Gemini 3.8 Flash
| Maker | |
|---|---|
| Family | Gemini Flash |
| Released | 2026-09-02 |
| Context | 1M |
| Max output | 65.5K |
| Open weights | Closed |
|---|---|
| Modalities | text, image, video, audio, pdf → text |
| Input / 1M | $0.75 |
| Output / 1M | $3.75 |
| Cached input / 1M | $0.075 |
| Review call | $0.0287 |
|---|---|
| Reasoning efforts | low, medium, high, xhigh, max |
| Hosts | Google Gemini, OpenRouter |
| Model id | gemini-3.8-flash |
- Rank
- #1=
- of 17
- Bugs caught
- 68%
- 54%–81%
- False alarms
- 1.0
- per clean PR
- Cost / review
- $0.0862
- Cost / bug
- $0.1166
- Time
- 427s
- per review
Gemini 3.8 Flash costs $0.0862 per review and catches 68% [54%–81%] of the known bugs. For that money there are better seats, so this is a model to skip unless you have a specific reason to want it.
It shares rank 1 of 17 only because its interval is wide: the run can't tell it apart from GPT-6.1 Sol or Gemini 4 Argon on recall, yet it measurably beats 3 ranked models where Sol beats 7 and Argon 6. Both cost less per review: a Gemini 3.8 Flash review costs 2.1× a Sol review and 1.6× an Argon review.
The list rate ($0.75 / $3.75 per 1M tokens) looks modest. The measured bill is higher because its reviews run long: it wrote several times the output the spec sheet's review-call price assumes, and a review took 427s. A review on GPT-6 Luna took 29s, a Gemini 3.8 Flash review costs 30× a Luna one, and the two cannot be told apart on recall.
Its one good trait is restraint. It averaged 1.0 [0.6–1.5] false alarms per clean PR, and 94% of its catches landed on the defect's own lines. If you must keep a Google model in a panel, that is the case for it. If you only want a second maker's view, Claude Haiku 5.5 gives one for less: a Gemini 3.8 Flash review costs 23× a Haiku one.
See the head-to-head with Luna for the bug-by-bug differences.
Figures follow the latest benchmark run and price check. Text checked Nov 5, 2026.
| Host | Model id | Input / 1M | Output / 1M |
|---|---|---|---|
| Google Gemini | gemini-3.8-flash | $0.75 | $3.75 |
Source: Google pricing, Oct 16, 2026
By kind of bug
- Logic 14
- 82%
- Security 9
- 67%
- Concurrency 8
- 79%
- Error handling 5
- 30%
- API misuse 3
- 57%
- Resource handling 3
- 50%
Bug by bug
- Known bugs
- 42
- Caught every run
- 26
- Caught sometimes
- 5
- Missed
- 11
- Clean PRs
- 13
- False alarms on them
- 18
| Result | Pull request | Language | Kind |
|---|---|---|---|
| Caught sometimes7/10 | 4.x: Enhance Streamable/Streamer and DisposableContainer APIsReactiveX/RxJava | Java | API misuse |
| Caught sometimes1/2 | Event Thread could stall when not notified of new queries on existing connections that are in a bad statec-ares/c-ares | C | Resource handling |
| Caught sometimes2/4 | Fixed #36540 -- Updated request.auser() in contrib.auth.alogin() and contrib.auth.alogout().django/django | Python | Logic |
| Caught sometimes1/2 | [12.x] Colocate Container build functions with the `Buildable` interfacelaravel/framework | PHP | Error handling |
| Caught sometimes1/3 | Refactor: MudDebouncedInput, ThrottleDispatcher and DebounceDispatcherMudBlazor/MudBlazor | C# | Concurrency |
| Missed | feat(v4): parse input containing reference cyclescolinhacks/zod | TypeScript | Resource handling |
| Missed | rpc : cache and reuse compute graphsggml-org/llama.cpp | C++ | Security |
| Missed | perf(gitdiff): optimize inline diff highlighting using cachego-gitea/gitea | Go | Error handling |
| Missed | Store cached child els in Attributes, not fieldjhy/jsoup | Java | Concurrency |
| Missed | Fix DataClassEq comparatorkotest/kotest | Kotlin | Logic |
| Missed | Refactor eventually to ensure wall-clock time usagekotest/kotest | Kotlin | Error handling |
| Missed | KTOR-9171 Redesign ByteReadChannel.readUTF8Line APIktorio/ktor | Kotlin | Logic |
| Missed | Add core API to enable Kotlin singleton mockingmockito/mockito | Java | API misuse |
| Missed | feat: add configurable maxPayloadSize for WebSocketnodejs/undici | JavaScript | Security |
| Missed | reorganize nushell configurationnushell/nushell | Rust | Error handling |
| Missed | [Security] Allow multiple OIDC discovery endpointssymfony/symfony | PHP | Security |
| Version | Released | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| Gemini 3.6 Flash | 2026-07-21 | $0.75 | $3.75 | 1M |
| Gemini 3.5 Flash | 2026-05-19 | $1.50 | $9.00 | 1M |
| Gemini 2.5 Flash | 2025-04-17 | $0.30 | $2.50 | 1M |
| Gemini 3.7 Flash | 2026-08-13 | $0.75 | $3.75 | 1M |
| Model | Maker | Input / 1M | Output / 1M | Review call | Context |
|---|---|---|---|---|---|
| GPT-5.4 Mini | OpenAI | $0.75 | $4.50 | $0.0305 | 400K |
| Qwen3.8 27B (Groq) | Groq | $0.80 | $4.00 | $0.0306 | 131.1K |
| Grok Build 0.1 | xAI | $1.00 | $2.00 | $0.031 | 256K |
| Step 5 Preview (OpenRouter) | StepFun | $1.00 | $2.70 | $0.0327 | 1M |
Updated Nov 5, 2026