Frontier, p. 6
Frontier
76 stories · page 6 of 9Shieldstral: Mistral releases an open guard model that reads your rules at the moment of the check
Mistral released Shieldstral - a European, multimodal safety classifier with 3 billion parameters and open weights under Apache 2.0. The moderation policy goes in as plain text at the call itself, with no retraining. According to Mistral, the model holds its own against systems up to 7 times larger, and does it on a single card with 16GB of memory.
Read →LFM2.5: 2.6 billion parameters trained for agent work right on the device
Liquid AI released LFM2.5-2.6B - a model built from the ground up for agents on phones and laptops: it follows instructions, handles tools, delivers 220 tokens per second on an Apple M5 Max CPU, and the weights are free on Hugging Face. According to the company, it leads its class on most agentic benchmarks.
Read →Alibaba showed its biggest model - and promised its weights for next week
Qwen3.8-Max carries 2.4 trillion parameters and a context of one million tokens, already available through the API. The real news is different: Alibaba says that in days it will release the weights freely - a top-class model anyone will be able to download.
Read →OpenAI's next model is called Astra, and it debuts with ten math results
OpenAI published ten results on longstanding open problems in mathematics - achieved, by the company's own account, by an internal version of Astra, their next big model. Every proof is formalized in Lean, so a machine can check it. The tokens for all ten cost about $2000.
Read →Grok Imagine learns consistency: one face and one voice in every scene
xAI added references to the Imagine Video 1.5 model: you send a photo of a character and a voice recording, and the model keeps them the same from scene to scene. Text-only video and native 1080p resolution are also coming. References launch first in the US, for the paid tiers.
Read →OpenAI cuts the price of its fast model by 80 percent
On 30 July OpenAI announced lower prices for GPT-5.6 Luna and Terra, and a new fast mode in the API. Their cheapest model gets five times cheaper. This isn't news about a model. This is news about the bill at the end of the month.
Read →AI is reshuffling who does what: almost half of profession-specific tasks come from another profession
On 27 July OpenAI released the first report in its new Work at the Frontier series: an analysis of over 800,000 work messages in ChatGPT shows that 43.5% of profession-specific tasks come from another profession. The designer handles his own budget. The salesperson analyzes data. The marketer fixes his own website. The division of labor is being rearranged before job descriptions have caught up.
Read →The biggest open model is now downloadable: Moonshot releases the weights for Kimi K3
On 27 July Moonshot AI uploaded the full weights for Kimi K3 - 2.8 trillion parameters, the largest openly available model to date. We wrote about its debut ten days ago. Now anyone can download it and run it on their own machine. In theory. In practice the weights weigh in at around 1.5 terabytes and need hardware that few people have.
Read →Anthropic ships Opus 5 - near-flagship intelligence at half the price, with an effort dial
On 24 July, Anthropic shipped Claude Opus 5 - by the company's own numbers, nearly as capable as the flagship Fable 5, at half the price. The real news is the effort dial: you decide how hard the model thinks on each task, and how much you pay for it.
Read →