Lots of information only exists on web pages, APIs don’t provide it. Traditional scraping can’t handle JS-rendered pages, and anti-scraping keeps getting stricter. AI browser automation can execute JS, fill forms, and understand page content.
On 13 February 2026 Google released Gemini 3 Deep Think. The vendor ARC-AGI-2 score is 84.6%, 0.4 points under the ARC Prize 85% line. That is test-time compute, not a knowledge lookup.
ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) is one of the benchmarks closest to testing “general intelligence.”
It doesn’t test knowledge—it tests the ability to solve new problems under unknown rules. Give a visual puzzle, and the model must infer the transformation rule and apply it to new figures.
Google released Gemini 3’s Deep Think dedicated reasoning mode, scoring 84.6% on ARC-AGI-2.
Early 2026, 17B parameter small models can replace large models in many scenarios.
Llama 4 Scout (17B) coding capability approaches GPT-4o, at zero cost.
Gemini 2.0 Flash Thinking is enough for simple-to-medium coding tasks; complex reasoning still lags Claude 3.7 and o3. The real edge is price: $0.1/M input tokens. No controlled eval, so no letter grades.
Anthropic drops Claude Opus 4.6.
Official positioning: “most capable model yet”—but the real new thing isn’t model parameters. It’s Agent Teams.
RAG (Lewis et al., 2020) splits parametric memory from a swappable index: retrieve only when the window cannot hold the corpus. Current Claude windows are in Anthropic context windows—1M by default on Fable / Opus / Sonnet 5, 200k on Haiku 4.5. If count_tokens says the bundle fits and the task needs cross-section reasoning, stuff it. Index when the corpus is larger than the window, or when documents must be replaced without resending the rest.
