Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Original Article ↗ Hacker News Discussion ↗

Google’s release of the Gemini 3.6 Flash and 3.5 Flash-Lite models is seen as a modest, speed-focused upgrade rather than a frontier breakthrough, with many noting that performance is roughly on par with popular rivals like GLM 5.2 while often being pricier. Commenters highlight that the Flash and Flash-Lite tiers can be very effective for cheap, high-volume, multimodal or agentic workloads, yet express frustration at Google’s frequent model deprecations, confusing product stack, and rising costs that make long-term reliance risky. The absence of a competitive new “Pro” or frontier model, combined with stronger open-weight and Chinese offerings, fuels concern that Google is ceding the high end of the AI market even as it doubles down on fast, scalable models for its own ecosystem and search.

Overall sentiment

  • Many see the release as incremental and “proof-of-life” rather than frontier‑moving.
  • Some are impressed by speed and practical usability; others call it underwhelming versus recent US and Chinese models.
  • Disappointment is amplified by the continued absence of Gemini 3.5 Pro / frontier models.

Benchmarks & intelligence

  • Benchmarks in Google’s post are criticized for being mostly self‑comparisons (against past Gemini) instead of against Fable, GPT‑5.6, GLM 5.2, DeepSeek, Kimi, etc.
  • Artificial Analysis and other custom benchmarks generally place 3.6 Flash around Sonnet‑5‑High / Opus‑4.8‑Medium territory, roughly comparable to GLM‑5.2, below top frontier models.
  • Some independent test harnesses and “pelican/MacBook SVG” style tests find 3.6 Flash surprisingly strong; others say it matches 3.5 Flash overall, just trading off strengths.

Speed, token efficiency & pricing

  • 3.6 Flash is repeatedly described as “very fast,” sometimes the fastest model on certain leaderboards.
  • Google claims higher token efficiency and lower cost per task vs 3.5; several users’ own benchmarks find it less token‑efficient, so real‑world costs can be flat or higher.
  • Flash‑Lite prices have increased sharply over generations; people running price‑sensitive or high‑volume workloads feel “boiled” as older, cheaper models get deprecated.
  • Multiple comments argue simple $/token comparisons are misleading; “cost per task” and verbosity matter.

Strengths & real‑world use

  • Flash/Flash‑Lite are praised for:
    • Low latency and stable, predictable tool use in long‑running agents.
    • Good multimodal performance, especially for image understanding and non‑English proofreading.
    • High value for cheap classification, extraction, and document processing tasks.
  • Some prefer Gemini for “Google this for me” / knowledge tasks due to Search and Knowledge Graph integration.

Weaknesses & regressions

  • Coding and agentic coding are seen as Gemini’s weak spots; several report 3.6/3.5 Flash or 3.5 Flash‑Lite breaking existing tool‑calling workflows or producing fragile code.
  • Many say Claude/Codex/other tools remain clearly superior for serious coding work.

Product, ecosystem & trust

  • Heavy criticism of:
    • Confusing branding (Flash vs Flash‑Lite vs Nano, Antigravity variants, consumer vs Workspace vs enterprise).
    • Abrupt deprecations and pricing changes that break production workflows.
    • Weak or awkward developer tooling (especially Antigravity and loss of older CLI).
  • Some enterprises still choose Gemini for compliance, geography, and “big vendor” comfort despite technical disadvantages.

Competition & strategy

  • Strong sense that Google is ceding frontier leadership to OpenAI/Anthropic and increasingly to Chinese open‑weight labs.
  • A common hypothesis: Google is optimizing for cheap, fast models to power Search and mass‑scale consumer features, not to win coding/frontier benchmarks.