July was the month the gap between open and closed models stopped being a comfortable assumption. A frontier-class model shipped with downloadable weights, the cost floor for capable inference dropped again, and Europe’s regulatory clock officially started ticking. For anyone deciding where AI fits on the roadmap, the planning math changed in ways worth taking seriously.
Llama 3.1 405B made open weights a frontier option
On July 23, Meta released Llama 3.1, including a 405-billion-parameter model that posted benchmark numbers competitive with GPT-4 and Claude 3.5 Sonnet. It is the first openly available model that credibly belongs in the frontier conversation, alongside an upgraded 70B and 8B with a 128K context window.
Why it matters for operators: “open vs. closed” is now a real architecture decision, not a hobbyist footnote. If you have data you can’t send to a third-party API, or you’re projecting high-volume inference where per-token API pricing stops making sense, self-hosting an open model is a legitimate path. The caveat is that 405B is expensive to run well — the smaller variants are where most teams will actually get value.
GPT-4o mini reset the cost floor
OpenAI shipped GPT-4o mini on July 18 at 15 cents per million input tokens and 60 cents per million output — roughly an order of magnitude cheaper than prior frontier models and over 60% below GPT-3.5 Turbo, which it replaced in ChatGPT.
Why it matters for builders: cheap, capable models change what’s worth automating. Workflows that were too marginal to justify at GPT-4 prices — classification, extraction, routing, first-draft generation at scale — flip to clearly profitable. The discipline here is matching model to task: route the easy 90% to a mini-class model and reserve the expensive frontier calls for the work that genuinely needs them.
Mistral Large 2 kept the pressure on
A day after Llama 3.1, Mistral released Mistral Large 2, a 123B model with strong code-generation performance and a license permitting research and non-commercial use, with commercial terms available.
Why it matters: the frontier is now crowded, and that’s good for you. Multiple vendors at comparable capability means pricing leverage, fewer single-vendor dependencies, and the freedom to design systems that swap models as the leaderboard shifts. Build your stack so the model is a component you can replace, not a foundation you’re welded to.
The EU AI Act officially became law
On July 12, the EU AI Act was published in the Official Journal, setting an August 1 entry into force and a phased compliance timeline — prohibitions in early 2025, general-purpose AI obligations in 2025, and the bulk of high-risk rules in 2026.
Why it matters for operators: if you touch EU users, the deadlines are now real and dated. You don’t need a compliance team tomorrow, but you do need to know which risk tier your use cases fall into and to start keeping the documentation trail the Act will eventually require. Retrofitting governance is far more painful than building it in.
DeepMind’s models earned a silver-medal math score
Around July 25, Google DeepMind reported that AlphaProof and AlphaGeometry 2 solved four of six International Mathematical Olympiad problems — performance equivalent to a silver medal, the strongest an AI has shown on this benchmark.
Why it matters: this is a signal, not a product. It shows that pairing language models with formal verification and search produces reliable reasoning on hard problems — the same pattern that will eventually make AI dependable in domains where being confidently wrong is unacceptable. Worth watching, not yet worth shipping. (OpenAI also previewed its SearchGPT prototype this month — another preview, not a production tool.)
The throughline for July: more capable models, cheaper inference, more vendor choice, and a real regulatory clock. The advantage goes to teams that turn that into deliberate decisions about what to build, what to buy, and what to skip. If you’d like a clear-eyed read on where AI actually belongs in your roadmap, let’s talk.