OpenAI DevDay: Dots Agents and $2 GPT-6.1 Sol — AI Digest

OpenAI DevDay: Dots Agents and $2 GPT-6.1 Sol — AI Digest

OpenAI launched dots, always-on agents wired into 4,000+ apps, shipped GPT-6.1 Sol at $2 per million tokens and cut Pro limits. Anthropic filed for an IPO valued above $2 trillion.

Top stories

OpenAI launches dots: always-on GPT-6 Astra agents wired into 4,000+ apps

At DevDay on September 29, 2026, OpenAI introduced dots — agents that run continuously, each on its own cloud computer and powered by GPT-6 Astra. An always-on agent is an AI assistant that keeps working on delegated tasks in the background instead of waiting for the next prompt. OpenAI announced that a dot connects to 4,000+ apps plus Slack and Teams, and users set the boundaries: what it may do alone, what needs approval and what it must never do. Connecting your own machine is optional. Dots ship to Pro, Business Premium and Enterprise; OpenAI's Tibo says the primary dot's own work does not draw on plan usage, while the Codex tasks it spawns do. One early tester reports a dot negotiating with customer service to cut about $500 a year in charges. The Rundown AI has a full breakdown. Ready-made agent workflows for leads, mailings and reports live in the AI SKILLS automation templates.

GPT-6.1 Sol: $2 per million input tokens, 1 point behind Astra

OpenAI released GPT-6.1 Sol at $2/$10 per million tokens, with cached input at $0.10 — a 95% discount on repeated context. OpenAI pitches it as "near-Astra intelligence for a fifth of the price". Independent testing by Artificial Analysis puts GPT-6.1 Sol 1 point below GPT-6 Astra on its Intelligence Index at $0.72 versus $3.26 per task; the hallucination rate fell from 60% to 54%, though it uses 10–30% more output tokens than GPT-6 Sol. In a planted-bug test, Pawel Huryn hid 105 bugs in two repositories: GPT-6.1 Sol found 44 for $6.56, GPT-6 Astra found 45 for $33 and Opus 5.5 found 41.7 for $58.53. For speed, OpenAI added Ultrafast mode — up to 8x faster in Codex and 6x in the API at 6x the price, which for Astra means $60/$300 per million.

ChatGPT plans re-tiered: the $200 Pro now gives half as much, and a $500 Pro arrives

OpenAI re-tiered ChatGPT usage to Plus 1x, Pro 100 5x and Pro 200 10x, and added a Pro 500 plan at 25x. OpenAI's Tibo posted the new multipliers. The old $200 Pro plan sat at roughly 20x, so heavy users lose about half its value, and the change drew heavy backlash. On Reddit, users call it the end of subsidized compute: the $500 plan now buys roughly what the $200 plan used to. At the same time, OpenAI opened Sign in with ChatGPT for partner apps, so plan quota can now be spent in Devin, Nous Portal and Hermes and T3 Code. The takeaway for anyone on a premium plan for the limits: check what you actually use — GPT-6.1 Sol through the API with caching may now cost less than the subscription.

Anthropic files for an IPO at a valuation above $2 trillion

Anthropic has filed for an IPO at a potential valuation above $2 trillion, with Q2 2026 revenue of about $11.5 billion and annualized revenue (ARR) above $65 billion. The filing is reported to list $518 billion in compute obligations and roughly 80 pages of risk factors. For comparison, OpenAI's annualized revenue is reportedly nearing $70 billion. A Reddit thread citing Reuters puts Anthropic's 2025 net loss at $42 billion and expected 2027 spending near $500 billion, and commenters question whether today's AI economics justify such a valuation. The Rundown AI covers the safety risks disclosed in the filing.

Numbers and facts

Different views: are GLM-5.3's open weights dangerous?

In a late-September 2026 report, Anthropic showed that Z.ai's open-weight GLM-5.3 built a working browser exploit in 50 of 410 attempts, versus 56 for the closed Claude Mythos Preview. The report is on Anthropic's site and under discussion on Reddit. Abliteration is the removal of an open model's refusals by editing its weights directly.

Tools and techniques

In brief