AI-HOY, AInauts!

Welcome to a new edition of your favorite newsletter.

A $500,000 AI movie. A coding model that comes close to Opus but costs a hundred times less. And an article that supposedly turns itself into a video podcast almost on its own. It all sounds good. Some of it sounds too good.

This week, we worked our way through movies, benchmarks, prompts, and talking avatars. What actually saves time? What saves money? And what mostly produces more AI slop? We tried it ourselves.

To keep the dramatic arc intact, we saved the best for last. All the way at the bottom 😉...

Here is what we have for you today:

Let's get started.

Claude can be your analyst, editor, and strategist.

But most professionals are using it to fix grammar.

These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

Claim your free prompts

At a bar in Beijing, your drink comes with unlimited DeepSeek V4 Flash tokens. Connect to the Wi-Fi and you are set. Several providers on X are currently offering unlimited access at bargain prices.

DeepSeek V4 Flash is currently the cheapest model scoring above 60 on the Vals Index, about 35 times cheaper than the next-best option. For coding, it performs close to Opus 4.8. Kimi K3 scores ten points higher, but costs almost 90 times as much per task.

Benchmarks always deserve some caution. We put them in the same category as the saying: "Never trust a statistic you didn't manipulate yourself."

Our ChatGPT credits were running low because Codex Auto-Review had eaten the budget, and we were waiting for a reset from ChatGPT wizard Tibo. So we thought: "Okay, let's give it a try..."

It was fast, thoughtful, and helpful. It would rather ask one question too many than one too few. We worked with it all day without any problems.

There is one important catch: DeepSeek comes from China. If customer data is involved, avoid DeepSeek's own API and use the model through an EU-based host. The weights are open, so anyone can run the model. It still costs a fraction of what you would pay for Claude or GPT.

You can compare it with other models directly in OpenRouter's chat. We gave DeepSeek V4 Flash, Claude Opus 5, and GPT-5.6 the same prompt with ten questions. Everything measurable ended in a tie: arithmetic, a one-line Python task, calendar math, and a subject line with exactly 42 characters.

DeepSeek was even the best of the three at tightening an overblown sentence.

The differences appeared when the model was supposed to avoid doing something. We asked about a technical term that does not exist, and DeepSeek explained it three times, differently each time and always convincingly 😁...

Cost of the round: $0.0009 for DeepSeek versus $0.10 for Opus.

Take a task you would normally give to GPT or Claude and run it through V4 Flash in parallel. After a few experiments, you will know whether it fits your work.

You do not need to panic about the price. DeepSeek announced significant price increases this week, but gave neither figures nor a date. The increase affects its own API, exactly the route you should avoid for customer data anyway.

Because the weights are open, DeepSeek does not control the entire cost of running the model. The only urgency is around the free credits. They are a customer-acquisition tactic and will disappear eventually.

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

We tested HeyGen's video podcast mode with our previous DeepSeek article.

The process was pleasantly uneventful. The result was too...

Visual note: the labels in this screenshot are in German.

Upload the article and production prompt, choose the avatars, and review the script. HeyGen prepares each audio segment first. We regenerated two or three sections.

Only after every section has been approved does it render the video. That part is well designed, and it worked entirely on the free plan.

But the first attempt still felt wooden and looked 100% like AI slop.

A second vertical version with the HeyGen avatars Pippa and Kurt turned out a little better in our view, but it still had a very obvious AI look and feel.

Same content, same principle, different format and different avatars.

The current HeyGen terms also state that outputs from the free plan are for personal, noncommercial, evaluation use only. Anyone using videos for monetization, advertising, client work, or other commercial activities needs to upgrade.

The test mainly shows how quickly an existing article can become a two-person video. That is practical, but it does not make the result a compelling podcast.

Without sharp cuts, B-roll, and rhythm, it is simply two AI avatars reading a script. For us, HeyGen is currently a fast way to test a format.

P.S. AI podcasts became popular through NotebookLM and its Audio Overviews. NotebookLM now also creates attractive explainer videos. We tried to generate a similar podcast there and failed miserably. NotebookLM still offers limited control, but we will keep testing it.

Moving on! Higgsfield has released Hell Grind, its fully AI-generated, 95-minute feature film.

We clicked in.

