AI Roundup July 2024: Llama 3.1 405B Goes Open, Mistral Large 2 Answers Back, GPT-4o mini Tanks Prices

TL;DR

July was the month open weights stopped being the polite runner-up. Meta shipped Llama 3.1 405B, the first openly downloadable model that actually trades blows with GPT-4o and Claude 3.5 Sonnet, and Mistral fired back the very next day with Large 2. OpenAI nuked the floor out of small-model pricing with GPT-4o mini and teased SearchGPT, while the EU AI Act got published in the Official Journal and became real law. If you build local or self-host, this was your month.

Meta Ships Llama 3.1 405B: The First Open Model That Plays at the Frontier

On July 23, Meta released Llama 3.1 in three sizes (8B, 70B, and the headline 405B), and the big one is the story. This is the first openly available model that genuinely competes with the closed frontier on general knowledge, math, tool use, and multilingual tasks. It was trained on over 15 trillion tokens across more than 16,000 H100 GPUs, and all three sizes ship with a 128K context window, a massive jump from the 8K of the original Llama 3 in April.

The 8B and 70B refreshes might matter even more for builders than the 405B does. Most of us are never going to self-host a 405-billion-parameter model (you are looking at multiple H100s just to serve it), but a 128K-context 8B that runs on a single decent GPU is the kind of thing that ends up in everyone's homelab stack by August.

The other quietly huge move: Meta updated the license to explicitly allow using Llama outputs to train and improve other models. Synthetic data generation and distillation from the 405B into something tiny just became a first-class, license-blessed workflow. That is a bigger deal than the benchmark charts.

Mistral Large 2 Lands One Day Later

Mistral does not let Meta have the spotlight to itself. On July 24, one day after Llama 3.1, Mistral dropped Large 2: 123 billion parameters, a 128K context window, strong multilingual support, and 80-plus coding languages. It posts 84.0 percent on MMLU and a state-of-the-art 92.0 percent on HumanEval, putting it in genuine Claude 3.5 Sonnet territory on code.

The interesting angle is size. While everyone else jumps from 70B straight to hundreds of billions, Mistral parked Large 2 at 123B, which they argue is the sweet spot for single-node inference throughput. The catch for self-hosters: weights are available, but commercial use runs under the Mistral Research License, so it is not the permissive grant Llama gives you. Test it free on La Plateforme and Le Chat, but read the license before you ship a product on it.

GPT-4o mini Drops the Cost Floor Through the Basement

On July 18, OpenAI released GPT-4o mini and immediately retired GPT-3.5 Turbo from ChatGPT. The pitch is brutally simple: better reasoning and vision than other small models, more than 60 percent cheaper than GPT-3.5 Turbo, at 15 cents per million input tokens and 60 cents per million output.

For anyone running high-volume pipelines (classification, extraction, routing, cheap RAG glue) this reset the math overnight. A lot of workloads that were borderline-too-expensive on 3.5 Turbo became trivially affordable, and the quality bump means you can hand mini jobs you would not have trusted to the old cheap tier. It is the unglamorous release that probably touched the most production code this month.

OpenAI Teases SearchGPT, Picking a Fight With Google

On July 25, OpenAI announced SearchGPT, a prototype search experience that blends its models with live web results and inline citations, launching to a small waitlist of testers. The framing was clearly aimed at Google's AI Overviews and at Perplexity, which had spent the year proving there was appetite for an answer-first search box.

It was a prototype, not a product, so temper expectations. But it signaled where the chat-versus-search lines are getting redrawn, and it is worth watching if you build anything that depends on referral traffic or retrieval.

The EU AI Act Becomes Law

On July 12, the EU AI Act (Regulation (EU) 2024/1689) was published in the Official Journal, which started the clock toward entry into force on August 1. After years of negotiation, the world's first comprehensive AI law is now real, with a staggered rollout: prohibited-practice bans hit in February 2025, general-purpose AI model obligations in August 2025, and most high-risk requirements in 2026.

If you ship anything that touches EU users, this is no longer theoretical. The general-purpose model rules in particular mean the providers of the models you build on (and you, if you are the provider) now have documentation, transparency, and copyright obligations on a real timeline. Worth a read before the deadlines sneak up.

Key Takeaways

• Open weights caught up. Llama 3.1 405B is the first downloadable model that genuinely rivals the closed frontier, which changes the build-versus-buy calculation for anyone with the hardware to run it.
• Distillation got a green light. Meta's license change explicitly permits training on Llama outputs, making synthetic-data and distillation pipelines a legitimate, mainstream workflow.
• The cheap tier got serious. GPT-4o mini made high-quality small-model inference cheap enough to throw at workloads that were not viable a month ago.
• 123B is a deliberate sweet spot. Mistral Large 2 bets that single-node-friendly sizing beats raw parameter count for real deployment, but its research license limits commercial use.
• Regulation is now a hard deadline. The EU AI Act is law, and the general-purpose model obligations arriving in 2025 mean compliance planning starts now, not later.