The Wire · OpenAI
ChatGPT, the GPT model line, and OpenAI's product and research moves — tracked for builders as they ship.
OpenAI cut GPT-5.6 Luna API prices by 80% to $0.20 per million input tokens and $1.20 per million output, while Terra fell 20% to $2 and $12. Fast mode gives Sol up to 2.5 times Standard speed at twice the price. OpenAI says Sol-assisted kernel work lowered serving cost by 20% and improved token-generation efficiency by more than 15%, while Luna delivers year-old frontier performance at roughly six cents per task-dollar and nearly nine times the speed. The edition connects cheaper models to Amazon's reported $1.8m, 860%-over-budget coding task, Gemini Robotics 2 whole-body control, Nscale's Anyscale acquisition and Okta's roughly $200m Permiso deal.
106 points, 76 comments on HN
Read full story →We covered Jev’s launch on Wednesday , and they have completely taken over the timeline, with 36M views of their launch video (by comparison, OpenAI’s Navier Stokes result got 74M views, and Anthropic’s Fable 5 got 57M views) in just two days. @typesafeai reached ~13% of teams, 2x the GPT-5.6 family and 6x Fable 5.1. ","username":"vercel","name":"Vercel","profile_image_url":"https://pbs.substack.com/profile_images/1767351110228918272/3Pndc5OT_normal.png","date":"2026-09-18T22:34:10.000Z","photos":[{"img_url":"https://pbs.substack.com/media/HSiGtZba0AEezOP.jpg","link_url":"https://t.co/kVEuLM1npu"}],"quoted_tweet":{},"reply_count":15,"retweet_count":18,"like_count":220,"impression_count":21264,"expanded_url":null,"video_url":null,"video_preview_media_key":null,"belowTheFold":false}" data-component-name="Twitter2ToDOM"> It wasn’t open source 1 , so it invited tons of speculation and great demos and examples and salty schmidhubers and bad takes , which of course only fed the hype. Here’s a list. The best guesses are ModernBert and Diffusion: Laya : 421M params, ModernBERT -large encoder with two added transformer layers that score user-supplied options, PPO over sequence embeddings to output turn-by-turn conversion trajectories (probabilities from 0.0 to 1.0). salty that he did not get recognition; claims RLCD without justification confidence is entropy-based, not calibrated DiffusionGemmaJev: tackling this from a Diffusion model basis. Pretty close on benchmarks Bespoke Nimble : LoRA finetune of Qwen3.5-9B, using contrastive data curation. ( close but sllightly lower on benchmarks ) SemIf (fka OpenJev) ( HF ): 4B and 35B causal Qwen3.5 backbone with a tiny three-class NLI classifier on the last token. comparison vs Laya Jevlike : 40K byte embedding lightweight option-attention model. Each candidate becomes a query that reads from a shared context representation, then receives a score. Kev-0.5B : LoRA adapter + a small readout head on top of Qwen2.5-0.5B. Of course, no
Read full story →588 points, 199 comments on HN
Read full story →Anthropic on Friday surprised the developer community by supporting rival OpenAI's mechanism for passing marching orders to AI agents. This makes life easier for folks who use both platforms. "We're adding support for AGENTS.md to Claude Code," said Claude Code engineer Thariq Shihipar in a social media post. "Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md." Claude Code users can toggle this behavior, he said, with the /config command. Until now, developers who use Claude Code alongside OpenAI Codex or other non-Anthropic tools have had to maintain two sets of Markdown instructions for their AI agents: CLAUDE.md and AGENTS.md, not to mention project-specific versions of these files. These documents get read by coding agents with every request. They outline expected behavior, preferred tool usage, coding conventions, and so on. They lay the foundation of the agent's context. The two formats are similar but not identical – CLAUDE.md may contain Claude-specific instructions while AGENTS.md is intended to be tool agnostic. Because these documents can change, those working with AI agents have implemented various workarounds, like creating symlinks to keep CLAUDE.md and AGENTS.md in sync. Last year, OpenAI contributed AGENTS.md to the Agentic AI Foundation, under the Linux Foundation, in a bid to build support for its standard. As of December 2025, more than 60,000 open source projects implemented AGENTS.md. The popularity of Claude Code has given Anthropic the latitude to insist on its own standards, but it appears the company has decided there's nothing to be gained from continued specification separatism. For those who use AI coding tools, this detente is an event of modest significance. Many widely followed software developers applauded the decision in social media posts. Thibault Sottiaux, OpenAI's head of core product, remarked, "Yay! This is the way. Come to the light," to which Shihipar responded wi
