Gemini 3.8 Flash Keeps Its Low Price—but Cyber Access Is Gated

Google launched Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, but gave the two variants different routes to users. Google’s launch announcement sets the standard model at the same introductory price as 3.7 Flash—$0.75 per million input tokens and $3.75 per million output tokens—while limiting Cyber to trusted defenders through the new Fairwind Program.

The standard model was released across Google’s developer and enterprise channels, with access also included in selected subscription products; Cyber was not made a generally selectable API model. Ars Technica’s September 2 launch report also identified this access split and noted that the standard model improved on 3.7 Flash in OSWorld-2.0 computer use but remained well behind the leading Claude Opus result.

The low price is introductory, not permanent

The headline rate applies only through December 31, 2026. From January 1, 2027, the listed API prices rise to $1.50 per million input tokens and $7.50 per million output tokens—twice the launch rates in both directions.

That deadline changes the economics of deployments expected to run into 2027. It also limits what can be inferred from the token price alone: Gemini 3.8 Flash can take additional reasoning steps and call tools repeatedly at higher effort settings, so a difficult agentic task may consume more tokens even when each token is temporarily inexpensive.

Lower effort settings remain available when token use or latency matters more than maximum benchmark performance. Gemini 3.7 Flash also remains supported, leaving developers with a cheaper-computation option rather than forcing an immediate migration to the more deliberative model.

A deployment matrix separates price, access and evidence

Treating the two releases as interchangeable obscures the decision that developers and security teams actually face. The relevant distinctions are:

• Gemini 3.8 Flash launch price: $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026.
• Gemini 3.8 Flash regular price: $1.50 per million input tokens and $7.50 per million output tokens beginning January 1, 2027.
• Standard-model access: available through the Gemini API in Google AI Studio, Gemini Enterprise and named Google developer tools, with separate subscription access in selected consumer products.
• Cyber access: application-based and controlled through Fairwind, with priority given to trusted government authorities, critical-infrastructure operators and software maintainers.
• Performance evidence: the launch scores are vendor-presented results. Independent publications have compared and contextualized them, but that is not the same as independently reproducing the evaluations.

The earlier Flash and Cyber product split helps explain the continuity: Google again pairs a broadly distributed workhorse with a defensive variant whose additional cyber latitude comes with tighter distribution.

The benchmark gains are real claims, but uneven ones

The strongest reported improvements cluster around terminal work and agentic software engineering rather than every form of reasoning. DataCamp’s comparison of Google’s published results lists 90.8% for Gemini 3.8 Flash versus 81.6% for 3.7 Flash on Terminal-Bench 2.1, but only 61.6% versus 60.4% on SWE-Bench Pro; it also records 45.4% versus 45.7% on the broader Humanity’s Last Exam comparison.

Those figures support a narrower conclusion than a universal upgrade. The model appears substantially stronger on one terminal-oriented test, moves only modestly on a repository-level benchmark and is effectively flat on the cited exam-style comparison. The OSWorld-2.0 result adds another boundary: computer use improved relative to the previous Flash release without reaching the leading result reported for Claude Opus.

The provenance matters as much as the scores. DataCamp is independent coverage, but the figures it compares come from Google’s developer materials and launch disclosures; the publication did not rerun the evaluations. For deployment decisions, the current evidence establishes what Google measured and how outside coverage interprets the gaps, not independently verified production performance.

Cyber exchanges open access for broader defensive capability

Gemini 3.8 Flash Cyber shares the standard model’s foundational intelligence but is specialized for autonomous vulnerability discovery and automated patching. Its cybersecurity mitigations are more permissive than those of the standard release, while its stated emphasis remains finding and fixing weaknesses rather than offensive exploitation.

Fairwind is therefore an access boundary, not merely a preferred queue for an otherwise public model. Organizations must apply, and eligibility depends on Google’s assessment of whether they are trusted defenders with a legitimate need for the broader cyber capabilities. Most developers can call the standard model without gaining access to its Cyber counterpart.

The disclosed security results include a success rate above 70% on an internal vulnerability-discovery evaluation spanning 20 programming languages and 47.2% pass@1 on the externally defined CWE-Bench patching task, compared with 47.8% for a leading model. The benchmark may be external, but the submitted result and the internal multilingual evaluation remain vendor evidence; neither establishes how the model will perform across an applicant’s production repositories.

The launch consequently leaves two different decisions. General developers can evaluate Gemini 3.8 Flash now, while budgeting for the scheduled price increase and measuring total task-level token use. Security organizations must first clear the Fairwind gate, and the broader market still lacks independent reproduction of the Cyber results or operating data from approved deployments.

Also read:

• Google’s Gemini 3.6 Flash, 3.5 Flash-Lite and Flash Cyber: What Changes
• M-Learning Expands Access—but Distraction, Design and BYOD Set the Limits
• Nano Banana 2 Lite Is GA, but Gemini Omni Flash Remains a Preview