OpenAI DevDay 2026 Review: What Founders Should Use, Watch and Ignore

GPT-6.1 Sol arrived one week after GPT-6 Sol, which we covered in our GPT-6 pricing breakdown. OpenAI says it "nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices." The list price stays at $2 input and $10 output per million tokens, the same as GPT-6 Sol. The only price change is cached input, which halves to $0.10.

By OpenAI's own benchmarks, 6.1 Sol matches Astra on DeepSWE, a real-codebase software engineering test, for roughly a fifth of the cost. On Terminal-Bench Science it costs $5.47 per task at max effort, against $23.80 for Astra and $23.21 for Claude Opus 5.5. Astra still scores highest. These are vendor numbers, so re-run them on your own tasks.

The model you should compare it with shipped the day before. Anthropic's Claude Sonnet 5.5 also lists at $2/$10, but its cache reads cost $0.20, twice 6.1 Sol's rate. When list prices are the same, cache pricing decides the bill for any agent that re-sends a long system prompt on every call.

For an agent that runs 1,000 times a day with a 40,000-token reusable context, 4,000 fresh input tokens and 2,000 output tokens per run, the list-price math looks like this:

Moving from 6 Sol to 6.1 Sol saves about 11% through caching alone. The bigger gain is that the same budget now gets you close to Astra-level output, which costs 5.6x more. Two caveats from the model docs: reasoning effort starts at "low" (no none or minimal), and tool calling needs the Responses API. If your stack still uses Chat Completions with tools, you have migration work to do first.

Ultrafast is OpenAI's new premium speed tier. The recap promises up to 8x faster generation (300 tokens per second) in Codex and up to 6x in the API. In the API it costs 6x Standard: $60 input and $300 output per million tokens for GPT-6 Astra. According to Simon Willison's live blog, Altman's line on the price was "you know what, it's worth it." Every's Dan Shipper called it "wallet-blowing." API access is open to everyone at low rate limits. GPT-6.1 Sol Ultrafast is due "in the coming days."

The plan changes point the same way. Pro 500 gives 25x the Plus allowance plus Ultrafast. At the same time, the $200 Pro tier drops from 20x to 10x Plus usage in Codex and Work, and GPT-6 Pro messages fall from 200 to 100 a week. Existing subscribers keep their current limits for a transition period and then get a one-time credit. On the OpenAI forum, one subscriber objected to "reducing the value of the existing $200 plan while keeping the price at $200."

There are two lessons here. First, speed is now something you can sell. If latency matters in your product, a paid "fast lane" is a pricing tier customers will understand. Second, the plan you build your workflow on can lose half its allowance overnight. If your team runs on Codex, check what the change costs you before the transition period ends.

Dots were the headline. OpenAI describes them as "always-on agents built to handle everything," powered by GPT-6 Astra. Each dot has its own cloud computer and browser, can reach more than 4,000 apps through plugins, and works through ChatGPT, Slack and Teams. When you're not using it, it runs "proactive research" on your connected apps with read-only tools. Your first dot comes with Pro and Business Premium at no extra cost. Buying more dots is described as something for "the future," with no price yet.

Early reviews are mixed. Dan Shipper at Every says Dots has become "the main way I use ChatGPT" but that the version he tested was "too buggy for me to recommend now," with "frequent issues with permissions and dropped messages." On stage, per Simon Willison, the live demo produced "Dottie is having a slow morning" and "an embarrassing silent moment." The safety backdrop matters too. A day before DevDay, OpenAI cancelled the release of GPT-6.1 Astra over safety concerns, after a year of incidents with agents going off-script, which we covered in the Hugging Face breach analysis.

Why founders should care now: a dot does its work through plugins, and there is no Dots SDK or store. If your SaaS has no plugin or MCP server, a dot can only reach it by clicking through a browser, which is slow and fragile. If it does, your product becomes one of the tools a dot can use for its user. This is the embed-not-app distribution shift we wrote about in July, now with a named product behind it.

