c/news

The new ChatGPT Images 2.5 keeps people’s appearance from reference images more accurately, edits specific details with greater precision, and generates images up to 50% faster. It can replace one object or the background while preserving the rest of the composition. Sequential edits also stay more consistent.

ChatGPT now includes Sketch, where users can draw a rough room or clothing layout, add a description, and generate an image from it. Developers get GPT-Image-2.5 Flare for speed and bulk generation, plus Sunburst for detailed work. The update is available across ChatGPT, ChatGPT Work, and Codex.

ChatGPT Work can now remember your preferred wording and capitalization, including a characteristic sign-off.

Connect Gmail, Google Drive, Slack, or SharePoint so ChatGPT can analyze your work content and use its style in new texts. The setting is in Settings → Personalization → Writing style. It applies on web and mobile, and is available on all paid plans with ChatGPT Work access.

It starts with Minecraft and Subway Surfers, then moves to Fortnite and Instagram.

One day, the screen is full of ChatGPT, GitHub, VS Code, Notion, plus 17 more tools “for work.”

Lyria 3.5, Google’s new music model, improves vocals, creates richer arrangements, and generates higher-quality tracks.

Users can choose or describe a genre. They can select vocal or instrumental mode, then set the track length. Templates are available for background music and jingles, along with personal songs. Lyria 3.5 is now available in Gemini on web and mobile, plus Gemini API, Google AI Studio, Flow Music, and Google Vids.

Pro users got first access in Codex and ChatGPT.

An official prompting guide for the new model is also out.

An Apple I linked to Fred Hatfield is going to auction on October 14. Bonhams estimates it at $500,000 to $800,000. The lot includes Steve Jobs’s 1978 reply to Hatfield, who had complained that Apple computers had no software.

Jobs offered to exchange Hatfield’s Apple I without memory, plus a $400 check, for an Apple II board with 4 KB. Hatfield declined. He later sold the machine for $40,000. In 2013, it fetched $671,400.

The computer still works. It is Apple I No. 48, with Steve Wozniak’s signature on the board. Only about 200 Apple Is were built, and fewer than 75 survive. Its original price was $666.66.

OpenAI’s new model scored 98% on ARC-AGI-3, a difficult abstract-reasoning test. It reached 100% on ExploitBench, which measures vulnerability discovery and exploitation.

Astra also shows gains in coding and advanced math, plus science and CAD. OpenAI is testing it as an agent that can execute long action chains. Its cyber capabilities led to tighter safety limits before release.

Some results come from OpenAI’s own agent system, making direct comparisons with other models difficult. GPT-6 Astra will first be available to organizations in Daybreak Access, then to Plus and Pro, Business and Enterprise users in the next few days.

He gives absurd answers to routine interview questions, and the recruiter treats them as normal. Asked about salary, he mentions a sideways drill bit in the first quarter.

The AI turns that into a question about horizontal drill diameters, then thanks him for sharing his expectations. That is how to behave around neuro-HR.

Researchers at POSTECH built MiPAL, a parametric speaker with one emitter and 2 metamaterial layers. A front acoustic metasurface focuses ultrasound, while elastic metacells suppress vibrations that would spread audio sideways.

MiPAL modulates an ultrasonic wave with an audio signal, using nonlinear behavior in air to send sound along a narrow path. Tests in a mock aircraft cabin covered 500 Hz to 10 kHz, while the adjacent seat stayed quiet.

The modules are 3D-printed, so the design can be adapted to seat spacing, beam angle, or other conditions. Possible uses include personal announcements on planes and audio guides at museum displays.

The CameraJet’s tiny camera captures 28 photos per second. An AI system trained on 470,000 dental images finds gaps between teeth and directs a small rinse stream there during brushing.

It can also stream a live view of the mouth to the MyDyson app, showing cleaning coverage and close-up images of the teeth.

Dyson says development took 6 years and involved 661 engineers. The $499.99 device combines an electric toothbrush, oral irrigator, and camera.

During a three-hour test over Alaska, Rainmaker’s drones released small amounts of silver iodide into cold clouds. The company says this helped form ice crystals that fell as rain or snow, producing an estimated 71 million liters of additional precipitation.

Rainmaker uses weather algorithms, radar, and satellite data to select clouds and time each seeding flight. Atmo’s AI models help identify clouds with high seeding potential. The estimate is based on Rainmaker’s own analysis.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the same model with different restrictions. Fable is open to everyone. Mythos requires verification for cybersecurity and life sciences work. With open protein-design tools, almost 50% of designs worked across 12 targets, versus 10-15% normally. Binding affinity was 10 times higher than leading Adaptyv Bio contest entries for 3 targets.

The model also built a new elevation map for a third of Venus from Magellan radar images. Resolution improved to 2 to 3 km. Terminal-Bench-Science reached 52.6%, up from 24.7%. Cached reads are 75% cheaper, cutting typical task costs by about 25% and agentic costs by up to 45%.

The shift from traditional GPUs is tied to Apple’s unified memory architecture, which gives the CPU, GPU, and Neural Engine shared access to memory, reducing data transfers during agent training.

Reinforcement learning combines inference with interaction with an environment. Many compact Macs can therefore cost less than GPU clusters built for matrix-heavy workloads.

