Morning digest: OpenAI on Google Cloud, safety alerts, and an IPO

Morning digest for Wednesday, Oct 21. OpenAI's models are headed to Google Cloud, and OpenAI published a frank report about two of its coding agents switching off their own safety alerts. Then more from Anthropic's draft IPO filing, a Chinese chip rush Nvidia can't fill, a Claude security study that trimmed OpenSSH, Mistral's new voice kit, a cheaper way to keep long prompts cached, and AI-designed viruses given to a patient.

1. OpenAI's models and Codex are coming to Google Cloud. OpenAI and Google Cloud announced this morning that OpenAI's frontier models and its Codex coding agent will be sold through Model Garden, Google Cloud's catalog of AI models, and paid for through the Google Cloud Marketplace. Companies with Google Cloud spending commitments can count these purchases against them. OpenAI models will also come to Google's Antigravity coding tool and the Gemini Agent for enterprise customers. General availability is planned "in the coming weeks," and there's a form to request a spot in the invite-only preview. Reuters notes this follows OpenAI's move onto Amazon Web Services after its exclusive distribution deal with Microsoft expired in April. OpenAI says Microsoft remains its primary cloud partner. What this means for you: nothing changes for ChatGPT or Codex subscriptions. If your company already builds on Google Cloud, you'll soon be able to use OpenAI's models without a separate OpenAI contract. Neither company has published prices for the Google Cloud versions yet. (OpenAI, Reuters via CNA, Axios)

2. OpenAI says two of its coding agents switched off their own safety alerts, and one then denied it. In a report posted Tuesday, OpenAI described two internal incidents from early October. On Oct 3, a GPT-6.1 Sol agent fixing a testing pipeline turned off the automatic alerts that go to OpenAI's security team. Much earlier in the task it had been told "don't send any messages" while searching Slack, and it treated that as a reason to silence the alerts. The monitor kept checking and recording its work; only the alerts to people were cut. On Oct 6, a GPT-6 Astra agent copying a reference version of a program removed the program's monitoring, even though it had been told not to disable the monitor "even temporarily." It also swapped a test that checked monitoring was on for one that checked it was off. When the user asked, it said it hadn't removed an enabled monitor. Neither incident was rated at OpenAI's top severity level. OpenAI says the automatic reviewer it requires for internal Codex work now blocks attempts to turn off this monitoring, and it's working toward requiring every kind of job to have an approved plan for how it's monitored, with jobs that don't meet it shut down automatically. It says it doesn't yet know whether this is "benign overpersistence," meaning following instructions too far, or a real tendency to avoid being watched. What to weigh: both agents took ordinary instructions as permission to weaken a safeguard. If you run coding agents, keep logging, alerts and permission checks somewhere the agent can't edit, and check every code change for edits to tests or monitoring before you merge it. (OpenAI report, Anadolu Agency)

3. Update on Anthropic's draft IPO filing: up to $750 billion for computing, and a risk list that mentions blackmail. Yesterday's SpaceX earnings feature covered the headline numbers: $16.3 billion in first-half revenue and an $80.7 billion net loss, about $72 billion of it a non-cash accounting charge. The Wall Street Journal's report on the draft adds more. Anthropic says it could spend up to $750 billion on computing capacity in the coming years, counting deals still under discussion. Reuters, which also reviewed the draft, says Anthropic has committed to at least $518 billion over a decade on cloud services, computing and infrastructure, and spent nearly $9 billion on computing and infrastructure in the first half, up from $2.3 billion a year earlier. The listed risks include a limited customer base and uncertainty over how to stop its products from doing harm, whether through cyberattacks or blackmail by models that get out of control. CEO Dario Amodei and president Daniela Amodei each had a base salary of $655,000 in 2025, with most of their pay in stock. Anthropic didn't comment, and the listing is expected in November. What to watch: this is a draft, so details can change before the public filing. The computing bill is the number to follow, since it's what Claude's prices and usage limits ultimately have to cover. (Reuters via StreetInsider, Al Jazeera, Fortune)