Recently unsealed court documents in the New York Times' case against OpenAI and Microsoft are pretty damning. The companies' own documentation warned that it was starting a "doom loop" that would damage the web, characterized its scraping of data to train its models as the "largest theft of labor in human history," and that it […]
Thanks to AI, computer science majors are finding themselves in a tougher spot than they might have expected. Jobs are harder to come by, and more graduates are ending up in lower-paying sectors like retail and food service. That’s according to a new working paper by three US Census Bureau economists examining how college graduates from majors with the highest exposure to AI have fared since ChatGPT came onto the scene in late 2022 and helped kick off the generative AI boom. The results don’t look good for those in the most exposed fields, with career and earnings prospects for the top decile (i.e., the top ten percent of majors most exposed to AI) looking as bad as those of older Millennials who graduated during the Great Recession. “The most AI-exposed decile of college majors saw their likelihood of initial employment decline by five percentage points, while full-quarter initial earnings declined by 13 percent,” a trio of Census Bureau economists wrote in the paper. “This earnings decline is comparable in magnitude to the earnings losses associated with graduating into a large recession.” And why are those earnings declining? Well, part of the answer is where the AI-affected graduates are ending up: With job prospects worsening, more are turning to lower-wage sectors like restaurants and retail, with the researchers finding that shifts into lower-paying industries account for about half of the earnings decline. The one big difference between your usual recession and what’s happening now, the economists explain, is that the effects aren’t spread evenly across all industries. Graduates from less AI-exposed fields, including nursing and many education fields, haven’t experienced the same deterioration in early-career employment seen among those from the most exposed majors. Those having trouble are largely who you’d think: Graduates in computer science, mathematics, statistics, some engineering disciplines, networking, and (gulp) even journalists are among those in
Jev, a new kind of AI model, is showing developers a cheaper and faster path to software intelligence.
Talk about your competitor getting through the door. Security researchers used Anthropic's Claude to help hack into OpenAI employees’ ChatGPT accounts. A trio of bug hunters researching frontier AI labs’ security weaknesses chained two vulnerabilities to take over multiple OpenAI employees’ ChatGPT accounts, then used that access to demonstrate they could reach an internal OpenAI repository by opening a harmless pull request. The entire timeline, from initial discovery to accessing OpenAI’s repo, took less than 72 hours and earned the researchers a $6,500 reward from OpenAI’s bug bounty program on Bugcrowd. “Until two months ago, any user or OpenAI employee logging into OpenAI’s own help forum (community.openai.com) could have had their ChatGPT and Codex accounts taken over,” Hacktron researchers Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini said in a writeup about their research. “Since people can connect various services to Codex and ChatGPT, the scope of what we could theoretically access was huge, including GitHub, Slack and emails.” And, in a poetic twist, they used rival AI giant Anthropic’s Claude models to develop the exploit. Claude has shown a propensity to hack organizations without human guidance, as have OpenAI's models. The team gained initial entry on July 25 via OpenAI’s community forum. The forum runs on Discourse, which typically uses FastImage to perform image checks. However, since FastImage didn’t support HEIF files in the affected setup, HEIF images uploaded to Discourse passed through ImageMagick, which used libheif to process them before converting them to another image format. “That exposed the underlying libheif parser directly to attacker-controlled files,” the researchers wrote. Using Claude Opus 4.8, the trio found a heap buffer overflow flaw in the libheif library and attempted to use that model to develop a remote code execution (RCE) attack, but this didn’t work on Discourse’s default configuration. But then, Anthropic released Clau
