Quick answer
Krea 3 image editing lets you select part of a picture, type what should change, and get back edited versions in seconds while the rest of the image stays the same. In Matt Wolfe’s hands-on test, edits took about 10 seconds and kept fonts, colours and untouched areas intact, which is why he rated it cleaner than ChatGPT’s image tool or Nano Banana for quick fixes. Krea says most edits finish in under 8 seconds at 2K and cost $0.06 per 2K image. The same week brought six new language models, including Meta’s Muse Wonder, Google’s Gemini 4 Argon and four open-weight releases, plus a quiet second week of OpenAI’s shipping pledge.
Key takeaways
- Krea 3: select a region, describe the change, pick from several variations. Best for small, precise edits such as thumbnails
- Open image models: Flux 3 Image Klein and Qwen Image 2.1 Turbo can both be tried free on Hugging Face
- Muse Wonder: strong benchmark scores and $4 / $20 per million tokens, but only middling in the creator’s own tests
- Gemini 4 Argon: $2 / $10 per million tokens until the end of November, and much cheaper per task than Opus 5.5
- Open-weight: Muse Spark Open, MAI Base 2, Nemotron 3.5 Super VL and Ultramarine 1 Flash, with very different levels of availability
- Also: two new decision models, Grok Voice 3.0, ElevenLabs instant soundtracks and AI moves at Xbox and Epic
In this guide
- Watch: the source video
- What is Krea 3 image editing, and is it really a Nano Banana rival?
- Which other image models launched this week?
- How good is Gemini 4 Argon in hands-on tests?
- Is Meta’s Muse Wonder really a top-tier model?
- Which open-weight language models arrived?
- What did OpenAI ship in week two of its 28-day pledge?
- What are decision models, and which new ones launched?
- What else happened in AI this week?
- What should small businesses do with this news?
- Frequently asked questions
- The bottom line
Most weeks, new image generators barely register any more. Most of them make good pictures. This week was different because the stand-out release was an image editor, not another generator.
Below we cover Krea 3 image editing first, then the six new language models, OpenAI’s latest updates and the smaller stories worth knowing about. Everything is split into what shipped, what the vendors claim and what is still unconfirmed.
🛠️ Practical guides
Want to act on this week’s news? These step-by-step guides answer two follow-up questions:
- How to Edit a Thumbnail With AI Without Redrawing the Whole Image — a step-by-step way to make precise, region-based edits to a thumbnail and check them before you publish.
- How to Build an AI Voice Support Agent That Only Answers From Your Website — how to scope, write, test and launch a voice agent that sticks to your own website content.
Watch: the source video
What this video covers: Matt Wolfe’s weekly AI news round-up leads with hands-on tests of Krea 3’s image editing, then covers Flux 3 Image Klein, Qwen Image 2.1 Turbo, Gemini 4 Argon, Meta’s Muse Wonder and Muse Spark Open, Microsoft’s MAI Base 2, Nvidia’s Nemotron 3.5 Super VL, Paper Instruments’ Ultramarine 1 Flash, OpenAI’s second week of daily shipping, new decision models, Grok Voice 3.0 and a rapid round of smaller news.
Video: Matt Wolfe (@mreflow). This summary and the opinions in it are our own. Vendor claims and unconfirmed details are labelled.
What is Krea 3 image editing, and is it really a Nano Banana rival?
Krea 3 is an image model from Krea that can generate pictures, but its real strength is editing. You upload an image, click to select an area, and type what you want changed. It then returns a few variations to choose from.
In the video, the creator edited one of his own old YouTube thumbnails. Here is what he tried:
| Edit requested | What happened |
|---|---|
| Make a phone’s background blue | Changed only the phone screen |
| Change text on the phone | New words appeared in the same font and colour |
| Swap a small logo, including one stitched on a cap | The new logo looked stitched into the fabric |
| Several edits at once | Returned two or three variations in roughly 10 seconds |
His verdict was that it was one of the best image editing experiences he has had, with cleaner edits than ChatGPT’s image tool or Nano Banana. That is one person’s hands-on opinion, not a benchmark. He also expects the bigger tools to catch up.
What does Krea say about Krea 3?
- Most edits come back in under 8 seconds at native 2K resolution.
- Pricing on Krea’s announcement is $0.06 per image at 2K and $0.03 at 1K.
