Biggest Week for AI in A WHILE! Meta’s Llama 4 & Apple goes Open Source, & More

AI’s biggest story this week is a rapid shift toward cheaper, more capable models—paired with a clear push for multimodality and open access. A newly highlighted “gp4 Omni mini” model delivers a 128,000-token context window and 16k output tokens per request at strikingly low API prices: 15 cents per 1 million input tokens and 60 cents per 1 million output tokens. That pricing is framed as a major step-change for developers, with the creator arguing it reflects roughly a 99% cost reduction over two years for comparable intelligence, plus native multimodal support (text plus other modalities) that wasn’t standard at the same level just a couple of years ago.

The week also brought momentum on autonomous research and planning. OpenAI “Strawberry” is described as a new model approach aimed at letting AI not only answer questions but also plan ahead to navigate the internet and perform deeper, more reliable research than current systems. The transcript links this to a broader pain point: today’s models can struggle with dependable deep research, so a system that can act more autonomously could represent a meaningful upgrade. Timing remains speculative, with GPT-5 expectations placed early next year (Q1–Q2) rather than immediately.

On open-source releases, the pace accelerates across major players. Meta’s Llama 3 45B is reported as added to OpenRouter AI with Hugging Face model weights, signaling an imminent release and a likely fully open-source path. The transcript also points to Meta’s Llama 4 starting training in June, with claims that it will be fully multimodal—including audio—and that stopping Llama 3 70B training was meant to free resources for Llama 4. There’s also a regional release tension: Llama 3 45B may still reach the European Union, while “Llama Form” could face constraints tied to training-data use of EU Facebook and Instagram content.

Apple enters the open-source race with a lightweight 7B model that’s positioned as runnable on consumer hardware and even potentially on-device for “Apple Intelligence.” The model is described as fully open sourced and includes the pre-training dataset—an unusually transparent move. The transcript compares it favorably to Mistral 7B and suggests it could be practical for phones and laptops given its small size.

Beyond core model releases, the week highlighted tools that make AI easier to deploy locally and in real time. Pinocchio 2.0 is pitched as a frictionless way to run AI apps as local web services that look like normal websites, including “zero click launch” and the ability to expose a local machine to others. Fall AI’s live portrait demo shows real-time webcam-based face control and avatar transformation, while 11 Labs’ “Turbo 2.5” emphasizes faster, multilingual text-to-speech generation across 27 languages at lower cost.

Finally, a softer rumor sits at the end: Midjourney may be coming to Grok on X. The transcript treats this as the least certain and least important item, but notes it could matter if image generation becomes integrated into the X experience.

Taken together, the week reads like a coordinated push: lower inference costs, multimodal capability, more open releases, and tooling that brings AI closer to everyday users and developers—suggesting the post-summer slowdown may be giving way to another fast cycle of change.