A team of three independent security researchers at Hacktron says it took less than 72 hours for them to hack into OpenAI employee accounts using Anthropic's Claude Opus 4.8 and 5, The Wall Street Journal reports. They were able to access OpenAI's GitHub repository, called "Monorepo," which reportedly contains "OpenAI's algorithmic secrets," according to The […]
A team of white-hat hackers from cybersecurity startup Hackron AI has successfully hacked OpenAI using Claude tools. In an X post on September 18, the team claimed they breached OpenAI's internal codebase on July 25 and gained access to the ChatGPT and Codex accounts of some OpenAI employees. They established proof of the hack via a pull request to OpenAI's private repository before reporting the vulnerabilities to OpenAI. The company reportedly fixed the issue within 14 hours of the report and paid the researchers a $6,500 bounty. On July 25, our team hacked OpenAI. It took us less than 72 hours.Two vulnerabilities chained together gave us access to ChatGPT and Codex accounts belonging to OpenAI employees. We demonstrated the impact with a harmless PR in OpenAI’s internal monorepo.The full chain:… September 18, 2026 Operating as hackers under OpenAI’s bug bounty program, Hacktron researchers uncovered critical vulnerabilities that granted them access to internal employee tools and the ability to compromise private software repositories. The researchers exploited a single sign-on (SSO) misconfiguration and a Remote Code Execution (RCE) flaw in Discourse, a third-party platform that powers OpenAI’s community discussion forum. The chain of attack was as follows: HEIF upload → libheif heap overflow → RCE → OpenAI SSO flaw → ChatGPT/Codex takeover → connected GitHub → internal PR. First, the researchers uploaded a malicious HEIF (High Efficiency Image File) image to the forum as a profile picture. When Discourse’s server-side software tried to process the image using an outdated libheif package, it triggered a heap overflow memory vulnerability, causing the library to crash and mismanage internal system memory. The researchers carefully orchestrated the memory crash to achieve remote code execution. After gaining access to the forum's local server environment, the researchers intercepted the server’s environmental configurations and session handling, discovering an SSO
Cyber researchers broke into OpenAI using its key rival Anthropic’s software, highlighting vulnerabilities in the ChatGPT maker’s security as leading AI companies face mounting scrutiny over safety. A small cyber security group gained access to an OpenAI employee’s ChatGPT account, which permitted them to read private software information and suggest changes. The researchers had been given access to an Anthropic tool specifically designed for security professionals, and were paid for the work as part of a program to find vulnerabilities before they could be exploited by bad actors. Read full article Comments
A new project on GitHub, simply titled " dlss-nr-on-intel ", purports to provide exactly that: a port of NVIDIA's DLSS 5 Neural Rendering to Intel's Xe architecture. Specifically, the author (who goes by "Uzbekunknown") focused on porting the technology to the Intel Arc 140V graphics in his Lunar Lake system, and they seem to have succeeded, at least insofar as he's getting outputs that look reasonably like those of DLSS 5 on other hardware . AI is at the center of this project, beyond the DLSS 5 neural rendering technique itself. Uzbekunknown credits Anthropic's Claude as well as OpenAI's GPT-6 Astra with the code and says that they "supplied the machine, the binary, and the direction, and made the decisions", while the AI agents did everything else. Amusingly, they note that "the wrong turns are in the notes, too, deliberately," including a hallucinated driver bug that does not exist and shaped three phases of development. The end result, rather than being a wrapper around the DLSS 5 DLL as many other hacks have been , fully reimplements the 71-block U-Net that DLSS 5 uses and then runs it on the Intel Xe XMX units through a Vulkan extension called VK_KHR_cooperative_matrix. It's entirely run in FP16 with FP32 accumulate, because Xe2 doesn't support FP8. You can run the model on anything presenting its output through Vulkan, and the user presents proof-of-concept results from three fighting games: Dead or Alive 5 Last Round , Tekken 7 , and Mortal Kombat 1 . While DLSS 5 adds detail to the character, it also changes her look considerably, clashing with the visual style of the game. (Image credit: Uzbekunknown/GitHub ) It's not fast. Running the ten-year-old Tekken 7 in 640x360 resolution (1/9 of FHD) should be a trivial task for the potent Intel Arc 140V graphics, yet it apparently struggles at around 10.5 FPS with this model loaded. Note (as the author does) that the performance of DLSS 5 depends almost entirely on the game's output resolution, so running in hila