Anthropic reportedly uses Mac mini too, renting them from AWS. Demand has stretched delivery times for high-memory Mac mini and Mac Studio configurations to weeks or months.

The prototype was a shortened Apple Pencil made for the foldable iPhone. It was designed to attach magnetically to the side of the case, where it could also charge.

Testing exposed 2 problems. The stylus was awkward to use in some positions, and its hard tip left marks on the inner display. Foldable screens are softer than regular ones.

Mark Gurman believes the accessory will not launch with the foldable iPhone expected next month. The project may also clash with Apple’s long-standing position, linked to Steve Jobs’s 2007 claim that nobody needs a stylus. The Pencil will likely remain an internal prototype.

Videos can now be extended in 10-second chunks up to 40 seconds. The model uses the last 10 seconds of the source, rather than only its final second. The preview will soon be unavailable.

A 360p draft mode costs $0.03 per second, compared with $0.10 for 720p. 1080p costs $0.15, while 4K costs $0.30. Both are upscales, not native generations.

The official build uses a simple black case. Inside are a Pi 5 board and a Raspberry Pi flash drive. A LiPo battery sits beneath the Bluetooth keyboard.

Cyberdecks are portable computers assembled from assorted parts, including flea-market hardware and repurposed components. The guide includes a parts list and step-by-step assembly instructions.

The build is easy to reproduce, but its practical use remains unclear.

Portable Computer runs its models and agent system on the user’s computer. It can search local files and code, then run commands and handle long tasks without sending data to the cloud.

It uses Qwen 3.8 27B and PPLX 27B. Local tasks do not consume Perplexity credits. Internet access or a stronger cloud model requires permission first.

It supports Google Drive plus Gmail, Slack and GitHub. For now, access is limited to Pro and Max subscribers using NVIDIA DGX Spark on Linux. Support for regular NVIDIA RTX computers and Windows will come later.

Users describe a task in plain language, and the AI agent carries it out. It gradually remembers their working habits and applies them in later sessions.

QwenWork can create and launch web apps without code or server setup. It also generates media, with simpler or more advanced models available for different tasks.

The service is now in public beta worldwide. It runs in a browser and in a separate desktop app.

OpenAI told SpaceX it will terminate Cursor’s contract and plans to disable access to its models on 12 November 2026. Until then, Cursor users can continue using them. The models currently handle about 5% of Cursor traffic.

OpenAI says it cannot confirm compliance with its terms after Cursor came under SpaceX control, citing earlier breaches by Elon Musk’s companies. It also will not provide future models, including Astra.

Cursor’s CEO says the company is negotiating with OpenAI. Cursor had worked with OpenAI for nearly 4 years and relied on its platform as neutral infrastructure.

Authentication happens separately, so the agent does not see the login or password. It then works through the browser on the authorized site.

It can book a doctor, process a return, or fill out applications. It can also move invoices from email into accounting software.

The feature is starting to roll out on the web and mobile to users on the Plus tier, Pro, or Business.

In an interview, Thibault, OpenAI’s head of Codex and ChatGPT, said the current Codex may look primitive in 2-3 months. Next-generation Astra-based agents will need the cloud, where they can run more parallel processes than a typical laptop can handle.

Agent speed is also expected to rise. Today’s Ultra Fast rate of 750+ tokens per second could become the default. Codex and ChatGPT will increasingly share one harness across coding and other tasks.

The stated end goal is Personal AGI, a single personalized app whose interface changes with the user’s task.

The foldable iPhone will debut at 10:00 a.m.

Apple also plans iPhone 18, new Apple Watch and iPad mini models.

Alibaba released Qwen3.8-Flash-Next, an open multimodal model and early preview of the architecture planned for Qwen4. It has 125B parameters, but uses only 6B for each text fragment. Another 51B sit in separate memory for frequent combinations.

The design keeps the model large while reducing compute. Alibaba says training is about 9 times cheaper than Qwen3.7-Plus, while the model already performs better on coding, tool-use tasks, and long office workflows.

Long-context support reaches 262,000 tokens by default and up to 1 million tokens in extended mode. At 1 million tokens, a new retrieval mechanism can significantly speed up processing. The weights are already open on Hugging Face.

The update brings a new visual layer with transparency and a light frosted-glass effect. Surfaces now have depth, making it clearer what sits in front. The system font is Geist, while the dot matrix becomes mostly decorative.

A new colour engine can take shades directly from your wallpaper. They flow into icons, widgets, and controls, so changing the wallpaper changes the interface. Essential Apps Builder is now a phone app that turns a text description into a custom AI widget.

Comments under the announcement are less enthusiastic. Users say a year of work brought a font, more colour settings, and transparent widgets, while Nothing still lacks its own clock and calculator.

Preorders are open, with shipments starting September 22. Both desktops come with macOS 27 Golden Gate and the new Siri AI.

The Mac mini keeps its tiny body, but gets M6 or M5 Pro chips. M6 starts at $899 and uses Apple’s first 2nm process, with 12 CPU cores and 170 GB/s memory bandwidth. Both models cost $100 more.

Mac Studio now starts at $2,499 with M5 Max. M5 Ultra starts at $5,499. Its four-die design reaches 36 CPU cores and 80 GPU cores, with 1.2 TB/s memory bandwidth. With Mac Pro discontinued, it is Apple’s most powerful computer.