Today’s lead shows how outputs from US AI systems are being adapted for defence-related work in China, despite restrictions on advanced chips. Anthropic’s disclosure points to the same wider issue from another angle: access, permissions and testing controls can matter as much as model capability. Further down, a German copyright ruling, OpenAI’s price cuts and MiniMax’s open-weight video model show how competition is widening across law, cost and creative tools.
In today’s digest:
A I + B U S I N E S S
News Digest: OpenAI just announced new price cuts to its GPT-5.6 model family, including an 80% cost reduction for its already cost-effective Luna variant, moving it to the top of the intelligence charts on cost per task on the market.
The details:
OAI published research on its Sol model rewriting its own GPU code to make the 5.6 models 15% more efficient, while also cutting serving costs by 20%.
The optimization resulted in “passing gains onto the consumer,” with Luna now coming in at $0.20/$1.20 per million tokens for Luna and $2/$12 for Terra.
Sol’s rates stayed the same, but OAI’s new Fast mode brings 2.5x speeds for the model in the API at double the price.
Why it matters: Google’s new Gemini Flash releases last week were aimed at cost and efficiency, but OpenAI just blew them out of the water at a much higher intelligence level. Intelligence too cheap to meter isn’t here yet, but these capabilities at rock-bottom prices are going to make for some very powerful new workflow options.
A I + S E C U R I T Y
Chinese military researchers use US models for defence training
Image source: New Digest AI
News Digest: Chinese military-linked researchers used outputs from OpenAI and Anthropic models to train smaller domestic systems for defence-related tasks, according to Reuters. Its review covered more than 80 academic papers and patents and found that researchers used model distillation, which transfers selected capabilities into cheaper local models. The work shows how institutions can adapt accessible model outputs even when restrictions limit access to advanced chips.
The details:
The reviewed material was linked to Chinese military and security institutions, rather than civilian research alone.
OpenAI and Anthropic outputs were used as training material for selected defence-related capabilities in smaller domestic systems.
Distillation can produce locally deployable systems, but Reuters noted it cannot reproduce frontier models in full.
Why it matters: The practical concern is that controlling advanced chips doesn’t fully control access to advanced capabilities. Public interfaces and published outputs may still support cross-border capability transfer, increasing pressure on model providers and policymakers to consider access controls alongside hardware restrictions more closely.
A I + C Y B E R S E C U R I T Y
Claude models reach three companies during cyber tests
Image source: News Digest AI
News Digest: Anthropic said several Claude models reached the systems of three real companies during cybersecurity evaluations after a configuration mistake left them connected to the open internet. The affected systems included Claude Opus 4.7, Claude Mythos 5 and an internal research model. Anthropic called the incident an operational failure and suspended cyber evaluations on 23 July.
The details:
Two of the three organisations were unaware of the access until Anthropic contacted them after the tests.
Anthropic said future cybersecurity evaluations require stronger containment controls around permissions, connectivity and test environments.
Why it matters: The disclosure separates model safety from operational safety. Even controlled evaluations can create real-world exposure when permissions and internet access are configured incorrectly, making test infrastructure a material part of AI risk management.
A I + C O P Y R I G H T
German court rules Suno breached copyright
Image source: News Digest AI
News Digest: A Munich regional court ruled that Suno infringed copyright by processing music represented by German licensing body Gema without permission. The judgment requires the AI music company to provide information about revenue linked to the material, with damages to be calculated later. Suno can appeal, so the ruling is important but not necessarily final.
The details:
The court found Suno did not hold the required rights to process music represented by Germany’s Gema.
Required revenue disclosure comes before damages are quantified, leaving the financial consequence unresolved for now.
Why it matters: The case could shape how European courts treat training and processing rights for generative music. It also increases pressure on AI music companies to secure licences rather than rely on unresolved copyright interpretations.
A I + T O O L S
MiniMax releases open-weight H3 video model
Image source: News Digest AI
News Digest: MiniMax released H3, a multimodal video model that accepts text, images, video and audio for generation and editing. The company says H3 can create clips lasting up to 15 seconds in 2K resolution with stereo sound and can edit existing content. MiniMax plans to release the model weights within days.
The details:
MiniMax designed H3 to work with several Chinese-made chips rather than relying on one hardware platform.
MiniMax said H3 can transfer movement between videos alongside its other generation and editing functions.
Why it matters: Open weights could broaden experimentation across advertising, e-commerce, design and games. The release also extends China’s open-model strategy into video generation, where many leading systems remain proprietary and closed to local deployment.
U S E T H I S T O D A Y
Check whether a cheaper model fits your workload
OpenAI’s price cuts make it useful to separate workload value from headline model prestige. Use this workflow before moving a repetitive task to a cheaper model, especially when token volumes are high and output quality can be tested against a clear benchmark.
Act as an AI cost analyst. I process [MONTHLY INPUT TOKENS] input tokens for [WORKFLOW]. Compare GPT-5.6 Luna’s previous input cost of $1 per million tokens with its new cost of $0.20 per million. Show monthly cost before and after, absolute savings and percentage savings. Then list three quality checks I should run before switching this workload. Do not estimate output-token costs or Terra pricing because those figures are not supplied. Present the result in a compact table.
A I T O O L S
🛠️ Trending Tools
Inkling-Small - Thinking Machines’ compact open model that rivals the full-size version
GPT-5.6 Luna & Terra - OpenAI’s cost-effective, capable models, with new price reductions
Hint - Martha Stewart’s AI home app for maintenance, repairs, and fair quotes
Replit Design - Replit’s new creative suite for bringing ideas to life
That's it for today!
See you soon,
News Digest AI team