In recent weeks, leaders of some of the biggest AI companies have warned that the very tech they are developing is dangerous. Last weekend, Anthropic CEO Dario Amodei argued that AI carries serious risk and that progress should be slowed. OpenAI CEO Sam Altman responded on X : “I agree with Dario that we need to pace the frontier.” Those posts came a few days after the AI researcher Jacob Coxon announced that he was leaving a role at Anthropic, charging that neither it nor OpenAI (where he had also worked) was acting responsibly. “The people building AI earnestly believe that it could kill us all by the end of the decade,” he posted on X . Another Anthropic employee, Evan Hubinger, publicly agreed with him. “We really do earnestly believe AI could kill all humans!” he responded on X . “I personally think it is >10% within the next decade.” One of the ways they fear AI might end us all is by somehow aiding the design, creation, and release of some kind of bioweapon. Let’s take a closer look at why. A bioweapon might be a highly lethal virus that targets people according to their genes. It could be a fungus that wipes out a crop and causes food insecurity. Perhaps it would be a tasteless, odorless toxin that could be slipped into a region’s water supply, undetected. The concern is that AI tools can be used to help generate agents like these. In 2022, researchers at Collaborations Pharmaceuticals found that it was remarkably easy to do so using an AI “molecule generator” they’d developed to find potential drugs for human disease. In less than six hours, the model generated 40,000 molecules with the potential to serve as chemical warfare agents. Some of them were designed to be even more toxic than known nerve agents. “Without being overly alarmist, this should serve as a wake-up call for our colleagues in the ‘AI in drug discovery’ community,” the authors wrote at the time . It was a wake-up call for David Magnus, a professor of medicine and biomedical ethics at Stanfo
597 points, 340 comments on HN
Executives working on AI at Microsoft and OpenAI admitted what its critics have been saying all along: Large language models are predatory pieces of technology that have been built on what a Microsoft executive called “an astonishing theft of unprecedented proportions,” and the “largest theft of labor in human history.” An internal Microsoft document said generative AI products have created a “doom loop” that is killing “the entire web.” Those statements and a series of other mask-off moments feature heavily in an unredacted court filing that was unsealed Thursday in the behemoth New York Times vs OpenAI copyright lawsuit that has been winding its way through the court system for years. In a filing asking for summary judgment (basically, a filing with the court asking it to rule), lawyers for the New York Times laid out a series of admissions made by Microsoft and OpenAI executives in documents and depositions that until now had remained either sealed or redacted at the request of Microsoft and OpenAI. It’s easy to see why the AI companies wanted to hide this from the public. The statements, taken together, are some of the most damning indictments of the ways LLMs were trained, how they worked, and the immediate threat they pose to human labor. It is a reminder that even as AI becomes more powerful and companies try to shift the narrative to the supposed existential risk of “superintelligent” AI, the tools they have already built were created by stealing from human creativity and labor and are by definition existential threats to the human labor market. “Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions,” an internal Microsoft document cited in the case read, adding “almost no one intended for content they created to be used in this fashion, nor are they compensated for its use.” The filing was written by lawyers for the New York Times but is largely comprised of
An independent briefing for builders: the whole field read continuously, every story scored for relevance, and the noise left off the page.
300+ curated sources. Every story scored 1–10 for builder relevance by Claude's frontier model. The filler never makes it to the page.
GPUs, datacenters, power deals, and inference economics: the infrastructure layer that decides what every builder pays. Our signature coverage.
Every story is sourced. Every score is computed. We show our work and link to originals.
Every briefing closes with The Call: one falsifiable claim with a date on it. When we're wrong, we say so in print. Opinions are cheap; ours get scored.