- Products, characters and text are meant to stay consistent from one edit to the next.
- It is available on krea.ai and, according to Krea, through partners including fal, Replicate and OpenRouter. An API is in early access.
The video describes Krea 3 as a public beta. Krea’s changelog says it became the default image model in the Krea app on 16 October 2026, so the exact labels may change as it rolls out.
🔎 Unconfirmed / rumour: The creator said he believes Krea 3 may be released as an open-weight model eventually. Krea has not confirmed that in the material we checked, so treat it as a guess.
⚠️ Warning: Swapping in another company’s logo is easy with tools like this. It is also a quick way to mislead viewers or break trademark rules. Only add logos and brand names you have the right to use, and don’t edit thumbnails to promise something your video doesn’t deliver.
Want to try this on your own thumbnails? Our guide How to Edit a Thumbnail With AI Without Redrawing the Whole Image walks through a safe, repeatable process.
Which other image models launched this week?
Two more image models arrived, both with open weights, which means you can download and run them yourself:
| Model | What it is | How to try it |
|---|---|---|
| Flux 3 Image Klein (Black Forest Labs) | A smaller open-weight image model for generating and editing. On an Nvidia DGX Spark, the published figures show text-to-image in 7.3 seconds and editing in 16.2 seconds. An ordinary RTX 4070 graphics card is only about 1.5 seconds slower for text-to-image | Download the weights on Hugging Face, or use the free Hugging Face Space (shared, so it can be slow at busy times) |
| Qwen Image 2.1 Turbo (Alibaba) | A faster checkpoint of Qwen Image 2.1. You probably give up a little quality for speed. It can also edit images | Free demo on Hugging Face |
The creator’s honest take was that image generation is “fairly solved”. What makes these two interesting is that they run on consumer hardware and are open, not that the pictures are better.
💡 Tip: If you want to run Flux 3 Image Klein in a business, read the licence on its Hugging Face page first. Press coverage reports that free commercial use only applies below a revenue threshold, and larger organisations need a separate agreement.
How good is Gemini 4 Argon in hands-on tests?
Gemini 4 Argon was announced on 30 September, so the benchmarks were already public. We covered it in depth in our Gemini 4 Argon review. Here is what was new from this week’s hands-on tests:
- Price: $2 per million input tokens and $10 per million output tokens through the end of November, after which the price roughly doubles. Our Argon cost estimate guide shows how to plan for that.
- Intelligence vs cost: on Artificial Analysis, which combines many benchmarks, Argon scores 53, five points below Opus 5.5. But Opus 5.5 is the third most expensive model per task, while Argon is far cheaper for similar results.
- Creator’s drawing test: about $0.34, 34,000 tokens and around 4 minutes, and it ranked second on his AI-judged leaderboard.
- One-shot game test: a playable 3D game with camera control, jumping and sprinting, but plain-looking graphics.
- 3D printing: from a single front-facing photo, it produced a printable STL file of a bust that looked good straight on but weaker in profile.
Where to use it: through the API, or in the Gemini app on the paid Google AI Pro and Ultra plans.
Not sure which of these models to test first? Pick one real task you do every week, write down what “good” looks like, and run it on two models side by side. Our guide to AI model routing for small business shows how to match each task to the right model.
Is Meta’s Muse Wonder really a top-tier model?
Muse Wonder is Meta’s new flagship model. Some well-known reviewers have been very positive, with one placing it alongside Opus 5.5 at the top of his personal rankings. The creator was less convinced, describing it as “kind of okay” in his own use.
| What we know | Detail | Type of evidence |
|---|---|---|
| Coding benchmark (DeepSWE) | 79%, which Meta presents as the best among top-tier models | Vendor chart |
| Other benchmarks | State-of-the-art or on par on legal, financial-research and general intelligence tests | Vendor chart |
| Price | $4 per million input tokens, $20 per million output, about the same as Claude Opus 5.5 | Published pricing |
| Artificial Analysis index | Above Gemini 4 Argon, GPT-6 Astra and Claude Fable 5.1. Tied with Sonnet 5.5, two points below Opus 5.5. Fifth most expensive per task | Independent aggregate |
| Creator’s drawing test | About $0.34, 17,000 tokens, just under 4 minutes. Ranked 14th | Hands-on test |
| One-shot game test | All the game mechanics worked, but the graphics trailed Argon, Opus 5.5 and GPT-6 Astra | Hands-on test |
The takeaway: Muse Wonder looks competitive on paper and capable at getting code to work, but it is priced like a premium model. Test it on your own tasks before paying Opus-level prices for it. Our guide on how to verify AI model claims explains how to separate vendor charts from real results.
