AI Articles

Google's Nano Banana 2.1 lowers image output prices and targets editing, text rendering, and infographics, with benchmarks and four tests against Nano Banana 2.

Gemini 4 Argon leads Google's benchmark table on knowledge work and long context, while Claude Opus 5.5 holds terminal coding and won our hands-on test.

Muse Wonder Full leads Meta's DeepSWE and SWE-Marathon comparisons, while regular Wonder brings a 1M context window to the API at 4/20 per million tokens.

Same list price, different bills: Haiku 5.5 leads on agentic benchmarks, while GPT-6 Luna matched its accuracy in our test for about half the cost.

Mistral’s trillion-parameter open-weight model leads on cybersecurity and legal benchmarks.

Explore OpenAI’s latest AI-generated mathematics results, the biggest discoveries, how the proofs were verified, and what they could mean for mathematicians and scientific research.

Learn what embodied AI is, how it differs from LLMs, how the perception-reasoning-action loop works, and why sim-to-real training matters for data scientists.

Anthropic's new small model costs $0.10 / $0.50 per million tokens, closes much of the gap to Sonnet on agentic work, and matched Haiku 4.5's perfect score in our test for a fraction of the cost.

Learn what tabular foundation models are, how in-context learning works, and when TabPFN or TabICL beat XGBoost on structured data, with a decision framework.

Reflection AI's Beam matches GLM 5.2 on reasoning with 3-4x less compute, but trails Kimi K3 on agentic coding. Here's how Reflection AI's open-weight model performs.

Aleph Alpha has released Kolibri 1, a 78B mixture-of-experts model with 3.46B active parameters, Apache 2.0 weights, and a 1M-token context, trained from scratch in Germany.

Google's table puts Gemini 4 Argon ahead of GPT-6 Astra on legal, finance, and long-context work at a fifth of the price. Astra still leads hard coding and is the one you can use.