Gemini 2.5

Original Article ↗ Hacker News Discussion ↗

Marketing, positioning, and versioning

  • Many see the announcement as following a now-standard template: “state-of-the-art,” benchmark charts, “better reasoning,” with diminishing excitement due to frequent, incremental releases.
  • The “2.5” naming sparks debate: some see it as pure marketing/expectation management; others argue .5 implies a substantial but not architectural jump (e.g., coding gains, Elo jumps).
  • Comparisons are viewed as selective: Google benchmarks against o3-mini rather than o1 or o3-mini-high, which some interpret as biased.

Pricing, rate limits, and “experimental” status

  • Model is available free in AI Studio/Gemini with low rate limits (e.g., ~50 requests/day, low RPM), which makes it hard to adopt as a daily driver or for large experiments.
  • “Experimental” label implies different privacy terms and permission to train on user data; some note that previous “experimental” models never graduated.
  • Lack of published pricing at launch frustrates people who want to plan production use.

Benchmarks, long context, and multimodality

  • Long-context scores (e.g., MRCR and Fiction.LiveBench) impress many; several report this as the first model that can reliably reason over 200k+ token inputs (e.g., 1,000-poem corpus, entire Dart codebase).
  • Some caution that Google has historically excelled only on its own long-context benchmarks and underperforms on others like Nolima/Babilong; they want independent confirmation.
  • Multimodal demos (video shot-counting, cricket match analysis, OCR from Khan Academy videos, slide-deck reconstruction from webinars, SVG/image generation) are seen as genuinely strong.

Reasoning performance and puzzles

  • Multiple users test it on hard logic/maths puzzles (e.g., a three-number hat-style riddle, prisoners-type problems); Gemini 2.5 often succeeds where other frontier models fail or loop.
  • Skeptics note that at least one flagship riddle is on the public web, so success may involve some training-data recall plus reasoning, not “pure” generalization.
  • Others share failures: incorrect physics explanations, broken interpreters with bogus “tail-call optimization,” and game-playing agents that still hallucinate environment state.

Coding and engineering use

  • On the Aider Polyglot leaderboard, it sets a new SOTA (73%), with especially large gains in diff-style editing; format adherence is still weaker than Claude/R1 but recoverable with retries.
  • Users report:
    • Finding a subtle bug in a ~360k-token Dart library.
    • Strong performance on engineering/fluids questions and multi-language coding.
    • But also serious tool-calling problems and infinite-text loops in some agent setups where Claude/OpenAI/DeepSeek work fine.

Guardrails, policy, and hallucinations

  • Google’s guardrails are seen as stricter than rivals: earlier refusal to answer benign questions (e.g., US political modeling, C++ “unsafe for minors”) still colors perceptions, though some note gradual relaxation.
  • In search “AI overviews,” older Gemini variants have produced egregiously wrong answers, reinforcing trust issues.
  • Political/election-related queries are sometimes blocked entirely, unlike other labs.

UX, integration, and workflows

  • Many feel Gemini’s raw model quality is catching up or leading in specific areas (long context, cost/performance), but UX lags OpenAI: weaker desktop/mobile polish, missing shortcuts, clunky message editing, and less seamless IDE integration.
  • Some users have moved significant workflows to Gemini (especially long-document analysis and RAG-like internal tools) and say it can substitute for junior analysts; others still see it as “backup” to ChatGPT or Claude.

Privacy and data retention

  • Consumer Gemini terms explicitly allow human reviewers (including third parties) to see and annotate conversations, stored for up to three years even if activity is “deleted,” albeit de-linked from the account.
  • This alarms some, especially around sensitive/business data; others note that paid tiers and API usage can avoid training-on-input, similar to other major providers.

Economic impact and industry race

  • Discussion touches on whether continual benchmark gains are translating into measurable productivity or GDP growth; consensus is that effects are real but hard to quantify and not yet visible in macro stats.
  • There’s ongoing debate over whether AI will displace many white-collar workers versus creating new roles, with some arguing that UX and workflow integration are currently a bigger bottleneck than raw model capability.