Google launched Gemini 3.8 Flash on September 2 with the same introductory token prices as 3.7 Flash and a qualification that deserves attention: on difficult jobs, the new model may consume more tokens. Google says its improvements come partly from additional reasoning and repeated tool calls. That makes this a release to evaluate against completed work, especially if an existing automation already runs cheaply and reliably. Google’s announcement also introduces a separate cybersecurity variant, Gemini 3.8 Flash Cyber, with restricted access.
For a developer paying the API bill, there are two questions here. Does the extra computation prevent enough failed attempts to pay for itself? And will the workflow still make economic sense when the introductory price expires?
The price has an expiration date
Google’s API pricing page lists standard paid-tier rates of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Output billing includes thinking tokens. On January 1, 2027, those rates become $1.50 and $7.50 respectively. The launch announcement confirms that schedule.
The arithmetic is straightforward. A hypothetical job using 100,000 uncached input tokens and 20,000 billed output tokens costs $0.15 at the introductory standard rates, excluding tool charges. The identical token usage costs $0.30 under the announced January rates. These are illustrative calculations, not measured Gemini workloads.
That difference belongs in a production forecast now. A pilot can look attractive on a small September bill and become a different purchasing decision once it runs continuously after the discount expires. Extra reasoning adds another variable: even before January, an unchanged rate card does not guarantee an unchanged cost per job.
Evaluate the work that currently fails
Google positions 3.8 Flash around software engineering and longer agent tasks. It also explicitly keeps 3.7 Flash supported for workloads that prioritize efficiency and suggests lowering effort when token consumption matters. Those qualifications are useful. An upgrade does not need to replace every existing model call to justify its place.
Our starting point would be a set of tasks the current system struggles to finish: a code change that needs repeated correction, for example, or a report whose evidence a reviewer keeps sending back. Run the same inputs through both versions, preserve the tool permissions, and have someone assess the outputs without seeing the model names. Record the total billed usage across retries alongside the review time.
Suppose a model spends more on its first attempt but saves a developer a second debugging pass. That could be a good trade. If it elaborates on an already correct classification and changes no downstream decision, the extra computation has little value. These are evaluation criteria, not claims that we have measured either outcome with 3.8 Flash. They follow the same approach as a useful AI proof of concept: choose an observable result before choosing the winner.
Flash Cyber has a separate access boundary
The Cyber variant is available through Google’s Fairwind Program, which vets applicants and prioritizes governments, critical infrastructure operators, and core technology platforms. Partners must restrict access to designated security teams, track employee use, and use phishing-resistant multifactor authentication. They cannot resell or redistribute model access.
Fairwind partners can use Flash Cyber directly or through CodeMender, Google’s security agent. Organizations outside the program can use CodeMender with publicly available models. A product carrying the CodeMender name therefore does not, by itself, establish that the restricted model is doing the analysis.
This is a procurement distinction familiar from OpenAI’s Daybreak access tiers. A buyer evaluating an AI-assisted security service should establish which model the provider can access and what evidence it returns for each finding. Access to a specialist model cannot substitute for a reproducible vulnerability report or a tested patch.
For general application teams, the immediate decision is narrower. Test public 3.8 Flash on expensive failures, compare cost per accepted result, and include January’s rates in the calculation. Adopt it where fewer corrections or more completed tasks justify the cost after the discount expires.
Sources
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google
- Gemini Developer API pricing — Google AI for Developers
- Fairwind Program — Google DeepMind