News, p. 6
News
440 stories · page 6 of 49GPT-6 Astra now has an Ultrafast mode: up to 8x faster generation, per NVIDIA
The mode runs on NVIDIA Blackwell and is available in the OpenAI API and to eligible ChatGPT Work and Codex users. OpenAI has no announcement of its own in its newsroom; the number is in NVIDIA's blog, with quotes from two OpenAI people, and in OpenAI's API documentation. OpenAI's documentation adds the condition that, for us, comes first: US or global processing only, no EU endpoints.
Read →Study at Apple: a strong model with a minimal harness keeps up with elaborate agent scaffolding on an equal time budget
A paper published on Apple Machine Learning Research and on arXiv compares the best open-source harnesses for autonomous ML engineering with a single coding agent that has only read, write and bash. Under an equal time budget and the same model, the scaffolding gives no advantage. The authors conclude that hand-crafted harnesses around strong models yield poor returns.
Read →Barclays expects half of its developers to be using Claude Code by the end of the year
On 1 October Anthropic said British bank Barclays is expanding their collaboration and rolling Claude out across the bank. Barclays expects Claude Code to reach 50 percent of its developers by the end of 2026 and a majority of its software engineers in 2027.
Read →Google releases Gemini 4 Argon to cyber defenders first, and without cyber brakes
On 30 September Google announced Gemini 4 Argon, its new top model. It is going to trusted defenders through the Fairwind programme, and for them and Google's internal teams it will come without cyber safeguards. The broad release is a plan: paid API customers and Google AI Ultra subscribers first.
Read →A 9.8 hole in Cisco SD-WAN Manager gives admin access without a password, and it is already in use
On 30 September Cisco published an advisory for CVE-2026-76504 in Catalyst SD-WAN Manager. Improper handling of URI encoding lets a remote attacker bypass the check and reach the API as admin. There is no workaround, exploitation is active, and CISA set a deadline of 3 October.
Read →OpenAI: people linked to Moonshot AI tried to extract its models' hidden reasoning
On 30 September OpenAI said it had disrupted a coordinated adversarial distillation campaign that began on 1 July. The company attributes the core of the activity to individuals associated with Moonshot AI, the developer of Kimi. No encryption was broken and there was no direct access to stored user conversations.
Read →DeepMind put a watermark inside the protein itself, without taking away its function in lab tests
On 30 September Google DeepMind introduced SynthID Bio, watermarking methods for synthetic biology. The signature goes into the choice of amino acids and into the 3D coordinates of predicted structures, and it can be verified even on the synthesised protein. In lab tests against three targets, watermarked proteins worked as well as unmarked ones.
Read →Anthropic: robots can already do three quarters of physical work, but beat human cost on only 0.3 percent
On 30 September Anthropic published a robot exposure index, with task ratings made using Claude. Robots today can do 74 percent of physical tasks in the US, mostly in controlled settings. They are cheaper than humans for 0.3 percent of tasks, and at the past pace of price declines it will take 40 years to reach 10 percent.
Read →Cloudflare launches a beta that pays you for each AI use of your text, not each crawl
On 30 September Cloudflare put Pay Per Use into beta: an AI company offers a price for a specific use of content, the publisher accepts or declines, the buyer reports each use, and Cloudflare bills and pays out. The same day Monetization Gateway entered closed beta for charging agents through HTTP 402.
Read →