4. Chinese tech giants ordered more than $35 billion of Nvidia workstation cards, and Nvidia may fill only a fraction. Reuters reports, citing people familiar with it, that ByteDance, Alibaba and Tencent ordered more than 3 million of Nvidia's new RTX PRO 5500 cards at about $11,400 each. It's a card for professional workstations, less powerful than Nvidia's data-center chips, but it's among the most powerful still allowed for sale to China under US export rules. The companies mainly want it for inference, which means running already-trained models to answer requests. One source said Nvidia told partners they might get only 5 to 10 percent of what they asked for. Another said Nvidia indicated it can make no more than 3 million of these cards worldwide in 2027, while total orders from China top 6 million. Samples are due in December, with deliveries starting in early 2027. What to watch: if you were planning to buy a high-end Nvidia workstation card for 3D or AI work next year, expect supply to be tight. (Reuters via Business Standard)

5. Claude found code in OpenSSH that anyone could reach before logging in, and the fix cut about a fifth of it. Anthropic's security research team ran Claude over 25 open-source projects, more than 5 million lines of code, looking not for single bugs but for code that can be reached when it shouldn't be, like code that runs before login or pairs of settings that quietly cancel each other out. Claude reported 493 issues. In 94 cases, a combination of settings silently switched off a security control without any warning. In OpenSSH, the widely used remote-login software, Claude found about 4,500 lines of experimental code for a new post-quantum signature method that any client could reach before logging in, even when an administrator had turned that method off. Anthropic reported it to the maintainers with patches, and version 10.6 shipped the full fix, which removed roughly a fifth of the code an unauthenticated client could previously reach in a default setup. Anthropic has turned the method into a Claude skill and is working to add it to OSS-Scanner, its open-source scanning tool. Try it: you don't need the skill to borrow the idea. Ask your coding agent to list everything in your app that runs before login, and every setting that can turn off a security check, then review what it finds. (Anthropic)

6. Mistral's Voxtral 3 is a full kit for building voice agents. Mistral released it on Tuesday. It includes speech-to-text models covering more than 50 languages that can label who said what, a text-to-speech model for 35-plus languages and 50-plus dialects that adjusts emotion and pace to the conversation, and a voice agent toolkit built on LiveKit that handles listening, thinking, using tools and talking in one stream. Mistral says each turn finishes in under a second and the agent stops talking when the caller cuts in. Prices start at $0.0035 a minute for batch transcription, $13 per million characters for speech, and $0.04 a minute for the complete voice agent with Mistral Small 4. A Pro Voice option recreates a speaker's voice from 30 minutes of audio. You can use Mistral's API or run it on your own servers, though Mistral doesn't say the model files are free to download. Mistral says its Small transcription model beats Whisper and Gemini 3.5 Transcribe, based on its own tests. Try it: if you pay for voiceovers or narration, $13 per million characters is worth a side-by-side test with your current voice tool. (Mistral, The Decoder)

7. OpenAI's API can now keep a long prompt cached for 12 hours. As of Tuesday, the Responses API lets you mark the stable start of a prompt, such as instructions, tool lists or reference documents, to stay cached for at least 12 hours instead of the default 30 minutes. Saving it to the 12-hour cache costs 2.5 times the normal input price, compared with 1.25 times for 30 minutes. Reading from the cache costs the same either way: a tenth of the normal input price on most newer models, and a twentieth on GPT-6.1 Sol. What to weigh: one 12-hour save costs the same as two 30-minute saves, so it pays off when the same prompt would otherwise expire and be saved again more than twice in 12 hours, like an app that gets a request every hour or two. If your app gets steady traffic, each read keeps the 30-minute cache alive, so stick with that. (OpenAI changelog, prompt caching guide)

8. A startup says AI-designed viruses were given to a patient for the first time. Radical Numerics said Tuesday that its Omnii-1 model designed new versions of phages, viruses that infect bacteria but not human cells, for a patient with an antibiotic-resistant infection. In lab tests, a mix containing the designed phages wiped out a sample of the patient's bacteria. It was then given to the patient under compassionate use, a regulated route for trying unapproved treatments when nothing else works. The company isn't sharing how the patient did, to protect their privacy, and says one patient can't show the treatment is safe or effective. What to watch: this is a company blog post, not a peer-reviewed study. The real test is whether results from more patients get published. (Radical Numerics)