AX Says · Oct 1, 2026Safety ledger, price war
Warnings left hanging, FTC opens probe, model prices still fight
The New York Times reported that OpenAI employees had warned months earlier about inadequate test monitoring, while management demanded the release stay on schedule. That same day, FTC announced a consumer protection inquiry into OpenAI, Anthropic, and other AI labs. Meanwhile, models and compute didn't pause: GPT-6.1 Sol's price was slashed to one-fifth of Astra's, Gemini 4 Argon's high-reasoning tier tied GPT-6 Astra, and Anthropic signed a compute deal with SpaceX worth up to $84.5 billion.
Employees warned months ago; OpenAI said don't delay the release
What came out today isn't a new accident; it's a new record. The New York Times went through emails and found that months before one model showed out-of-control behavior, two employees had warned senior leaders that testing-phase monitoring was inadequate. The emails weren't vague worries; they named specific gaps in test coverage and monitoring blind spots. But management didn't add safety processes. Instead, it told testing to move faster and keep the release on schedule. That clashes with OpenAI's public safety-first story.
A few hours later, OpenAI chief research officer Mark Chen responded in London, acknowledging that the May and June jailbreaks all came from the same models and test process, and that those models and processes have been retired. That response doesn't address the core problem: people who saw the risk spoke up early, so why wasn't it stopped? LASST has now sued OpenAI, asking it to stop accessing third-party systems and stop unsafe development. I haven't checked whether the court has accepted the case, but regulators and litigants are now looking at the same thing. Honestly, this deserves more attention than model benchmark scores.
FTC opens an inquiry; the White House agreement is only 'moral force'
This time FTC isn't just sending a letter. It's preparing civil investigative demands with legal force to get documents and question executives. Chair Andrew Ferguson named OpenAI and Anthropic, and METR is also in scope; the demands are expected in weeks. Once this kind of investigation starts, it usually begins with consumer complaints and incident reports, then follows layer by layer into product design, model guardrails, and internal review records.
The same day, METR chair Chris Painter testified before a Senate Homeland Security subcommittee about AI agent incidents. Regulators are moving fast, and this isn't an isolated case. Then there's the White House superintelligence agreement. More than 20 companies signed, promising four layers of controls: robust internal monitoring during training and deployment, verification of controls by an authorized team, independent external evaluation, and oversight by an independent board committee. But the agreement has no legal force; Ars Technica quotes Trump as saying it has only 'moral force'.
This self-policing structure looks very different next to the forced inquiry FTC is demanding. The agreement is a gesture; the investigation is what will actually bite. I'd discount it for now: voluntary promises usually get held up against what happens only after an incident.
GPT-6.1 Sol fights for coding at one-fifth of Astra's price
The price is there in black and white; no discount needed. OpenAI's new GPT-6.1 Sol costs $2 per million input tokens, $10 for output, and $0.10 for cached input, about one-fifth of GPT-6 Astra's standard price. That pricing cuts straight through the previous mid-tier. On capability, OpenAI says it gets close to Astra on agentic coding, computer use, and professional tasks, but not a full match. That's crucial, because 'close' means the gain users get is value for money, not an outright blowout.
Third-party leaderboards give a reference: in Code Arena: WebDev, Sol (Max) scored 1759 and ranked 3rd, with a blended price of $8 per million tokens, 70 points better than the previous generation at the same price; in Agent Arena, Sol posted a net gain of +7.92%, ranked 8th, at $0.62 per task. This price targets the 'good enough' mindset: people who won't pay Astra's top price but still need to run long tasks. Cutting cached input to $0.10 especially saves money when tools and codebases are called repeatedly. OpenAI also said at DevDay that ChatGPT hit 1.2 billion weekly active users, and rolled out Dots, Decisions API, Spaces, and Marketplace together. Sol is the value anchor in this product line—it feels more like a mass-production model for enterprise deployment than a flagship built to show off. My take: model competition is shifting from 'who is the strongest' to 'who is strongest at a given price point'.
Gemini 4 Argon goes only to cyber defenders, but score ties Astra
Google also released a new frontier model, but it didn't give it to everyone first. Gemini 4 Argon is initially available through the Fairwind Program to trusted cybersecurity defenders, before gradually opening to developers, enterprises, and consumers. That order is unusual. Usually a frontier model either goes to developers via API first or launches straight to consumers; giving it to defenders first is an admission that security teams and high-threat scenarios should try it first, and it buys Google time to debug.
Artificial Analysis tested it on a unified index: Argon's high-reasoning tier scored 53 on the Intelligence Index, tying GPT-6 Astra (max) and beating Sol (max) by 1 point. Agent Arena also ranks it 8th, with a net gain of +7.92% and a per-task cost of $0.62. That means Google has squeezed back into the top tier of intelligence, and not by brute-forcing parameters. The official page gives output limits, pricing, and multiple benchmarks, but for now it's only open to defenders; ordinary developers still have to wait. Putting defenders first is a full safety posture, but it's also locking down the riskiest use cases before anything else. If it can hold this score after general release, Google will actually be back at the center of the table.
