Google Gemini 3 Makes a Huge Leap on the ARC‑AGI Benchmark

Tue, 25 Nov 2025 09:00:00 -0700

Google’s Gemini 3 has posted a standout result on the ARC Prize Leaderboard – ARC‑AGI‑1, scoring about 87.5% in its Deep Think preview — a very strong showing on a benchmark focused on abstract reasoning. ARC‑AGI‑1 (Abstraction & Reasoning Corpus for AGI) is designed to test fluid intelligence: each task provides a few input/output examples, and models must infer underlying rules rather than rely on memorization. The benchmark emphasizes skill-acquisition efficiency, rewarding reasoning and generalization over brute-force performance. (ARC Prize Technical Report)

This 87.5% score suggests Gemini 3 Deep Think is effectively reasoning with abstract rules, not just pattern matching, and positions it as a model capable of structured, AGI‑style problem solving. Beyond ARC‑AGI‑1, Google has publicly confirmed that Gemini 3 Deep Think achieves 45.1% on ARC‑AGI‑2, a more challenging follow-up benchmark that also tests reasoning and code execution skills. (Google Blog, VentureBeat)

If these results hold, they represent a meaningful step forward: Gemini 3 is not just a larger LLM but a reasoning-first model capable of abstract problem solving. High performance on ARC‑AGI‑1 signals efficient learning and generalization — core aspects of intelligence that many benchmarks don’t test — and marks a clear signal that AI systems are beginning to handle tasks previously out of reach for conventional models.

Google Quietly Ruins Gemini’s Free Tier

Thu, 17 Jul 2025 09:00:00 -0600

Google has quietly chopped the Gemini API free tier down to size—and now it’s about as useful as a screen door on a submarine. Once upon a time, you could fire off 1,500 requests per day, but now Gemini 2.0 Flash free tier is capped at 200 requests per day (Google Rate Limits “Free Tier” table; previously 1,500/day according to Zapier: "Free up to 15 requests/min; 1M tokens/min; 1,500 requests/day")

Meanwhile, if you switch to Gemma—which on paper allows 30 requests per minute and a whopping 14,400 requests per day (Google Rate Limits “Gemma 3 & 3n” row)—you’ll still hit a ResourceExhausted error long before you approach those limits (Google Support).

I doubt I’m even burning through as many requests as Google claims. For example, my test script throws “ResourceExhausted” after just a handful of calls—despite staying well under the documented RPM and TPM caps. That means prototyping or smoke‑testing your app without buying tokens is now effectively impossible. Want to validate your UX? Better hand over your credit card. Moreover, you'll hit this Gemma limit even if you have a paid tier!

For anyone chasing budget‑friendly LLM options, it might be time to head back to Hugging Face or other open‑source alternatives. Because if Google won’t let you play in the sandbox for free, you’ll have to build your sandcastle elsewhere.

Other Related Links:

Mindfire Technology

Google Gemini 3 Makes a Huge Leap on the ARC‑AGI Benchmark

Google Quietly Ruins Gemini’s Free Tier

Google Gemini 3 Makes a Huge Leap on the ARC‑AGI Benchmark