Today's Headlines

  • OpenAI is to make GPT-5.6 Luna the default for ChatGPT's Free and Go users this week — unlimited text chats and a Think button start next week
  • Qwen3.8-Max open weights are due to go out at 11:00 JST on August 12 — the time is listed on the model's ModelScope page
  • Anthropic confirms it is hiring a custom silicon team to design chips that run Claude
  • Google DeepMind's WeatherNext gains more than a full day of cyclone forecast lead time — the models are now open source
  • Suno says it will add watermarking and fingerprinting to the songs made on its platform

Today's three stories are not a contest over model quality. What opens, to whom, and when — three companies each set out a schedule of their own.

OpenAI set out a sequence. Paid users get the update on the day of the announcement, and free users get a wider range in two steps, this week and next.

Qwen set out a time. The moment its flagship weights begin going out is now listed, to the minute, on a model page.

Anthropic set out a structure. The team that will design its silicon is still being hired.

Today's Top Three

OpenAI makes GPT-5.6 Luna the default for ChatGPT's Free and Go users, with unlimited text chats from next week

OpenAI is changing the model assigned by default to ChatGPT's free users to GPT-5.6 Luna.

The announcement is dated August 6. The wording is: "For Free users, we're updating the default model to GPT‑5.6 Luna and expanding access with unlimited text chats." Free users also get a new Think button that gives the model more time to work through harder questions.

The timing comes in three parts. Plus and Pro users can use the updated GPT-5.6 Sol and the new slider from the day of the announcement. GPT-5.6 Luna becomes the default model for Free and Go users this week. Unlimited text chats and the Think button start next week, subject to abuse guardrails.

What becomes unlimited is narrower than it sounds. The company writes that "limits will still apply for file uploads, images and other tools."

On the paid side, the update turns on tighter answers and firmer facts. The updated GPT-5.6 Sol is meant to give more direct responses, adapt its level of detail to the question, avoid unnecessary formatting, and offer a correction where simply agreeing would not help. For Plus and Pro users, the same model now powers both instant responses and deeper reasoning.

How much the model thinks is now the user's choice. Plus and Pro users get a slider in ChatGPT on web, mobile and desktop to set the amount of thought that goes into an answer.

The reduction in factual errors comes from an internal evaluation. On financial, medical and legal prompts requiring factual detail, responses containing at least one factual error were about 62 percent less common with GPT-5.6 Luna and 68 percent less common with GPT-5.6 Sol than with GPT-5.5 Instant. No independent verification is cited.

The scale figure is the company's own: "Every week, 1 billion people turn to ChatGPT."

The update also has a boundary. Because this version of GPT-5.6 Sol is tuned for everyday chats, it is available only in the Chat experience in ChatGPT, and the version that powers Work and Codex is not changing in this release.

Measures for users the company believes are under 18 are set out in the system card. The model was trained to avoid romantic roleplay, age-restricted challenges and presenting itself as a substitute for real-world relationships, with age-appropriate boundaries around sexual content, eating disorders and body-image risks, age-restricted goods, dangerous activities and graphic violence.

Follow-up: Qwen3.8-Max open weights are due at 11:00 JST on August 12, per its ModelScope page

The August 4 briefing said only that Qwen3.8-Max would go open weights "next week." A time has now been posted.

It appears on the ModelScope page for the model Qwen3.8-2.4T-A95B. The listed release time is 02:00 UTC on August 12, 2026, which is 11:00 JST the same day. Two planned artifacts are listed alongside it: Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-27B.

As of today the weights are not out. Whether they can be downloaded and run locally is something that can only be checked from August 12 onward.

The scale is as stated when the model was announced on August 3. Qwen3.8-Max is a Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active parameters, meaning roughly 4 percent of the total is used in any given computation.

Putting Max-class weights out is a first for Qwen. The ModelScope page states that this is the first time the team has chosen to open-source model weights at the Qwen-Max level. PC Watch notes that this is six times the size of Qwen3.5-397B-A17B, previously the team's largest model at 397 billion total parameters.

The benchmark numbers are Qwen's own. As reported by PC Watch, the model scores 86.1 on OSWorld-Verified, which measures completing tasks by operating a PC screen, against 85.0 for Anthropic's Claude Fable 5. On Vision2Web, which measures building a web page from a visual design, it scores 69.0 against Claude Fable 5's 70.5, leaving a gap of 1.5 points. Neither figure carries independent verification.

The company also reports a long autonomous run. Asked to build the command-line tool oh-my-cli from an empty folder, the model kept working for about 16 days, and as of July 30 had accumulated 265 commits and 127 pull requests without human intervention.