Which open-weight language models arrived?
Four language models in this week’s video were either open-weight or presented as open. Availability varies a lot, so check the last column:
| Model | Key facts | Can you use it now? |
|---|---|---|
| Muse Spark Open (Meta) | Meta calls it the most capable US open-weight model. 816B parameters, 1M-token context. Optimised with Apple to run on an M5 Ultra Mac Studio, a setup the creator put at around $10,000. Leads most open-weight benchmarks Meta shared, but trails Kimi K3 on coding. In the creator’s drawing test via OpenRouter: 1 min 45 s, 9,280 tokens, no cost | Yes. Weights and hosted APIs. See our Muse Spark Open breakdown |
| MAI Base 2 (Microsoft) | Pre-training finished. 80B active parameters. Microsoft says it beats every open-weight model it tested | No. Not post-trained or released yet |
| Nemotron 3.5 Super VL (Nvidia) | 120B parameters. Scores 26 on Artificial Analysis, well behind DeepSeek, Qwen 3.8 and GLM 5.3, but much faster at output. Drawing test: about 2.5 minutes, 14,400 tokens, no cost | Yes |
| Ultramarine 1 Flash (Paper Instruments) | Built for knowledge work. Priced at $0.15 per million input tokens and $0.50 per million output, close to GPT-6 Luna Max | Through Paper Instruments’ API. Not on OpenRouter when the video was recorded |
🔎 Unconfirmed / rumour: Paper Instruments claims Ultramarine 1 Flash is “Fable-class” on knowledge work at 60 times cheaper than Claude Fable. The benchmarks were picked by the vendor and compare it with models that are no longer each lab’s best. Its own site says the weights are “coming soon”. Treat “open” and “Fable-class” as claims until independent tests and the weights appear.
⚠️ Warning: The video groups MAI Base 2 with open-weight models. We have not seen Microsoft confirm that the weights will be released. It is also a base model, so its final abilities may look quite different after post-training.
What did OpenAI ship in week two of its 28-day pledge?
On 4 October, OpenAI’s Codex team promised that for 28 days it would either ship one clear improvement each day or give users a full usage reset. Week two was quieter than week one:
| Day | What happened |
|---|---|
| Day 5 | Composer predictions (it suggests your next prompt), creating a Dot from your phone, and delegating Codex work to your Dot |
| Day 6 | No update, so a usage reset |
| Day 7 | Default context for GPT-6.1 Sol in Codex and Work raised from 272K to 500K tokens |
| Day 8 | Reset |
| Day 9 | Storage changes in the ChatGPT and Codex desktop apps, cutting stored data on disk by about 60% |
| Day 10 | Reset |
| Day 11 | Changes and more compute to cut “model at capacity” errors |
Some users argued that the day 9 and day 11 changes were bug fixes and basic reliability rather than new features. That seems fair. The larger context window is the most useful change if you run long coding or document tasks.
What are decision models, and which new ones launched?
Decision models are not chatbots. They read text or look at an image and pick an answer from options you give them, such as yes/no, multiple choice, a rating or a rubric score. Instead of a long reply, you get a probability for each option, which is handy for automation.
- Microsoft Decision 1: gives a calibrated probability for each answer option, and can grade AI responses and agent actions against a rubric. Microsoft claims slightly higher accuracy than Jev.
- Nimble V4 and Nimble V4 Max (Bespoke Labs): the same kinds of answers, with fewer errors in the vendor’s benchmarks. Bespoke Labs also says they beat Jev.
The creator’s view: they all do much the same job, with small differences in accuracy. If you want to see how this works in practice, our OpenAI Decisions API Tutorial: Automate Lead Scoring and Spam Filtering uses the same idea for lead scoring and spam filtering.
What else happened in AI this week?
- Grok Voice 3.0. A new voice model with a free sandbox. The creator built a small support agent that answered questions using only his Future Tools website. Its replies came almost instantly, though it talked over him at one point. We turned this into a step-by-step guide: How to Build an AI Voice Support Agent That Only Answers From Your Website.
