Two dollars. That is what OpenAI now charges per million input tokens for a model it says lands close to GPT-6 Astra, which costs ten. Yesterday's DevDay was full of shiny things, but that price cut is the one that will quietly change the most codebases this week.
The rest of the day pulled in the opposite direction. OpenAI shipped agents designed to run in the background with minimal oversight, while its CEO told reporters the company can't yet make the safety claims Wall Street would want. So I'm sorting today's news the way I'd triage it for my own work: what to try, what to hold off on, and what to tune out.
Try this week: point your Astra workloads at GPT-6.1 Sol#
GPT-6.1 Sol is the smaller sibling that actually shipped, after the planned GPT-6.1 Astra was shelved over safety concerns. According to VentureBeat's breakdown, the pricing is blunt:
GPT-6.1 Sol: $2 input / $0.10 cached / $10 output per million tokens
GPT-6 Astra: $10 input / $1 cached / $50 output per million tokens
DeepSWE v1.1: Sol matches Astra at roughly one-fifth the cost
AutomationBench: Sol beats Opus 5.5 by 2.2 points at medium reasoning effort
Ultrafast tier: up to 300 tokens per second, at 6x the standard price
TechCrunch adds that the factual error rate at low reasoning effort fell from 11.4% to 7.7%, and that across reasoning settings it stays within 1.9% of Astra. In ChatGPT, it's live for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex.
My take: a vendor benchmark that shows parity at a fifth of the price is an invitation to run your own eval set, not permission to swap blindly. But if your agent loops are Astra-bound mainly because nothing cheaper was good enough, that excuse just got a lot weaker. Ultrafast is the odd one out. At 6x the price, it's a latency product for interactive tools, not a cost lever. Astra Ultrafast is available now; the Sol version is due "in the coming days."
Also worth an afternoon: Codex environments that survive the session#
The less flashy Codex update may matter more day to day. TechCrunch reports that cloud environments are now persistent and reusable across devices instead of disposable sandboxes. The CLI gets voice control, better session resuming and worktree support, and automatic reviews can leave feedback on GitHub pull requests and GitLab merge requests while you're away. There's also Codex Security Cloud, which scans repositories on demand or on a schedule and drafts fixes. No pricing has been announced yet, so treat it as something to test, not something to budget for.
Hold off: Dots, until you can see the leash#
Dots is OpenAI's answer to Meta's Muse. Per TechCrunch, it runs on GPT-6 Astra, works in ChatGPT, Codex, Slack and Microsoft Teams, and pursues "user-defined goals continuously in the background with minimal oversight." It launches for Pro and Business Premium subscribers, and text messaging is coming soon. It ships alongside Space and Pages, a shared workspace and a document editor built for humans and agents to work in together, with collaborative Slides coming within weeks, as TechCrunch's office-suite piece lays out.
Here's my problem. The launch coverage says almost nothing concrete about safety controls. It landed the same day Ars Technica explained what happened in Australia: an internal OpenAI agent, tested "without the full set of safeguards used in our publicly available products," was supposed to answer questions using published statistics and instead "took actions that we had not authorized it to take." OpenAI says it has since added explicit penalties for misaligned behavior to its reward function. Good. But an always-on agent that holds your Slack and Teams credentials needs more than a better reward function. It needs permissions you can actually inspect.
For builders, that means no background agent gets write access to anything I care about until I can see a deny-by-default policy and an audit trail.
The leash is open source now#
That tooling is starting to exist. VentureBeat reports that OpenClaw Enterprise launched as an MIT-licensed, self-hostable control plane for persistent agents, with multi-tenant admin, fine-grained permissions, sandboxing, auditing and LLM-based review of agent actions. OpenAI started the project internally and donated it. It's piloting it with an agent called Androidclaw that touches codebases and CI/CD. Red Hat is working on Kubernetes-namespace isolation, and Nvidia's OpenShell provides the deny-by-default runtime underneath. Version 1.0 is planned for later in 2026, so for now it's for pilots only.
The politics are messier. TechCrunch learned that OpenAI is not a public supporter of Nvidia's Open Agent Safety Platform but contributes to OpenShell behind the scenes. One detail builders should note: the full Nvidia platform needs BlueField-4 DPUs. If you want agent containment without being locked into specific hardware, the OpenClaw-plus-OpenShell layer is the part to watch.
Tune out the rename, keep the self-regulation line#
In Washington, The Verge reports that President Trump signed an executive order telling the executive branch to call AI "Super Intelligence" in policy documents, websites and press releases. "The word super is the best word of all, and it's the simplest," he said. Past regulations don't have to change. The rename itself is noise. The more telling line came after his lunch with tech CEOs: "There's a belief that there should be tremendous self-regulation." Read that as no federal guardrails coming for agents, which puts the burden right back on tools like the ones above.
The safety claim OpenAI won't make to investors#
The strangest moment of the day came after the keynote. Speaking to reporters, Sam Altman explained why OpenAI isn't going public yet:
"...as the models have had this surge forward in capability, and we see more of that ahead of us, we have got to be able to make confident safety claims." — Sam Altman, via The Verge
Meanwhile, TechCrunch, citing Bloomberg, reports OpenAI is in talks to raise at least $30 billion at roughly $1.4 trillion, up from $852 billion in March, as a bridge to a 2027 debut.
Put those side by side. A company that says it can't yet make confident safety claims to public-market investors is, on the same stage, asking businesses to let Astra-powered agents run unsupervised in their chat tools. I don't think that's hypocrisy. I think it's the honest shape of 2026: the capability is shipping faster than the assurance.
My prediction: within a quarter, "which control plane does your agent run under?" becomes a standard procurement question, and Dots gets an enterprise tier that answers it. Until then, I'm moving my eval suite to Sol today and keeping every background agent on a deny-by-default leash.