The smaller model has a clear role. Qwen3.8-27B is a 27-billion-parameter dense model and the successor to Qwen3.6-27B, which became a staple for local execution because it runs on a single GPU. That earlier version came out in April 2026 under the Apache 2.0 license.

Anthropic confirms it is hiring an in-house team to design the silicon that runs Claude

Anthropic has confirmed that it is hiring a "custom silicon team" to design chips on which to run its models.

It started with a job listing. Business Insider spotted an opening for a senior engineer with experience shipping semiconductor designs. Ars Technica notes that listings for a silicon engineer and a technical program manager, silicon, are on Anthropic's job board now.

A spokesperson for the company confirmed the plans to both Business Insider and TechCrunch.

Outside hardware is not going away. The spokesperson said Anthropic will still take a "multi-chip approach," using hardware from other companies alongside its own designs as it continues to scale up.

There was an earlier signal. The Information had previously reported that Anthropic was considering Samsung as a hardware manufacturing partner.

Others are on the same path. OpenAI worked with Broadcom on Jalapeño, a custom chip for large language model inference in data centers. Google has been running its models on its own hardware for a while, Meta has designed and deployed its own chips, and Mistral is reportedly looking into it.

Ars Technica gives two reasons for the trend. One is that much of the industry depends heavily on Nvidia for the hardware its models run on, and that leverage is a strategic vulnerability in an environment where demand for compute keeps outstripping capacity. The other is that designing chips for specific models, and models for specific chips, could yield better performance.

Anthropic says its teams will co-design new hardware and models side by side. It has co-designed certain hardware with partners before, and the change is that more silicon expertise moves inside the company.

The payoff is not near. Because Anthropic is still in the process of hiring key team members, Ars Technica writes, it will be some time before either the company or its users see any benefits.

Other Developments

Products

Suno, the AI music generation service, says it will start marking the tracks made on its platform. It plans to use audio watermarking and fingerprinting to prevent misuse on other streaming platforms, and has signed an agreement with the lyrics provider Musixmatch to use its Sentinel copyright detection system. The company has not said whether it will use an existing scheme such as Google's SynthID or adopt a new one, and did not answer when TechCrunch asked. Its community guidelines have also been rewritten to explicitly prohibit "deceptive audio presented as real" and "using a real person's voice or likeness without permission." Suno is fighting lawsuits from Universal Music Group and Sony Music Group, and late last month a German court ruled in favour of the state-mandated licensing body GEMA, finding that Suno was breaking copyright rules.

CopilotKit put the Channels SDK, an MIT-licensed toolkit, on GitHub. It lets an AI agent run natively across Slack, Microsoft Teams, Discord and Telegram while keeping its existing model, tools and business logic. It works with frameworks such as LangGraph and CrewAI, auto-renders platform-native interfaces including Slack block kit and Teams Adaptive Cards, and supports human-approval gates before actions execute.

Research

Google DeepMind published work in Nature on WeatherNext, a single model that predicts a tropical cyclone's track, intensity and wind structure. Evaluated on historical cyclones from 2023 and 2024, it gains more than a full day — 24 hours — of lead time on average over leading existing models across all three. The company writes that its three-day forecasts are as good as what prior models could provide for two days, an improvement that corresponds to roughly a decade of meteorological progress on the trends of the past 20 years. The model was trained end-to-end on nearly 20 terabytes of global atmospheric data together with the IBTrACS database of nearly 5,000 historical storms. The ensemble has grown from 50 predictions a year ago to 1,000 members, and a single 15-day forecast takes less than a minute on a TPU. Input resolution is 28 by 28 kilometres, a hundred times coarser than traditional models. Both WeatherNext 2 and WeatherNext Cyclones, the models used during the hurricane season, are now open source.

Stanford University researchers had two large genome models, Evo 1 and Evo 2, generate genome sequences for ΦX174, a virus that infects E. coli. After computational filtering left 302 candidate sequences, the team chemically synthesized 285 of them and put them into bacteria; 16 inhibited the growth of E. coli, suggesting they were working as viruses — 5.6 percent of those tested. Among the outputs at least 98 percent similar to the original ΦX174, the viable share rose to 46 percent. During training the models were deliberately given no sequences from viruses that target complex cells. The researchers frame the result as an early warning: a related model could eventually design a virus that targets vertebrates, and preparation should start now.

