AI Model Updates

Every model update.
What it means for your prompts.

We track every major release across image, video, text, and audio AI models and update HonePrompt's prompting engine within days of each release.

March 2026

Mar 6, 2026
Nano Banana Pro
● High impact

Nano Banana 2 launches with Google's official prompting guide

Google officially released Nano Banana 2 (Gemini 3.1 Flash Image) alongside a comprehensive prompting guide from the DeepMind team. Key structural changes affect how prompts should be written.

What changed for prompting
Verb-first formula is now official: [Subject doing Action] in [Location]. [Composition]. [Lighting]. [Style]. Start with a strong verb.
Remove quality spam. Tags like "4K", "masterpiece", and "trending" are actively ignored or misinterpreted. Delete them.
Named real light sources outperform abstract descriptions. "Golden hour sun from the left" beats "dramatic lighting" every time.
Edit mode: describe changes conversationally. "Change the jacket to burgundy leather. Preserve everything else." No masking needed.
HonePrompt updated: Nano Banana Pro system prompt fully rebuilt with these findings.
Hone a Nano Banana Pro prompt →

February 2026

Feb 4, 2026
Kling AI
● High impact

Kling 3.0 adds multi-shot generation, native audio, and 15-second videos

Kling 3.0 is a category-defining release. The Omni Model architecture processes text, image, and audio simultaneously, enabling coherent multi-shot videos with synchronized dialogue in a single generation.

What changed for prompting
Multi-shot prompting: describe up to 6 labeled shots in one prompt. No more stitching clips together manually.
Native audio: add an AUDIO section to every prompt. Characters can speak with specific voice tones using [Character A: description, voice] labels.
Motion intensity: numerical values from 0.1 to 1.0 now produce predictable results. Use 0.4 for subtle and 0.9 for high energy.
Temporal flow is required. Describe beginning, middle, and end. The model understands narrative arc.
Named light sources produce better results. "Neon signs", "candlelight", and "golden hour" outperform "dramatic lighting".
HonePrompt updated: Kling system prompt fully rebuilt for Kling 3.0.
Hone a Kling 3.0 prompt →

January 2026

Jan 22, 2026
Runway
● Medium impact

Runway Gen-4 Turbo becomes the production default

Runway promoted Gen-4 Turbo to the production default model. Gen-4 Turbo introduced reference image mode. When an image is uploaded, prompts should describe motion only, not appearance.

What changed for prompting
Reference image mode: describe camera movement and subject motion only. "Slow push in as she turns her head. Preserve all visual elements." Appearance is handled by the image.
Text-to-video: temporal arc is now mandatory. Describe the start state, what happens, and the end state.
Color grade references improved significantly. "Warm bleach bypass", "teal-and-orange Hollywood", and "cold blue clinical" all produce consistent results.
HonePrompt updated: Runway prompt rebuilt for Gen-4 Turbo with both modes.
Hone a Runway prompt →
Jan 8, 2026
Midjourney
● Minor update

Midjourney V7 style tuning improves photorealism

Midjourney pushed a series of model tuning updates to V7 improving adherence to --style raw and photorealistic prompts. No structural prompting changes required but stylization values needed recalibration.

What changed for prompting
The --s (stylization) sweet spot shifted lower for photorealism. Values between 150 and 250 now outperform 400 and above for editorial and portrait work.
Film grain and analog references improved. "Kodak Portra 400 grain" and "Fuji Pro 400H color cast" now produce much more accurate analog simulation.
Hone a Midjourney V7 prompt →

November 2025

Nov 12, 2025
Nano Banana Pro
● High impact

Nano Banana Pro (Gemini 3 Pro Image) launches

Google DeepMind launched Nano Banana Pro, the community nickname for Gemini 3 Pro Image, with a reasoning process that thinks through your prompt before generating. Immediate viral adoption driven by unprecedented photorealism and accurate text rendering.

What changed for prompting
Reasoning-first model. Elaborate, detailed prompts outperform short ones. The model handles complex multi-element scenes reliably.
JSON schema prompting is uniquely effective for this model. Structured key-value pairs produce more reliable results than prose alone.
Text in images is now reliable. Put exact text in quotation marks anywhere in the prompt.
Up to 14 reference images in a single prompt for style or character consistency.
Hone a Nano Banana Pro prompt →