AI Roundup June 2025: o3 Drops 80%, Meta Buys Into Scale AI, Hollywood Sues Midjourney

TL;DR

June 2025 was the month the frontier got cheaper and the lawyers got busy. OpenAI cut o3 by 80% and launched o3-pro, Google pushed Gemini 2.5 Pro and Flash to general availability, and Meta dropped $14.3B for nearly half of Scale AI to stand up a Superintelligence Labs. Meanwhile Disney and Universal filed the first big Hollywood copyright suit against an AI company, suing Midjourney days before it shipped its first video model.

OpenAI Cuts o3 by 80% and Ships o3-pro

On June 10, OpenAI did the thing it almost never does: it made its flagship reasoning model dramatically cheaper without changing the model. o3 dropped to $2 per million input tokens and $8 per million output, an 80% cut from the previous $10 / $40. Same weights, optimized inference stack, much smaller bill.

Alongside the cut, OpenAI introduced o3-pro, a higher-compute variant aimed at the hardest math, science, and coding problems, priced at $20 input and $80 output per million tokens. It posted strong numbers on GPQA Diamond and AIME, trading benchmark blows with Gemini 2.5 Pro and Claude Opus 4.

For builders, the o3 price drop is the bigger story. An 80% cut moves a top-tier reasoning model out of the demo-budget tier and into something you can actually put in a loop. If you shelved an agent idea because o3 was too expensive to call repeatedly, June was your cue to dust it off.

Gemini 2.5 Pro and Flash Go Generally Available

Google moved Gemini 2.5 Pro and 2.5 Flash out of preview and into stable, generally available status across the Gemini API, Google AI Studio, and Vertex AI. The company also expanded the family with Flash-Lite, a cheaper, faster tier for high-volume, latency-sensitive work.

GA matters more than it sounds. Preview models can change under you or get deprecated on short notice, which makes them hard to build a product on. Stable Gemini 2.5 means teams can commit to it for production without the rug-pull anxiety. Pair that with a tiered lineup (Pro for hard reasoning, Flash for the middle, Flash-Lite for cheap bulk) and Google has a credible answer to OpenAI's pricing moves the same week.

Meta Pays $14.3B for 49% of Scale AI

On June 12, Meta confirmed a $14.3B investment for a 49% stake in Scale AI, the data-labeling company that quietly powers a huge share of the industry's training and evaluation pipelines. Founder Alexandr Wang left the CEO role to join Meta and lead its new Superintelligence Labs, while staying on Scale's board.

This is a data-and-talent play, not a model acquisition. After a lukewarm reception to its open releases, Meta is buying its way back to the frontier conversation by locking up evaluation data and a high-profile leader. The awkward part: Scale is supposed to be a neutral vendor to everyone, and several major customers reportedly started reconsidering once Meta owned half of it. Watch the second-order effects on who labels data for whom in the back half of the year.

Disney and Universal Sue Midjourney

On June 11, Disney and Universal filed a joint copyright infringement suit against Midjourney in federal court, the first time major Hollywood studios have gone after an AI image company directly. The 110-page complaint accuses Midjourney of scraping their libraries to train a model that happily generates Darth Vader, Elsa, and other iconic characters on demand.

This is a different fight than the text-model copyright cases. The plaintiffs are pointing at outputs that are unmistakably their characters, which is a harder thing to wave away as fair-use abstraction. If the studios win or force a settlement with real guardrails, every image and video model that trained on the open web inherits the precedent. For anyone shipping a generative-media product, this case is now required reading.

Midjourney Ships Its First Video Model

Roughly a week after getting sued, Midjourney launched V1, its first video model. It animates generated or uploaded stills into 5-second clips, extendable up to about 20 seconds, available to its roughly 20 million users through the web app.

The timing was either bold or oblivious depending on your read, but the product itself is notable: Midjourney priced video at around eight times an image job while claiming it is far cheaper than prior options on the market. Image-to-video as a cheap, accessible feature (rather than a research demo) is exactly the kind of thing that pulls indie creators in. The legal cloud overhead just got a lot heavier on the same day.

Key Takeaways

• The price floor is falling. An 80% cut on o3 plus a GA Gemini 2.5 lineup means top-tier reasoning is now cheap enough to loop, which unlocks agent workloads that were previously uneconomical.
• Stable beats shiny for builders. Gemini 2.5 going GA is a bigger deal than any single benchmark, because you can finally commit a production app to it without preview-model rug pulls.
• Data and talent are the new moat. Meta's $14.3B Scale AI deal shows the frontier race is increasingly about owning evaluation data and people, not just shipping weights.
• Hollywood entered the copyright fight. The Disney and Universal suit targets recognizable character outputs, a harder case to dismiss than the text-training disputes, and it could reshape how every media model trains.
• Generative video is going mainstream. Midjourney putting cheap image-to-video in front of 20 million users moves the category from research demo to consumer feature, lawsuit notwithstanding.