SYLIN

Gemini Flash

by Google · Model · Language & code models

strong

Gemini 3.8 Flash ranks near the top for image and video understanding

Gemini 3.8 Flash released; beats Opus 5 and GPT 5.6 Sol on finance, legal and agentic terminal coding, state of the art in chart and long-video understanding

Why we say so

Arena's vision leaderboard puts 3.8 Flash third twice (1315, then 1317), behind only Claude Fable 5 (1326) and Claude Opus 5.5 (1324), a small gap. AI Search uses it as the baseline that Qwen 3.8 Omni Flash and GPT-6 Luna are compared against, and reports Qwen beats it on audio at a lower price. Google's claim that Gemini 4 Hydrogen is 2.5x faster is a maker claim about a different model.

Versions

VersionReleasedRank
Gemini 3.8 FlashLATEST
strong Third on Arena vision, close behind the Claude models
Sep 2
Gemini 3.7 FlashAug 16
Gemini 3.6 FlashJul 21

Evidence

  1. Nov 9
    #3 on Arena vision (1317)
  2. Oct 19
    #3 on Arena vision (1315)
  3. Sep 27
    GPT-6 Luna scores below Gemini 3.8 on the intelligence index
  4. Sep 6
    Gemini 3.8 Flash released, third Flash model in six weeks, via Google AI Studio, Gemini API, Antigravity and Gemini app
  5. Sep 6
    Ranked number one on DeepSWE for long-horizon software engineering at much lower average cost, but uses by far the most output tokens
Show all 19
  1. Sep 6
    Artificial Analysis index ranks it one point below open GLM 5.3 and Kimi K3 and a few below frontier Claude and GPT; about 348 tokens/s
  2. Sep 6
    Ranked eighth on Arena but outside the top 10 on LiveBench and below Gemini 3.7 Flash, suggesting benchmark overfitting
  3. Sep 5
    Commenters say Gemini 3.8 Flash ranks too high; one reports it gets confused about what it did, another doubts it beats Astra or Fable at code
  4. Sep 4
    Driven by Hermes Agent, found and fixed a planted one-line sort bug in a Docker/Postgres/FastAPI stack in 2m12s, with self-verification
  5. Sep 2
    On expert-level reasoning benchmark (HLE-Verified), 3.8 Flash narrowly edges out frontier models; Google reports 54.9%
  6. Sep 2
    Reported top of the DeepSWE leaderboard, ahead of Opus 5
  7. Aug 16
    Gemini 3.7 Flash tops a frontier-code benchmark among small models; 340 tokens/s; in Antigravity, AI Studio, Android Studio, Gemini app
  8. Aug 16
    Beats GPT Luna Max on AA index but costs 40 cents per task versus 5 cents for Luna
  9. Jul 26
    Google released Gemini 3.6 Flash as the workhorse Flash upgrade, in Gemini API, AI Studio and Antigravity
  10. Jul 26
    Artificial Analysis ties 3.6 Flash with 3.5 Flash at slightly lower cost, still pricier than GPT Luna and GLM 5.2
  11. Jul 22
    Claimed to beat 3.5 Flash on DeepSWE, OSWorld and MLE-Bench while using fewer tokens per task
  12. Jul 22
    Does not claim to beat 3.1 Pro on the hardest long-context reasoning tasks
  13. Jul 22
    Via Hermes Agent it found and fixed an unseen berth-depth bug in a FastAPI app without hints
  14. Jul 22
    Reviewer thinks its coding quality does not match Fable 5 or Opus, though results were still good

Alternatives

Claude Opus#2 on Arena vision (1324)
Claude Fable#1 on Arena vision (1326)
GPT AstraAiBenchy SVG showcase comparison cited as strong SVG generation against Claude Fable 5.1 and Gemini 3.8 Flash