Google Gemini Pro API Available Through AI Studio

Original Article ↗ Hacker News Discussion ↗

Google has opened its Gemini Pro large language model to developers via AI Studio and a generous free tier, prompting interest in its pricing, features, and how easily it can replace or complement OpenAI’s GPT APIs. Commenters note attractive limits (up to 60 free queries per minute), multimodal support, and function calling, but many report poor UX, regional access restrictions, and performance roughly comparable to GPT‑3.5 rather than GPT‑4, especially on reasoning tasks. There is also concern about Google using free-tier data to improve models, potential future price hikes, and the broader impact of tools like auto‑generated blog post creators on web quality.

Access & Availability

  • Gemini Pro is available via Google AI Studio / MakerSuite and through Vertex AI, with support for many countries and languages.
  • Several users report “Access restricted,” “internal error,” or Early Access permission issues, even when admin settings appear correct.
  • Some can only access MakerSuite via a US VPN; Gemini is reportedly not yet available in the EU, and Bard there may still run on PaLM 2.
  • Firefox users see large numbers of blocked popups on the AI Studio / MakerSuite site.

Pricing & Free Tier

  • Free tier allows up to 60 queries per minute, which many see as very generous and a strong incentive to experiment.
  • Suspicion that this is a loss-leader, with worries about later price hikes similar to Google Maps.
  • Free-tier data is used to improve the product.
  • Paid pricing is per character; rough comparisons suggest overall cost parity with GPT‑3.5 Turbo when accounting for characters vs tokens.

Model Quality & Comparisons

  • Multiple simple reasoning puzzles (books in a house, siblings, orange under plate) expose frequent logic errors from Gemini Pro, roughly comparable to GPT‑3.5 and weaker than GPT‑4 and some newer open models in these tests.
  • Gemini Pro vision makes notable mistakes (e.g., misidentifying a thumbs-up as thumbs-down with grandiose explanation).
  • Several users describe Bard/Gemini as underwhelming; enthusiasm is mostly about price and access, not capability.
  • Gemini Ultra (positioned by Google as GPT‑4 competitor) is not yet released.

UX & Onboarding

  • Common complaints about confusing flows between AI Studio, MakerSuite, and GCP/Vertex, broken back buttons, popups, and API key creation errors.
  • Broader criticism of Google’s cloud UX, language/locale handling, and organizational incentives that deprioritize onboarding polish.

API Design & Integration

  • API is seen as an easy swap for OpenAI’s chat API: “messages” → “content,” roles slightly different (no explicit system role), plus built‑in moderation controls.
  • Function calling, embeddings, semantic retrieval, and multimodal inputs are available; Qdrant has example docs for embeddings.
  • Some confusion over minor differences between Node.js and Web SDK examples; clarified as documentation/overload issues.

Privacy & Ads

  • Concern that Bard/Gemini queries immediately influenced personalized ads across Google properties; contrasted with ChatGPT usage not doing so (as observed by one user).

Content & Web Quality

  • Example “blog post creator” prompt template seen as legitimizing AI‑generated web spam.
  • Counterpoint that structured text generation (e.g., press releases) can be a valid use, though some prefer raw fact lists over filler prose.

Branding & Naming

  • Some note the name collision with the existing Gemini internet protocol; others argue “Gemini” is generic and Google’s usage references the space program and internal team structure.