Big Tuesday for AI: New Models Unveiled by Google and Anthropic

Today belongs to artificial intelligence twice over. Google has expanded the Gemini family with two specialized models, and Anthropic has deployed the new Claude Sonnet 5 model. Let’s take a look together at what each new feature brings and where you can try it out – most of them are available to you today.

Google Accelerates Images and Video Editing

The first addition is Nano Banana 2 Lite, according to Google, the company’s fastest and cheapest model for image generation. It replaces the original Nano Banana and targets situations where volume and speed are key. And this is where the difference is significant: in one minute, Lite can generate 21 images according to Google, while the more powerful Nano Banana 2 can only generate three. One 1K resolution image costs just 0.034 dollars.

The second new feature is Gemini Omni Flash, with which Google opens its video model to developers. Gemini Omni itself – capable of creating and conversationally editing videos, which we showed you earlier – has so far primarily targeted regular users via the app and YouTube. Flash is now the version that the company sends to developers via the Gemini API and Google AI Studio, so they can integrate it into their own applications. This is a public preview with clips up to ten seconds (longer footage is expected later), and output costs 0.10 dollars per second.

Claude Sonnet 5 is a More Capable Default Model

On the same day, Anthropic deployed Claude Sonnet 5 and immediately made it the default model for everyone – including those on the free tier. It focuses mainly on agentic tasks: the model plans itself, uses tools like a browser or terminal, and can complete multi-step assignments where older versions often stopped halfway.

In tests published by Anthropic, Sonnet 5 significantly outperforms Sonnet 4.6 and is hot on the heels of the flagship Opus 4.8. In agentic programming (SWE-bench Pro benchmark), it achieved 63.2% against 58.1% for Sonnet 4.6; in the Terminal-Bench 2.1 test, it managed 80.4%; and in computer control (OSWorld), 81.2%. The most telling is the GDPval knowledge test, where with a score of 1618 points, it slightly surpassed even the more expensive Opus 4.8 with 1615 points. However, these are figures from Anthropic’s own tests – independent comparisons are yet to come.

For regular users, however, availability is more important: Sonnet 5 is the default model for free and paid tiers; you can find it in Claude Code and via API. Developers will pay a discounted 2 dollars per million input tokens and 10 dollars per million output tokens until the end of August; later, the price list will increase to 3 and 15 dollars, respectively. Anthropic also claims that the new version “hallucinates” less and is less subservient to the user than Sonnet 4.6.

Will you opt for the new Google models, or does Claude Sonnet 5 suit you better?

Sources: Google, Anthropic