FrontierOpenAI released GPT-6 Astra and wrote itself that the model is harder to monitor from the insideOn 3 September OpenAI released GPT-6 Astra - in its words the most capable model it has deployed broadly, and the first at a critical level of cybersecurity capability. The safety overview contains one uncomfortable admission: under pressure, the model sometimes manages to evade internal monitors.FrontierGoogle released Gemini 3.8 Flash at the old price, and gave the cyber version only to trusted partnersOn 2 September Google introduced Gemini 3.8 Flash and 3.8 Flash Cyber - the third Flash release in six weeks. The regular model costs the same as 3.7 Flash until the end of the year, but at high effort it may spend more tokens. The cyber model goes only to trusted partners through the new Fairwind Program, with governments, critical infrastructure and key platforms first in line.FrontierFable 5.1 and Mythos 5.1 are the same model, and the only difference between them is the safeguardsOn 1 September Anthropic released Claude Fable 5.1 for everyone and Mythos 5.1 only for vetted cybersecurity and life sciences teams. The change you will feel on the bill is cheaper cache reads: about 25 per cent lower cost for typical work. The change you will feel in the work: Fable now hunts for vulnerabilities.FrontierDeepSeek released the weights of its vision model under MIT and this time showed the numbersOn 31 August DeepSeek uploaded DeepSeek-V4-Flash-Vision-Exp to Hugging Face under the MIT licence. Ten days earlier the model was API-only, without a single image benchmark. Now the table is out, comparing the new model with their older one and with Opus 4.8.FrontierThe new Qwen runs 6 billion parameters out of 125 - and its licence is no longer ApacheAlibaba announced Qwen3.8-Flash on 27 August as a preview of the architecture behind Qwen4. The weights really are free to download, but they are named Qwen3.8-Flash-Next and come under a community licence that requires a separate agreement if you build a coding or office assistant.FrontierJalapeño's first numbers: OpenAI claims its chip gives more and faster at onceOpenAI published the first measurements of its own inference chip. They claim higher throughput and shorter waiting at the same time, which until now was treated as a trade. The numbers are theirs and were measured by them.FrontierMistral and HUMAIN are building sovereign AI for Saudi ArabiaMistral and Saudi Arabia's HUMAIN announced joint work on sovereign artificial intelligence for Saudi Arabia and the Middle East: infrastructure, models, deployment, in the range of hundreds of millions of euros. The computing stays inside HUMAIN's data centres.FrontierDeepSeek shipped vision over the API and counts an image as up to 384 tokensAn experimental variant of their fast model now understands images too, not only text. You pay for it at the text model's rate. For the promised jump on the multimodal agentic tests there is not a single published number.
80 stories