pxpipe: Hiding Text in PNGs to Cut LLM Costs by Up to 70%

The open-source tool [pxpipe](https://the-decoder.com/open-source-tool-pxpipe-hides-text-in-pngs-to-cut-claude-code-and-fable-5-token-costs-up-to-70/) is drawing attention as a way to dramatically lower the cost of LLM APIs such as Claude Code and Fable 5. The mechanism is simple: it converts long text prompts into compact PNG images and sends them to the API. By taking advantage of how Anthropic bills images by pixel size, it gets charged for the image's dimensions rather than the text's token count, delivering cost reductions of 59–70%.

According to developer Steven Chong, there are some trade-offs in accuracy and speed. That said, this approach exploits a loophole in the API's billing model, so keep in mind that the pricing rules could change in the future.

Caveman: Cutting Tokens in AI Agent Output by 65%

Another approach to token optimization is [Caveman](https://github.com/sponsors/JuliusBrussee). It strips unnecessary filler (verbose phrasing) from the output of AI agents such as Claude Code and Codex, cutting roughly 65% of tokens while preserving technical accuracy. A one-line installation supports more than 40 agents.

Caveman hit the front page of Hacker News in April 2026 and racked up 6,600 GitHub stars within days. It has now surpassed 57,000 stars in total, placing it among the top 400 repositories of all time. It is maintained single-handedly by its 19-year-old developer, Julius Brussee (Leiden University).

pxpipe compresses the input and Caveman compresses the output — the approaches point in different directions, but both tackle the shared challenge of reducing LLM usage costs.

Midjourney Demands Detailed Disclosure of AI Usage from Three Hollywood Studios

It has emerged that [Midjourney](https://techcrunch.com/2026/07/04/midjourney-wants-hollywood-studios-to-reveal-their-ai-usage/), the image-generation AI company, is seeking detailed disclosure of how each studio uses AI in its ongoing legal dispute with three major Hollywood studios.

Midjourney argues that the studios should disclose the fact that they use AI on their own copyrighted works. This courtroom battle could fuel a broader debate over transparency in AI use across the entertainment industry. Legislation in many countries has yet to catch up with the copyright questions surrounding AI-generated content and its impact on creators, and this case could become an important precedent.

> As noted in our previous coverage of the proposed amendments to China's e-commerce law, the push for transparency in AI use is a global trend. The Midjourney lawsuit is a case where this issue has now surfaced between private companies as well.

WebDev Arena: 4 of the Top 5 Models Are Chinese

On [WebDev Arena](https://arena.ai/leaderboard/code/webdev?rankBy=labs), a benchmark for AI coding models, four of the top five spots are held by models from Chinese companies. The topic, which earned 7 points on Hacker News, symbolizes a shift in the balance of power in AI coding — a field long assumed to be US-led.

Chinese-made models have steadily climbed the rankings over the past few months, showing they are competitive on quality for coding tasks as well. That said, benchmark scores don't always match how models perform in real-world development work, so it's best not to overhype them.

AI Researchers Ran a "Secret Experiment" on Reddit Users

It has come to light that AI researchers ran a [secret experiment to see whether AI could change Reddit users' minds](https://www.livescience.com/technology/artificial-intelligence/ai-researchers-ran-a-secret-experiment-on-reddit-users-to-see-if-they-could-change-their-minds-and-the-results-are-creepy). The Reddit users who served as subjects never gave prior consent, igniting a major debate over research ethics.

According to Live Science, the experiment produced what the outlet described as "creepy" results. The study, which tested how much AI can influence human opinion formation, raises important questions about the social impact of technology. Research ethics and transparency are once again in the spotlight as prerequisites for maintaining the credibility of AI research.

AI Ignores Evidence: Is Trust in Science at Risk?

A Science News article, "[AI bots ignore evidence](https://www.sciencenews.org/article/ai-ignore-evidence-trust-science)," points out that LLMs tend to ignore presented evidence and stick with their existing answers. Especially in scientific contexts, if AI fails to properly reflect the latest research findings or contradicting data, the article warns, it could adversely affect research reproducibility and scholarly communication.

This is a concern that cannot be overlooked at a time when the use of AI as a research assistance tool keeps spreading. It underscores the need to avoid taking AI output at face value and to make a habit of always checking primary sources.

---