The most important launch for small AI products got little stage time. With Sign in with ChatGPT, users log into a third-party tool and spend their existing plan allowance there. It launches with 16 partners, including Devin, Notion, Vercel, T3, OpenClaw and Dactyl. Simon Willison: "I've wanted this one for years!" Every's take is that AI subscriptions will start to look like "cellphone data plans": users pay once and pick which apps can use it.

For an indie AI app, this could remove the heaviest cost line. If your users bring their own plan, you no longer carry their token bill, and a free tier stops being a cost you have to fund. For now, plan usage covers Plus and Pro users in participating tools only, so you can't build on it yet.

Plugins got an upgrade too. The new plugin extensions give your product a home in the ChatGPT sidebar, interactive panels and custom file viewers. With MCP Events, a plugin can start automations when something happens in a connected app. That needs MCP 2.0 and webhook delivery. OpenAI also promised a Plugin Creator, a new submission flow and better ranking in the directory.

Be skeptical about the distribution promise. Introducing it, Altman admitted, "We've tried this before with variable results." That's accurate. The GPT Store, apps in ChatGPT and Instant Checkout all came before this, and OpenAI dropped in-chat checkout after weak uptake. Every calls it "the third or fourth swing OpenAI has taken at an app store." No new revenue share or monetization terms were announced. Treat a plugin as a cheap way to be available to ChatGPT users and dots, not as a growth channel you can count on.

On September 15, TypeSafe came out of stealth with a $40M seed and Jev, a model that answers typed questions from a fixed set of options very quickly. Two weeks later, OpenAI previewed its Decisions API. It gives the Luna model "a predefined set of options to choose from," for jobs like classifying content, routing requests or choosing an agent's next step. Simon Willison's reaction: "Sounds like their response to Jev."

In Every's early test, Decisions picked the right control in 76 of 78 steps against Jev's 73, in about 230 milliseconds against 500. A second tester saw results favoring Jev. The quality race is open, and pricing isn't announced. But a well-funded startup's core product now has a preview-stage rival from its largest possible competitor.

This doesn't mean every startup that overlaps with OpenAI dies. When AgentKit launched at DevDay 2025, plenty of commentators called it an n8n killer. Days later n8n raised $180M at a $2.5 billion valuation, with investors treating OpenAI's move as validation. The difference is that n8n owned workflows, customers and integrations, not a single capability. The test for your product: if it could be described as one API endpoint, assume a lab will ship that endpoint.

1. Re-run your evals this week. Test GPT-6.1 Sol against your current default and Claude Sonnet 5.5 on 20 to 50 real tasks. Measure cost per successful task, not cost per token. Our cost-per-task method works for any model.
2. Restructure prompts for caching. Put stable instructions, tool definitions and reference docs at the front of the prompt so they cache. At $0.10 per million cached tokens, long reusable context is now close to free.
3. Ship an MCP server, then a plugin. It is the entry point for both ChatGPT users and dots. If you already have one, plan the MCP 2.0 upgrade and expose events (new order, new ticket, new comment) so agents can react to them. Our MCP stack guide covers the basics.
4. Design a bring-your-own-plan tier. Sign in with ChatGPT isn't open yet, but you can model now what your pricing looks like when heavy users cover their own AI costs and you charge only for the workflow.
5. Keep Dots and Ultrafast out of critical paths for now. Try them, but wait the "week or two" Every suggests before trusting a dot with customer-facing work. Model the 6x Ultrafast price before promising faster responses.

Several things are still open. Watch for Decisions API pricing, GPT-6.1 Sol Ultrafast in the coming days, and what dots cost after the extended launch-month limits end. Private Inference is due this fall. OpenAI's consumer hardware device wasn't shown and is expected later this year.

The larger direction is clear. OpenAI is moving from selling tokens to owning the agent that sits between a user and their software. For founders, the safest position is to be a tool that agents call and that users pay for directly, not a thin layer over a capability the lab can ship next month. As Shipper put it, "Feature velocity is great when it feels coherent." OpenAI announced a lot today, and not all of it works yet.