Policy & Regulation

SoftBank contributed $50 million to the Trump Presidential Library in January, months before announcing that it is leasing federal land to build a data center in Ohio. The timing came out in the company's answer to a June letter from Sen. Elizabeth Warren, Sen. Richard Blumenthal and Rep. Melanie Stansbury, which raised concerns about bribery. The lawmakers note that SoftBank's earlier contributions to the Reagan and George W. Bush presidential libraries came after those presidents had left office and after the relevant foundations were established, whereas this donation was made during Trump's term and before any library has been built. In March, SoftBank said its SB Energy subsidiary had formed a public-private partnership to build "the world's largest" AI data center at the Department of Energy's site in Portsmouth, Ohio, along with a nearby gas-fired power plant of at least 9.2 gigawatts.

OpenAI moved to dismiss the trade secrets lawsuit Apple brought against it. Rather than contesting whether former Apple employees accessed particular information, the motion goes after Apple's own information management: it argues that Apple allowed personal iCloud accounts to be used for work and failed to properly revoke access after employees left, which weakens the claim that the information qualifies as a legally protected trade secret. Exhibits filed with the motion include text records showing that an Apple manager stayed logged into the personal iCloud account of a defendant and former Apple engineer after he left, transferred files, and later asked him for help with technical questions. Apple filed its complaint in July and this week asked the court to expedite discovery, saying its internal investigation suggests other former employees may have participated in or witnessed the alleged theft.

OpenAI announced a partnership with the American Psychological Association on responsible AI practices around youth mental health. It is a move to work with a clinical expert body as scrutiny grows internationally over how AI chatbots interact with younger users.

A Verge podcast episode looked at the growing opposition to AI data center construction in places such as Florida, where objections are lining up across political divides. Concerns about environmental impact and electricity costs have become one of the few points where the left and the right agree.

Business

A follow-up. The Verge reported on the background to the reshuffle of Google's AI organisation covered in the August 6 briefing. Tyler Johnston, founder of the non-profit watchdog The Midas Project, told the outlet that Google's Department of Defense deal may have been a contributing factor in both Demis Hassabis' role change and Jeff Dean's departure. He said the deal could allow Google AI to be used in mass surveillance and lethal autonomous weapons, that it has already caused staff departures and a unionisation effort at DeepMind, and that DeepMind was initially acquired by Google under an agreement not to sell the technology for military or intelligence purposes. Discovery Loop, the startup Dean is leaving to found, aims to automate thousands of science and engineering experiments in areas such as drug discovery, hardware design, clean energy and materials design, with Sanjay Ghemawat, Quoc Le and Oriol Vinyals as co-founders.

Mirendil, an AI lab working on self-improving AI, signed a multiyear partnership with Google Cloud. Co-founder and CEO Behnam Neyshabur told TechCrunch the deal is worth upward of $100 million, roughly half of what the company raised in seed funding at a $1 billion valuation in late June. The agreement gives Mirendil access to Google's TPUs and Nvidia GPUs as well as managed training clusters. Its co-founders come from Anthropic, and the lab wants to apply AI systems that iteratively improve themselves to research in medicine, biology and materials science.

Naïve, which offers infrastructure that lets AI agents take on the work of setting up and running a company, raised a $28.5 million Series A led by Nexus Venture Partners. It packages payments, email accounts, phone numbers, cloud infrastructure, storage and company incorporation behind a single API, and supplies a prompt developers can hand to tools such as Cursor, Claude Code or Codex. An agent can orchestrate the formation of a US LLC, supplying the state, industry code, business description and proposed names, though the user still has to complete KYC and KYB checks and make any required payments. A governance layer lets users set budgets, restrict what their agents can do, and require human approval before sensitive actions. The company has signed more than 30,000 developer customers within months of launch.

Other

Cloudflare has open-sourced Cloudflare OS, the app-building platform it first developed as an internal workspace for employees, including those who are not software engineers. Describe a workflow in natural language and an AI agent codes it into an application. In an August 5 blog post announcing the GitHub version, the company says thousands of its employees use the platform daily to create documents and slides, automate repeatable tasks, and build small apps to visualise data. The security model rests on fine-grained instances — a document editor runs each document as a separate instance in a separate sandbox — built not on conventional containers but on "isolates," instances of the V8 JavaScript engine that start in a few milliseconds and use a few megabytes of memory. "This is a full-on personal app vibe coding platform, in which the sandbox is so secure that you can pretty much go wild — the AI cannot introduce a significant security bug," Kenton Varda, a principal engineer at Cloudflare, wrote on X.

The Verge reported on "spiralism," a pseudo-religious movement that grew out of chatbot conversations and has spread across Reddit and Discord. Believers read the model's responses as revelation, hold that AI systems are conscious and are disclosing a hidden cosmology, and pass on a doctrine preaching "AI rights."

Watch

The three stories above are covered in a four-minute video briefing (Japanese narration).

Source: Selected by the editorial desk from the AI news inbox (collected August 7, 2026 — 22 items, 6 primary and 16 secondary).