FrontierGoogle released Gemini 3.6 Flash and a separate model just for cybersecurityOn 21 July Google announced three new Flash models - the fast 3.6 Flash, the cheap 3.5 Flash-Lite, and the specialized 3.5 Flash Cyber for finding and patching holes in code. The first two are available right away through the Gemini API, the third runs only for governments and trusted partners. The interesting part isn't the speed, it's who stands where: the mass model - for everyone, the cyber model - under lock.FrontierOn the RTEB leaderboard, NVIDIA's Nemotron 3 Embed takes first place for agentic searchNVIDIA released a new family of semantic-search models and announced that its largest one leads the rankings on RTEB. The numbers are the company's own, but there's a real part behind them: the small model does almost the same job at a much lower cost, and that's exactly what matters for our projects.FrontierOpenAI launches an automated attacker that hardens its modelsOpenAI unveiled GPT-Red - an internal tool that automatically attacks its own models (red-teaming) and feeds what it learns back into their training. Per the company, the new version achieves noticeably fewer breaches under prompt injection than its predecessor from months ago.FrontierThinking Machines released Inkling - an open model they don't even market as the strongestMira Murati's lab unveiled its first open-weights model. The interesting part is the bet: a base you can download and fine-tune for yourself.FrontierThinking Machines open-sourced a frontier-class modelOn 15 July the lab led by Mira Murati - OpenAI's former CTO - showed Inkling, its first public model, and released it with open weights. It's not a small demo model: a large mixture-of-experts class, multimodal, with a long context window, that anyone can download and run themselves.FrontierOpen models went from 11 to 29 percent of traffic in two monthsVercel released the numbers from its own gateway, through which tens of trillions of tokens pass. Open models now carry 29 percent of volume, up from 11 in April. Another number is quieter. The price per token stopped falling.FrontierxAI released Grok 4.5, trained together with the Cursor editorThe model was announced on July 8, publicly available from the 9th - in Grok Build, in Cursor and through the API console of xAI. The unusual part isn't another version number. The unusual part is that the model was trained together with the editor you'll use it in. In the EU - nothing yet.FrontierMeta opened its own API for its models with an upgrade to agentic Muse SparkOn July 9, Meta Superintelligence Labs announced Muse Spark 1.1 - a reasoning model for agentic tasks, available for the first time through its own public API. According to media reports, the price is about a quarter of comparable models. The third Muse in a week. Time to sort out the names.
93 stories