benchr
Google made Gemini 3.8 Flash generally available on September 2, 2026, twenty days after Gemini 3.7 Flash. benchr rechecked the pricing table, the model page and the launch post on September 7. The comparison that matters here is unusual, because the two endpoints are not separated by price, context, or output limit. They are separated by behavior, and behavior is the part no rate card prices.
The rate card is the same one, to the cent
Every published lane matches its 3.7 Flash equivalent exactly. This is not an approximation from a summary table; it is what both pricing entries state.
| Lane | Through 31 Dec 2026 | From 1 Jan 2027 | Differs from 3.7? |
|---|---|---|---|
| Standard input | $0.75 | $1.50 | No |
| Standard output | $3.75 | $7.50 | No |
| Cached input | $0.075 | $0.15 | No |
| Cache storage / 1M / hour | $0.50 | $1.00 | No |
| Batch and Flex input | $0.375 | $0.75 | No |
| Batch and Flex output | $1.875 | $3.75 | No |
| Priority input | $1.35 | $2.70 | No |
| Priority output | $6.75 | $13.50 | No |
The input window is 1,048,576 tokens and the maximum output is 65,536 on both. A free tier is offered on both. So a migration cannot be justified or rejected on the price sheet, and a cost estimate that only swaps the model ID will return the same number for two models that do not behave the same way.
What Google changed is how much the model spends
The launch post is direct about it. Google writes that 3.8 Flash works harder
, and that on complex tasks it exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively
. Thinking is exposed at low, medium and high levels on the model page.
Read as a cost statement rather than a quality one, that sentence says the model will, by design, produce more output tokens and more tool calls for the same prompt. Output is priced at five times input on this rate card, and thinking tokens are billed as output. A workload that was comfortable on 3.7 Flash can therefore cost materially more on 3.8 Flash at identical published rates, and the increase will not appear anywhere except the token counter.
Google says as much in its own guidance: for applications where compute efficiency is the primary constraint, developers can use lower effort levels or continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads
. That is a vendor telling you the newer model is not automatically the right one.
One published number, and the ones that are not numbers
Google publishes a single figure for the general model at launch: 54.9% on HLE-Verified. That is a provider-reported result, not a benchr measurement, and benchr has not reproduced it.
The rest of the launch claims are comparative statements without published figures — that 3.8 Flash outperforms most larger frontier models on DeepSWE v1.1, and beats 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark. Those may well be true. They are not usable in a procurement comparison, because a claim with no number cannot be checked, ranked, or entered into a spreadsheet. benchr records the one figure and leaves the others as what they are.
The Cyber variant is not something you can plan around
The same announcement introduces Gemini 3.8 Flash Cyber, a specialized variant for vulnerability detection and automated patching, with 47.2% pass@1 on CWE-Bench and a stated real-world vulnerability discovery rate above 70%.
It is reachable only through Google's Fairwind Program, which is restricted to trusted government authorities, critical infrastructure operators and software maintainers. It has no entry in the API pricing documentation and no separate API model ID published there, so benchr carries no record for it. If you are not already in that program, treat it as unavailable rather than as a model you might adopt later.
What the endpoint does and does not do
The model accepts text, image, video, audio and PDF input and returns text. Function calling, structured outputs, code execution, file search, URL context, Search grounding, Grounding with Google Maps, caching, Batch, Flex and Priority inference are all listed, and computer use is listed as Preview.
Image generation, audio generation and the Live API are explicitly not supported for this model. If a stage of your pipeline produces images or audio, or holds a live session, that stage belongs on a different endpoint regardless of how the reasoning stage performs.
Choosing between 3.7 and 3.8
| If your constraint is | The record points at | Why |
|---|---|---|
| Cost per completed task, on work the current model already handles | Gemini 3.7 Flash | Google names it as the efficiency-first choice and keeps it fully supported |
| Multi-step agentic work that currently stalls or gives up early | Gemini 3.8 Flash | Extra reasoning steps and iterative tool calls are the stated change |
| A fixed per-request token budget | Gemini 3.7 Flash, or 3.8 at a low thinking level | Higher effort levels spend output tokens, which are billed at five times input |
| Generated images, generated audio, or a live session | Neither | Those output paths are not listed for either model |
| A budget that crosses into 2027 | Model both periods first | Every tier doubles on January 1, 2027 on both models |
What a pilot has to measure
Because the prices are identical, a token-rate comparison tells you nothing. The measurement that separates these two models is cost per completed task, not cost per token, and it needs the same representative cases run through both endpoints with the tool calls counted.
Record, per case: total input tokens, total output tokens including thinking, the number of tool calls, whether the task completed without human intervention, and the wall-clock time. A model that costs 40% more per task and completes 20% more of them is a different decision from one that costs 40% more and completes the same set. Neither of those outcomes is visible on the rate card, and neither is predictable from the launch post.
If a workload uses repeated prefixes, price context caching properly on both: the cache-hit rate is a per-model behavior, and the estimate has to carry cache storage at $0.50 per million tokens per hour rather than treating every hit as free.