- ElevenLabs instant soundtracks. Upload a video to ElevenLabs Creative and it generates a soundtrack with synchronised sound effects.
- Facebook feed controls. You can now type in your interests to shape what you see, and the ranking is said to adapt in real time. The creator hasn’t tested it.
- Google scam check. On Android, Circle to Search can check a text message and an AI overview flags any warning signs. This is also untested in the video.
- Kylon. A new tool where several AI agents and several people work together in one shared chat. The creator had a demo and was impressed.
- Xbox Labs and Epic Games. Microsoft has started an experimental Xbox group exploring AI game mash-ups and AI-made games. Epic’s Tim Sweeney says the company is reorganising to speed up development with AI, aiming to build more rather than cut staff. Many gamers are not happy about it.
- Apple. A new Apple TV 4K with Siri, and a countertop smart-home screen with Siri built in.
- GenSpark Design. An AI design tool shown in a sponsored segment of the video. We haven’t tested it.
What should small businesses do with this news?
| Story | Status | What we would do |
|---|---|---|
| Krea 3 image editing | Released | Try it on one old thumbnail or product image and compare it with your current tool |
| Flux 3 Image Klein / Qwen Image 2.1 Turbo | Released, open weights | Test free on Hugging Face. Check the licence before any commercial use |
| Gemini 4 Argon | Released | Test now while the launch price lasts. Budget for roughly double from December |
| Muse Wonder | Released | Compare it with your current model on your own tasks before paying premium rates |
| Ultramarine 1 Flash | Released (API), weights “coming soon” | Try it on low-risk knowledge work. Don’t rely on vendor benchmarks alone |
| MAI Base 2 | Pre-training only | Background only. No action needed |
| OpenAI 500K context | Released | Useful for long Codex or document jobs. Re-test long tasks |
| Decision models | Released | Consider them for yes/no or scoring steps in automations |
The usual rule applies. Act on what has shipped, test it on your own work, and keep rumours and vendor claims on a watch list. Our guide on when to switch AI models gives you a simple re-test rule so you don’t chase every launch.
Ready to put this week’s news to work? Start small: edit one thumbnail with AI using region-based edits, or build a voice support agent that only answers from your own website.
Frequently asked questions
What is Krea 3?
Krea 3 is an AI image model from Krea that focuses on editing. You select part of an image, describe the change, and it returns edited versions while keeping the rest of the image the same. It can also generate images from text.
How much does Krea 3 cost?
Krea lists $0.06 per image at 2K and $0.03 per image at 1K. In the Krea app it uses compute units like other models. Check Krea’s pricing page for your plan.
Is Krea 3 better than Nano Banana?
In Matt Wolfe’s hands-on test, Krea 3 made cleaner region-based edits than ChatGPT’s image tool or Nano Banana. That is one person’s opinion from a few tests, so try both on your own images.
Is Muse Wonder better than Claude Opus 5.5?
Not overall. On Artificial Analysis it sits two points below Opus 5.5 at a similar price ($4 / $20 per million tokens). In the creator’s tests it got game mechanics right but produced weaker graphics than Opus 5.5.
How much does Gemini 4 Argon cost?
$2 per million input tokens and $10 per million output tokens until the end of November 2026. After that, the price roughly doubles.
Is Ultramarine 1 Flash really open-weight?
Not yet. Paper Instruments says the weights are coming soon. For now you can use it through its API, and its “Fable-class” claim rests on benchmarks the company chose.
The bottom line
Krea 3 image editing was the most practical release of the week. Fast, precise region edits are useful for anyone who makes thumbnails, product shots or social graphics. On the language model side, Muse Wonder and Gemini 4 Argon are both strong, but Argon is the better value right now. Most of the open-weight releases are worth watching rather than switching to. As always, test on your own work before you change anything.
Related posts
- Muse Spark Open: Meta’s Open-Weight Model, Opus 6 Rumours and This Week’s AI News
- Gemini 4 Argon Review: Has Google Finally Caught Up?
- How to Verify AI Model Claims Before You Believe the Hype
[…] This guide turns that demo into a practical plan for building an AI voice support agent for your own business, whichever voice platform you use. For the full news story, read our Krea 3 and new AI models round-up. […]