OpenAI held DevDay 2026 at Fort Mason, San Francisco on September 29, packing more than twenty announcements into a single keynote. The releases fall into three groups: product forms (Dots, Space, Pages), price and speed (GPT-6.1 Sol, Astra Ultrafast, the $500 tier), and distribution (Agents API computer use, Sign in with ChatGPT, Plugin extensions).
A note on sourcing: OpenAI's own recap page was unreachable at the time of writing; figures below come from CNBC, The Next Web, 9to5Mac, The Decoder, BGR, and Engadget coverage, plus an explainx.ai roundup. Pricing and availability may shift once docs are complete.
The always-on category: what Dots are
Dots are each user's resident agent. Running on GPT-6 Astra, each comes with a dedicated cloud computer and browser, connects to 4,000+ app connectors, and keeps working toward a goal you set, with rules for what it can do without asking.
Rollout starts with Pro and Business Premium users; enterprise editions (including Edu and Healthcare) require admin opt-in. Each plan includes one Dot; Free and Plus are excluded this round.
Autonomy has boundaries: spending money and sending messages trigger confirmation prompts. A persistent Dot sitting between a browser and app credentials makes prompt injection a long-term risk—connector permissions should be governed like production IAM.
The hardest numbers: GPT-6.1 Sol pricing
GPT-6.1 Sol upgrades GPT-6 Sol. OpenAI says it approaches GPT-6 Astra on agentic coding, computer use, and professional work, at one-fifth of Astra's standard I/O price:
| Token type | GPT-6.1 Sol | Comparison | |---|---|---| | Input | $2.00 / 1M | Sonnet 5.5 also $2 | | Cached input | $0.10 / 1M | 95% below standard input; half of GPT-6 Sol's old cache price | | Output | $10.00 / 1M | Sonnet 5.5 also $10 |
The real significance is that it lands on the same price tag as Anthropic's Claude Sonnet 5.5, announced the day before. Opus 5.5 (September 22) is $4/$20. On the third-party Terminal-Bench Science 0.1 benchmark, GPT-6.1 Sol scored 57% vs. Opus 5.5's 58.7%—one point apart at half the price.
With prices aligned, model selection shifts from "who is cheaper" to "who is more reliable on your workload." Three factors deserve separate modeling: cache hit rate (the $0.10 cached-input price only pays off when your prompt prefix is stable), tool-call reliability, and latency.
Speed becomes purchasable
Ultrafast is a paid speed tier. GPT-6 Astra runs up to 8x faster in Codex (~300 tokens/second) and up to 6x on the API side, available now in the API, ChatGPT Work, and Codex Pro 500/enterprise. API surcharges are expected; check the pricing page before switching.
The companion $500/month tier is OpenAI's highest usage-limit tier and exclusively includes Astra Ultrafast—previously leaked as "Pro Max."
Reading the September 22 Sol/Luna price cuts, this event, and the previous day's cancellation of an October GPT-6.1 Astra release together: OpenAI's frontier weights are pausing while the price-performance ladder keeps compressing downward.
Distribution moves worth watching
- Agents API computer use. The Agents API (public beta since September 10) can now operate software, not just call structured tools.
- Sign in with ChatGPT. Users link their OpenAI account to third-party apps like Notion and spend their own ChatGPT quota in-app. Inference cost moves from your API key to the user's subscription—arguably the biggest distribution rule change of the day.
- Plugin extensions. Developers can ship "full apps" inside ChatGPT and Codex: sidebars, custom file viewers, editors, dashboards. OpenAI cited 32 Marketplace partners.
- Codex Security Cloud. Scans repos for vulnerabilities and prepares fixes; a cloud version of the open-source Codex Security from August.
- Decisions API (limited preview). Described as "smart if-then statements" for routing logic; details pending docs.
- ChatGPT Space and Pages. Team workspaces shared with Dots; Pages are documents co-produced by humans and agents, with charts and visualizations.
- Private Intelligence. A preview of stronger data controls; architecture and availability undisclosed.
- Dots: only one per plan, no expansion timeline.
- Plugin extensions: distribution terms unpublished.
- Decisions API: limited preview only.
- Private Intelligence: no architectural details.
- Ultrafast API surcharge: no numbers given.
- The Terminal-Bench Science 0.1 scores come from one-sided citations; a single benchmark won't represent your workload.
- explainx.ai, "OpenAI DevDay 2026: Every Announcement, Explained for Builders," 2026-09-29
- 9to5Mac, "OpenAI makes 20 announcements at DevDay event including always-on agents called dots," 2026-09-29
- The Next Web / CNBC / The Decoder / BGR / Engadget on-site coverage (GPT-6.1 Sol pricing and cache price)
- Anthropic Claude Sonnet 5.5 (2026-09-28, $2/$10) and Opus 5.5 (2026-09-22, $4/$20) pricing pages
Why "Sign in with ChatGPT" matters most: in traditional SaaS, AI features are billed to the vendor and costs scale linearly with users. When quota follows the account, marginal inference cost shifts to the user's subscription—a business-model change, not just an integration change.
Why it matters
1. Aligned price tags push competition onto evaluation—cache hit rate, tool-call success rate, and latency must now be benchmarked on your own workload. 2. Always-on agents change the risk model: something is running continuously with your browser and app credentials. Permission governance becomes a prerequisite, not an option. 3. Cost structures change: Sign in with ChatGPT can move inference costs to the user side, rewriting unit economics for a class of AI apps. 4. Compute-for-speed is now explicitly priced, signaling that agent-loop bottlenecks have moved from "can the model do it" to "how long does it take."
What remains unclear
Bottom line: this DevDay shipped no new frontier model. It made concrete something already underway—agents are becoming resident, the price ladder keeps dropping, and speed is now something you can buy. What genuinely needs recalculating is your evaluation methodology and permission boundaries.
---