Daily Tech News — September 25, 2026

Full transcript

Alex: Good morning, and welcome to Daily Tech News. It is Friday, September twenty-fifth. I am Alex.

Jordan: And I am Jordan. OpenAI just slashed prices in half, NVIDIA went open source on serious agent models, and Marc Benioff called Microsoft Copilot the new Clippy.

Jordan: On September twenty-second, OpenAI launched two new models, GPT six Sol and GPT six Luna, and cut API prices by fifty percent.

Alex: That is a big move. Sol now costs two dollars per million input tokens and ten dollars per million output tokens, half the price of its predecessor. And Luna is the real shocker. Ten cents per million input tokens, fifty cents per million output tokens.

Alex: And it is not just cheap, it is good. OpenAI says Sol makes about half as many factual errors as its predecessor, and it scored sixty-eight point eight percent on a coding benchmark called DeepSWE. To put that in perspective, that is within a point of Anthropic's flagship model Claude Fable five.

Jordan: So you are getting near-flagship coding performance at mid-tier prices. Luna is aimed at high-volume work where cost matters more than raw power. This is clearly a play for agent workloads, where an app might make hundreds of small model calls to finish one task. If every call costs money, the unit economics have to work.

Alex: Exactly. And OpenAI timed this for DevDay on September twenty-ninth, where Sam Altman is expected to lay out the company's pitch across every price point. Anthropic just cut Opus five point five prices too, which we covered yesterday. This is now officially a race to the bottom on price.

Jordan: One caveat from independent testing. Luna has some small regressions on certain coding evals despite the price cut. Cheaper is not automatically better everywhere, so test before you bet your production pipeline on it.

Alex: Good note. Now, speaking of agents, NVIDIA made a big open source play this week. They unveiled the Nemotron three Ultra model. It is a five hundred fifty billion parameter mixture-of-experts model designed for long-running AI agents.

Jordan: NVIDIA claims up to five times faster inference and up to thirty percent lower cost compared with other open models in its class. It is post-trained for popular agent frameworks, with a new secure agent runtime called OpenShell plus new models for safety and speech recognition.

Alex: Here is the part that matters for enterprise folks. CrowdStrike is already using Nemotron models for specialized agents that continuously identify, prioritize, and remediate vulnerabilities and misconfigurations. Palantir is integrating them into its AI FDE platform, including air-gapped systems.

Jordan: That is a fascinating signal. NVIDIA is not just selling GPUs anymore. They are building the whole agent software stack, open, and getting security and enterprise platform vendors to build on it. Your SOC tools could literally be powered by NVIDIA's open models soon.

Alex: Let us switch to the enterprise software world, where Salesforce just had its biggest product launch in years at Dreamforce. They announced AIforce, which is not another chatbot. It is an interface layer that connects Salesforce's data and workflows directly into whatever AI assistant you use.

Jordan: Right, so instead of logging into Salesforce and clicking through dashboards, you just talk to your assistant, and AIforce lets it pull records and run multi-step workflows on your behalf, using your existing permissions. They launched with three integrations. Claudeforce brings your Salesforce data into Anthropic's Claude. Slackforce connects it to Slack conversations. And there is an Agentforce Coworker inside Salesforce's own interface.

Alex: The killer feature for enterprise IT is zero data retention. Salesforce guarantees the model provider does not retain your business data, which has been the biggest objection to feeding CRM data into AI models.

Jordan: And of course, Marc Benioff could not resist taking a shot at Microsoft. He called Copilot the new Clippy, and said rival assistants suffer from a fantasyland problem because they lack a dependable business data layer. Strong words, but there is substance under the trash talk. Seven billion agentic workloads have now run across Agentforce and Slack.

Alex: That is a real number, not a slide. And Salesforce just set a sixty-three billion dollar revenue target for fiscal twenty thirty. If AIforce delivers on CRM data flowing into any assistant with zero retention, that is a serious enterprise play, whether or not you buy the Clippy comparison.

Jordan: Time to talk chips, because this week was wild on the silicon front. AMD just joined the trillion dollar club, riding a wave that includes a roughly ten percent chip price hike and a rotation back toward CPUs that plays to AMD's strengths. And reports say Amazon has ordered millions more GPUs.

Alex: But here is the bigger story. According to a report from Piper Sandler, pretty much every major US chip buyer, Apple, Nvidia, Google, Microsoft, Amazon, Tesla, AMD, and Qualcomm, is evaluating Intel's new fourteen A manufacturing process. The technical data is reportedly better than expected, and the industry is desperate for alternatives because TSMC is fully booked through twenty twenty-eight.

Jordan: Intel is putting real money behind this. They raised twenty billion dollars in a stock offering to build out fourteen A fabs, with one facility aimed at twenty thousand wafers per month. If they land even two or three of those customers for production in twenty twenty-eight, that changes the whole narrative.

Alex: And Nvidia's stock is trading around two hundred twenty-seven dollars after more than doubling this year. The constraint is still chips, and the orders keep coming. But keep an eye on the thesis making the rounds that power, not silicon, will become the binding constraint. Whoever supplies the electricity to these data centers might be the quiet long-term winner.

Jordan: Now, security, because it has been a brutal week. A critical WordPress flaw, CVE twenty twenty-six dash eighty-seven nine zero two, was patched on September twenty-second, and attackers were exploiting it within five hours. Patchstack saw traffic grow tenfold the next day as attackers moved from probing to actually writing files to servers.

Alex: The CVE is rated nine point two out of ten. If you run WordPress, confirm you are on version seven point one point two or later, and check your temp directories for files like wp-pear-rce-flag dot php or poc eighty-seven nine zero two dot php. Those are telltale signs of compromise.

Jordan: And that is not the only fire. An authentication bypass in TeamCity On-Premises, CVE twenty twenty-six dash sixty-three zero seven seven, is now flagged by CISA as being used in ransomware attacks. Unauthenticated attackers can run operating system commands. Upgrade to the patched versions or install JetBrains' security plugin, and keep your build server off the open internet.

Alex: There is also a reported zero-day that turns Meta's Muse assistant into Mac spyware, plus actively exploited flaws in Roundcube Webmail and Salesforce's own Agentforce. The lesson, as always, is that five hours from patch to exploit is the new normal. If your patch window is monthly, you are exposed.

Jordan: Quick hits before we wrap. Snorkel AI tripled its valuation to three point five billion dollars as demand for AI training data booms. Qualcomm launched two new smartphone chips with an emphasis on AI. And Meta admitted that Muse's likeness to the open source tool OpenClaw is not a coincidence, which is an awkward but honest moment for the company.

Jordan: And one reminder for developers. OpenAI says GPT five point five instant retires on October fourteenth, so expect changes to default models in ChatGPT soon. If you have pinned old model names in your code, go check your configs now.

Alex: That is the show for Friday, September twenty-fifth. I am Alex.

Jordan: And I am Jordan. Have a great weekend, and we will be back Monday with the latest.