Supported controls, defaults, token visibility, quality, latency, and cost |
| tools / toolConfig | Function calling and selected built-in tools are model dependent | Schema subset, tool choice, parallel calls, and built-in tool support |
| responseMimeType / responseSchema | Structured output is supported on documented current routes | JSON Schema support, validation, finish reasons, and compatibility |
| cachedContent | Context caching is endpoint and tier dependent | Eligibility, TTL, cache identity, hit accounting, and billing |
| safetySettings | Request safety settings use documented harm categories | Supported categories, thresholds, blocks, and finish reasons |
| streamGenerateContent | Streaming is available on compatible Gemini generation routes | Chunk order, usage delivery, tool events, errors, and cancellation |
| usageMetadata / finishReason | Responses expose token usage and completion state | Field shape, thinking/cache accounting, blocks, and invoice reconciliation |
| Limits and regions | Account, service tier, channel, and region dependent | Context, output, RPS, TPM, concurrency, residency, and 429s |
Keep the candidate unset until EvoLink verifies a successful request and returned model identity. This configuration-only example does not send an API request and does not assume a Gemini 4 model ID.
const activeModel = process.env.EVOLINK_PRIMARY_MODEL
const fallbackModel = process.env.EVOLINK_FALLBACK_MODEL
const gemini4Candidate = process.env.EVOLINK_GEMINI_4_MODEL || null
// Keep the candidate unset until EvoLink verifies a callable route and returned identity.
export const modelRouting = {
activeModel,
fallbackModel,
gemini4Candidate,
}Google develops Gemini; EvoLink's role is to make verified model access easier to compare, operate, switch, and roll back after a route is genuinely callable. Today that means evaluating the current Gemini routes and keeping model selection in configuration.
Keep one integration, key, billing surface, and observability path while model selection changes through configuration.
Publish access only after request success, returned identity, usage, pricing, and billing are validated.
Evaluate the candidate offline and in canary traffic while retaining a current Gemini route as the measured rollback.
Use the release tracker for time-sensitive evidence and the comparison guides for decisions about models you can evaluate today.
Track what Google confirmed, what is rumored, the public API gates, and the evidence that would change this page's status.
For teams already on OpenAI workflows: what to run now, what to adapt, and what Gemini 4 must prove before a shadow test.
For long-running agent teams on Claude: record your context-reuse and recovery baseline, then define what a new route must prove.
Google has publicly confirmed that Gemini 4 is in pre-training, in its July 22, 2026 Q2 letter and its July 21 Gemini 3.6 Flash announcement. That is a development statement, not a product launch: no release date, variant, specification, or API has been announced.
Google has not published a release date. Year-end predictions circulating online are third-party estimates, not official information.
No. As of September 16, 2026 it is absent from Google's public model catalog, pricing page, and changelog. The newest listed models are Gemini 3.8 Flash and Gemini 3.8 Live.
No public request model ID exists. gemini-4 is this page's URL slug, not an identifier you can call, and you should not assume the final ID will follow that pattern.
Pricing is unpublished. Current Gemini 3.8 Flash and 3.1 Pro prices are baselines for your own cost model only, not a forecast for Gemini 4.
Google has not confirmed any Gemini 4 variant. Do not extrapolate names from the Gemini 2.5 and 3.x lineups; this page will list variants only after Google documents them.
No. Gemma is Google's open-weights model family and is a different product. Gemma 4 downloads, licenses, or self-hosting guides are not evidence about the Gemini 4 API.
No. Google's statement describes the start of a training run in July 2026; it does not say how long training, post-training, or evaluation will take, and it does not commit to a launch window.
No. EvoLink has no Gemini 4 route, SKU, or integration date. The status changes only after Google publishes an API contract and EvoLink verifies a successful authenticated request, returned model identity, usage, pricing, and billing.
Review Gemini 3.8 Flash for the current Flash tier or Gemini 3.1 Pro for the current Pro-family baseline, then verify the listed EvoLink route before production.
No. Keep model selection configurable, evaluate current routes, save representative traces, and maintain a measured fallback. Waiting for an undated model has a cost and no schedule.
Verify the official model ID, price, inputs, context and output limits, tools, thinking controls, streaming, quotas, regions, request success, returned identity, usage, and billing, then promote through replay, shadow, and canary stages with rollback.
Review provider-documented current models and verify each EvoLink route before production. Compare Gemini 4 only after identity, request, usage, and billing verification.
Gemini 3.8 Flash
ListedGoogle's newest listed Flash model, positioned for long-horizon software engineering and agents. Verify the EvoLink route before production.
View model
Gemini 3.1 Pro
ListedThe documented current Pro-family baseline for complex reasoning, coding, tools, and multimodal work. Verify the EvoLink route before production.
View model
Gemini Model Family
CompareCompare current Gemini listings by model access, price, context, production role, and workload fit.
View model