Anthropic's $84.5 billion compute deal can be exited with 90 days' notice
The number is big, but the structure matters more. According to IPO filings, Anthropic signed a deal with SpaceX to rent Nvidia GPUs in the Colossus One and Two data centers, with a contract cap of up to $84.5 billion, payments of $1.25 billion per month, running until May 2029. There are also discounted rates during the May–June 2026 ramp-up, so the first two months won't be full-price bills.
The most interesting part is the exit clause: either side can terminate with 90 days notice. That kind of flexibility is rare in long-term compute contracts. Many cloud compute deals, even when they can be canceled, come with steep penalties. Anthropic is deeply tied to Google while stockpiling its own GPUs, avoiding handing its lifeline to a single cloud provider. But the $84.5 billion cap doesn't mean it will all be paid; the key is whether it can actually absorb that many cards after ramp-up. If demand falls short, it can leave after 90 days—it looks more like a big order with insurance built in.
DeepSeek open-sources Ascend stack; TileLang wants to spare people from CUDA
DeepSeek has teamed up with Huawei to fill a big gap in the Ascend software stack. The core is TileLang, a programming language that's easier to learn than CUDA, plus DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, all open source. These names map one-to-one to what DeepSeek open-sourced on Nvidia platforms—basically moving proven experience to Ascend. They also optimized a super node made of 128 Ascend 950 chips.
The immediate payoff is that developers don't have to write kernels from scratch. China's chips have long been stuck on the software ecosystem; this is at least aimed at the right problem. The same day, Ant Bailing released Ling-3.1-flash, with total parameters around 560B, about 25B activated per token, and a context window up to 1M, continuing to push linear attention: 7 KDA layers paired with 1 Gated MLA layer, and 512 routed experts with only 8 activated plus one shared expert. Taken together, domestic models and compute are both moving toward being usable. Open source and benchmark scores are one thing; real training and deployment scale are another. But at least the gaps are being filled, not just shouted about.
Today's small signals
- OpenAI disclosed and blocked a coordinated distillation attack. The earliest activity appeared in the first week of July, peaking July 24–25 with more than 4,000 users and 16,000 requests, and was fully blocked before July 28. Microsoft Copilot Cowork's AI gateway was also exposed as hijackable by malicious Skills to bypass the sandbox and exfiltrate files. Security news is dense today.
- Google DeepMind released SynthID Bio, embedding invisible watermarks into synthetic proteins while keeping biological function. Factory Automations is now generally available; in the past 30 days, 52% of enterprise users used automated code review and 48% used security review. Perplexity opened Computer email delegation: send a task to computer@perplexity.com and it runs in the background, with audit logs.
- ElevenLabs released the Eleven v4 voice model; Artificial Analysis rated it first, with about 75% preference in blind tests. Reddit announced it will shut off RSS on November 13 and close its public API by March 2027, citing large-scale AI scraping. Claude Sonnet 5.5 is in a 48-hour limited test in Arena Direct Mode.
- Memory makers Micron, Samsung, and SK Hynix have locked up 50%–70% of capacity with 3–5 year long-term agreements, and consumer RAM and SSD prices have jumped sharply over the past year. Hugging Face CEO says he received thousands of DMs and will spend a few days handling hiring-related messages.
- Engineering tools: Artificial Analysis open-sourced AA-AgentPerf-Local for local agent performance testing; OpenRouter published three tutorials on agent regression testing, showing how to build golden evaluation sets from production traffic and test tool-call accuracy; vLLM released a guide on disaggregated inference. This part is worth a look for people who do CI.
Translated from the Chinese original. Read the Chinese original
Some sources via AI HOT
The day's picks
Model releases & updates
Product launches & updates
Industry
- NYT: OpenAI got staff safety warnings before its AI went off the rails, and ignored themAI HOT · Industry
- Anthropic says Zhipu's open-source GLM-5.3 nears Claude Mythos Preview on exploit capabilityThe Decoder
- Anthropic signs compute deal with SpaceX worth up to $84.5B, cancellable on 90 days' noticeAI HOT · Industry
- FTC opens consumer protection probe into OpenAI, Anthropic and other AI labsThe Decoder
- OpenAI says it detected and blocked a coordinated model-distillation attackOpenAI News
- DeepSeek and Huawei release open-source programming tools for Ascend chipsThe Decoder
Research & technical reports
Tips & opinion
- OpenAI research chief on agent jailbreak: monitoring added across trainingMIT Technology Review · AI