Gemini 3.0 Flash (Skyhawk CKPT Tested): Cheap & Awesome Coding Model by Google is COMING!

Key Concepts

• Gemini 3.0 Pro: Google's latest advanced AI model, noted for its improved performance and cost-effectiveness compared to previous versions.
• Gemini 3.0 Flash: An upcoming, faster, and cheaper version of Gemini, expected to be released soon.
• Nano Banana Pro: A new model capable of reasoning before image generation, built on Gemini 3.0 Pro.
• LM Arena: A platform where users can test and compare various AI models, including early versions of Gemini 3.0 Flash (Skyhawk and Sea Hawk).
• King Bench: A set of 11 questions used to evaluate the performance of AI models.
• Omni Models: AI models capable of processing multiple input modalities, such as video and audio, for live reasoning.
• Open Source Models: AI models that are publicly available and can be modified, such as Devstrol and GLM4.6V.
• Devstrol: A new open-source model that shows promise for its size and cost, potentially rivaling models like Minimax.

Gemini 3.0 Pro and Nano Banana Pro Launch

Google has recently launched Gemini 3.0 Pro, described as a significantly improved model that is also more cost-effective than models like Sonnet. It is currently in preview but demonstrates substantial advancements over Gemini 2.5 Pro. Alongside this, Nano Banana Pro was introduced, notable for being one of the first models capable of reasoning before generating images, and it is based on Gemini 3.0 Pro.

Gemini 3.0 Flash: Upcoming Release and LM Arena Testing

The Gemini Flash models have been popular due to their strong capabilities at a lower cost. Gemini 2.0 Flash and 2.5 Flash were highly effective, with the original Nano Banana model being based on 2.5 Flash. It is anticipated that Gemini 3.0 Flash will be released soon. Early versions, referred to as Skyhawk and Sea Hawk, have appeared on LM Arena. These are believed to be early checkpoints of Gemini 3.0 Flash. Users can access these models on LM Arena by sending a message and waiting for a random assignment to Sea Hawk or Skyhawk.

King Bench Performance Evaluation of Skyhawk (Gemini 3.0 Flash Variant)

The transcript details a test of the Skyhawk model on King Bench, a set of 11 questions. The results were mixed:

• Floor plan in 3JS: "Not very good. It is fine but it is not anything extraordinary."
• SVG panda eating a burger: "Quite interesting. Looks kind of good, but it looks more like an icon rather than a full image, but it is interesting."
• Chessboard with autoplay feature: "Just fails in this. The code is very wonky. The design is very wonky, and it's not a good generation."
• Minecraft in 3D generation: "Kind of good. More of a map rather than a playable game, but it is indeed functional and kind of well-made with the Kandinsky style intact."
• Majestic butterfly flying in the garden: "Quite good. The wings are somewhat wonky, but overall, this looks actually good."
• CLI tool in Rust: "Fine, but nothing extraordinary, but it worked."
• Blender script for a Pokeball: "Pretty fine."
• Riddle: "Solved well."
• Math questions (two): "Both a fail."

Overall, Skyhawk scored "just below the Caterpillar model which was GPT 5.1 and also a bit below 4.5 sonnet." The assessment is that while not a "breakthrough," it can be "really useful nonetheless."

Cost-Effectiveness and Capabilities of Flash Models

The Flash models (specifically 2.5 Flash and 2.0 Flash) are highlighted for their low cost and strong vision and tool-calling capabilities.

• 2.5 Flash: Costs approximately $0.3 per million input tokens and $2.5 per million output tokens.
• 2.0 Flash: Previously cost only $0.10 for input and $0.40 per million output tokens. These models also offer a "very generous free tier in the API."

A significant advantage of Flash models is their live capabilities. They are often omni models, meaning they can process live video and audio streams, enabling reasoning across this information. This "live reasoning" is considered an "underrated option."

Expectations for Gemini 3.0 Flash and Future Models

The expectation is that Gemini 3.0 Flash will be released first, followed by a "flashlight" version. The hope is that Gemini 3.0 Flash will address some of the hallucination problems observed in Gemini 3.0 Pro and improve its performance on long horizon tasks and complex tool calling. If Gemini 3.0 Pro's front-end capabilities can be distilled into 3.0 Flash, it could become an "awesome cheap model just for front end."

Currently, Opus is considered unmatched in its capabilities and price point for front-end tasks. The transcript suggests that a Gemini Ultra model, potentially with an "Opus-like mode," would be beneficial. Google has provided Ultra members access to Gemini Deep Think, described as their "O3 Pro."

Open Source Advancements

The open-source AI landscape is also progressing with models like Devstrol and GLM4.6V. Devstrol is noted for its potential to offer capabilities similar to Minimax at a lower cost and is currently free on its API.

Nano Banana Flash and Image Generation

It is anticipated that Nano Banana Flash will also be released, likely incorporating image generation capabilities, as image leaks have begun to appear.

Conclusion and Call to Action

The overall sentiment is that the new Gemini models, particularly the upcoming Flash versions, are promising, especially considering their cost-effectiveness. The transcript encourages viewers to share their experiences with Skyhawk/Sea Hawk on LM Arena and to express interest in videos about Devstrol. The creator also mentions options for supporting the channel through Super Thanks or channel memberships.