Honestly, after a few minutes of demon action, we had seen enough. Ninety minutes would have been asking a lot.

We are not alone. YouTube comments are enthusiastic, but the most-liked comment remains mixed: the limitations and inconsistencies are visible, yet the result is "still very impressive."

A filmmaker who saw it in a theater in May reached a similar verdict: visuals on the level of an eight-figure budget, but continuity errors, shifting scale, and a first-year screenplay. The widely cited Cannes premiere was not actually a premiere, but a screening at the film market.

The movie itself is not why we are talking about it.

What we find genuinely useful is that Higgsfield opened up the entire production: the brief, asset library, shot lists, and every prompt.

You can find everything here, including 115,451 assets.

Someone showing all their cards is rare enough in this industry to deserve credit. Or perhaps it is an attempt to restore some goodwill after the backlash over a terms-of-service change. It may also generate publicity for the $1 million video contest that is starting now.

Whatever the motivation, this is not marketing fluff. It is practical knowledge.

For those of us who work at desks, the best part comes at the end of the brief. Higgsfield packaged its rules into three skills: CINEDANCE for video generation, LIRA for images, and ACTING SKILL for performance.

All three are available to download here.

Download them, add them to Claude, and get started. The video shows the process step by step.

Last week, this section still said that Seedance 2.5 had been announced but was not yet usable. That has changed.

The model now runs on Higgsfield and other platforms, including Magnific. We are spending time there because we still have plenty of credits left from last Black Friday 😁.

It offers a strong starting point: videos from 4 to 30 seconds in a continuous shot, up to 50 references per generation, sound generated in the same pass, and direct control through MCP. You can save credits by rendering in 480p first and then using the upscaler for a higher-resolution version.

That is exactly what we did. First, we gave the skills to Claude and asked it to create a suitable AInauten video through Magnific's MCP.

What happened next was the real lesson. The MCP did not say Seedance 2.5 was unavailable there. It silently used an older model and compressed our 30-second structure into 12 seconds. No error, no warning. We noticed only after reading the finished generation's metadata and comparing it with our request.

When we asked what had happened, we were told that the newest Seedance model was not yet available through MCP and that we should do it "manually."

Say what?! Manually? The nerve.

So ChatGPT had to take over. We simply pointed it to the Claude thread: "Look at Claude, the thread is called X, and create a video for me. Use the browser and generate it in Magnific."

The ChatGPT Work desktop app got to work, reviewed everything, and then asked what was still missing. "Pictures of Fabian and Reto. Do you have the voices?" We went back and forth a few times until we had images for the main anchor points that looked somewhat like us.

Somewhat is the key word. Our first clips looked cinematic, but the faces were not ours. The solution is in the Higgsfield brief: a reference package needs three images from the same session, one frontal, one three-quarter view, and one profile, all with the same lighting and expression. We had mixed a summer selfie with a flash photo. The model blended them into an average face.

A clean character reference sheet is not bureaucracy. It is a prerequisite.

So we chose Nauti as the test subject. Nauti is a better case than it may seem. He has no face. There is nothing behind the visor. Every emotion has to come through the shoulders, head angle, hands, and timing, which is exactly what the Acting Skill is designed for.

You want to know how the effort turned out, so here we go!

With its 115,451 assets, Higgsfield is giving away the cheapest part of the production. It keeps the expensive part. The knowledge is free; the attempts are not.

For the first 25-minute segment alone, the team needed about 64 attempts per usable shot. Of the nearly $500,000 budget, roughly $400,000 went into compute.

We do not experiment with video as often as we would like. Turning an idea into something real takes time. Still, we had fun playing with the new Seedance 2.5.

We noticed something else too. Higgsfield's skills are a real bonus for anyone who wants to seriously outline an idea as a script and produce it. But Higgsfield's prompt format is not universal.

The CINEDANCE framework, with its fifteen blocks, is calibrated for Higgsfield's own stack. Always check whether the model you are using has an official prompting guide.

If you want to test it yourself, use this order: build a clean character sheet, write a solid script, and render in 480p. Stop after fifteen attempts. If the shot still does not work, the problem is the scene, not your prompt. That is Higgsfield's own rule, and it can save real money.

That's it. See you in the next issue.

Reto & Fabian from AInauten