Gemini 3.8 Flash: Pricing, Model ID, and a Safe 3.7 Migration

Gemini 3.8 Flash is a production model, not a rumored alias. Google released it on September 2, 2026, and the stable Gemini Developer API model string is gemini-3.8-flash. It is a sensible first candidate for a new general-purpose Gemini integration. For a service already running 3.7, it should begin as a reversible challenger, not an automatic fleet-wide replacement.

The reason is easy to miss. Gemini 3.8 Flash has the same current per-token rates as 3.7 Flash, but Google says 3.8 may take more reasoning steps and make more tool calls on difficult work. Equal rates do not guarantee equal token consumption, equal latency, or equal cost per task that actually passes review.

The launch also includes Gemini 3.8 Flash Cyber. That name does not identify a second public model row. Cyber is reserved for approved defenders in Google's Fairwind Program. Migrating a normal API application and applying for a restricted cybersecurity program are separate decisions.

What the public 3.8 model actually includes

As of September 3, Google's Gemini 3.8 Flash specification lists these details:

Multimodal input does not imply multimedia output. If the product must generate images, synthesize audio, or run a real-time voice session, select a purpose-built model instead of assuming the higher version number adds those outputs.

Google's current deprecation table gives neither gemini-3.8-flash nor gemini-3.7-flash a shutdown date. That is not a promise of permanence. It does mean 3.7 remains a legitimate rollback and efficiency option today; there is no published retirement deadline that requires a rushed cutover.

The rate card is unchanged; the economics may not be

The Gemini Developer API pricing page gives 3.8 and 3.7 the same introductory prices:

Output billing includes thinking tokens. Context caching storage, Priority processing, Search grounding, and Maps grounding have additional rows. These are USD rates for the Gemini Developer API, not Vertex AI, an enterprise agreement, taxes, or a third-party gateway. The broader Gemini API pricing and credential guide covers account funding and provider boundaries.

Suppose a request uses 50,000 input tokens and 8,000 output tokens including thinking. At the Standard introductory rate, it costs about:

text
(50,000 × $0.75 + 8,000 × $3.75) / 1,000,000 = $0.0675

If 3.8 uses 14,000 output and thinking tokens on the same input, the call costs $0.09. If that extra work turns a two-attempt 3.7 task into a one-attempt success, 3.8 can still be cheaper. The useful denominator is therefore:

text
cost per accepted result = total request cost / results usable without repair

This captures retries and rejects that a raw price comparison hides.

A clean 3.7 integration may need only one code change

For an application already following the Gemini 3 requirements, the functional change can be as small as replacing the target model with:

text
gemini-3.8-flash

Still, Google's current migration checklist is worth auditing before the canary. It calls for these constraints:

• Remove temperature, top_p, and top_k from generation configuration.
• Replace thinking_budget with the string enum thinking_level; do not send minimal to 3.8.
• Remove candidate_count, which Gemini 3 does not support.
• Use server-side previous_interaction_id for multi-turn work and remove prefilled model turns.
• Preserve thought signatures and check the format of function-call payloads.
• With generateContent, include call_id and name in each FunctionResponse.
• If pre-tool text causes Malformed_Function_Call, use the documented workaround rather than masking it with retries.

Many well-maintained 3.7 applications already satisfy most or all of these rules. They are not all new 3.8 breaking changes. An audit can still catch old compatibility code, a gateway transform, or a legacy prompt path that reintroduces unsupported fields.

Put the exact ID behind one configuration value:

bash
GEMINI_MODEL=gemini-3.8-flash

Keep gemini-3.7-flash as the known rollback value until the rollout completes. Avoid using a moving alias such as gemini-flash-latest as the acceptance-test baseline: a test is reproducible only when the model under test stays fixed.

Let a matched workload authorize the migration

Build a compact acceptance set from recent successes, known failures, and the boundary cases that would hurt if they regressed. Run 3.7 and 3.8 against identical prompts, input files, tool schemas, timeouts, retry rules, and thinking levels. Record at least:

1. Whether the result is usable without manual repair.
2. Whether every tool choice and argument validates.
3. Input, output, and thinking tokens.
4. First-token and end-to-end latency, including timeouts.
5. Formatting errors, safety blocks, and retries.
6. Total cost per accepted result.

After offline or shadow evaluation passes, send only a small share of eligible production traffic to 3.8. Expansion criteria should be observable: acceptance rate no worse than the current baseline, no rise in tool failures, and P95 latency plus accepted-result cost within budget. “The new model feels smarter” is not a release criterion.

The expected advantages of 3.8 are most relevant to long-horizon software engineering, autonomous multi-step agents, and complex professional workflows. Simple classification, short extraction, or high-volume transformation may not benefit from additional reasoning. If 3.7 already meets the quality bar, keep it for that workload or retest 3.8 with thinking_level: "low". Google explicitly keeps 3.7 supported for efficiency-first work.

Gemini 3.8 Flash Cyber has a different front door

Google describes the Fairwind Program as limited access for approved trusted partners. It prioritizes governments and national cyber authorities, critical-infrastructure operators, core technology platforms, and academic labs focused on defensive benchmarking. Google performs organizational due diligence and does not promise that an application will be accepted or processed by a fixed date.

Approved partners receive exclusive access to Gemini 3.8 Flash Cyber, either as a standalone model or with CodeMender. They may grant it only to internal cybersecurity, incident-response, or penetration-testing teams. Participating organizations must use individual authentication, phishing-resistant MFA, access controls, and usage tracking. Authorized threat simulation, reverse engineering, and malware analysis are permitted for defensive or academic research; malicious work such as creating malware is not. Partners may not share, resell, or redistribute access.

Those conditions lead to three practical conclusions:

• Do not treat gemini-3.8-flash-cyber as a documented public API identifier.
• Do not apply the public 3.8 Flash price table to Cyber; Google has not published a general public Cyber rate card.
• Do not buy supposed “Cyber access” from a reseller. Fairwind explicitly prohibits partners from redistributing it.

If your organization is not eligible, Google points to CodeMender with publicly available models and other AI Threat Defense products. Their price, region, and product requirements still need a separate check; they are alternatives, not equivalent access to the restricted model.

The practical choice

For a new application, evaluate gemini-3.8-flash first: it is GA, has a clear public specification, and currently carries the same unit rates as 3.7. For an existing 3.7 deployment, audit API compatibility, compare a matched workload, shadow the new model, and canary a small traffic share. Expand only when accepted work, tool reliability, latency, and total cost clear your thresholds.

Keeping 3.7 for stable, efficiency-sensitive tasks is not a failure to upgrade. It is a supported workload decision. And if the real requirement is Gemini 3.8 Flash Cyber, start with organizational eligibility and governance—not a search for a hidden model string.