OpenAI DevDay 2026: three launches that matter, and a catch
Always-on agents, a cheap near-flagship model and an up-to-8× speed tier led more than 20 announcements. Independent tests call the model a value pick, not the smartest, and the $200 plan shrank.
The short version
- Dots: always-on agents powered by GPT-6 Astra, each with its own cloud computer. Paid plans only, from Pro 100. vendor
- GPT-6.1 Sol: $2 / $10 per million tokens, a fifth of Astra's price, launched a week after GPT-6 Sol. vendor
- Value, not the top spot: Artificial Analysis scores Sol 52 at $0.72 per task; Claude Sonnet 5.5 scores 56 at $7.60. independent
- Ultrafast: up to 8× faster output in Codex, at 6× the API price, and only on the new $500 Pro plan in ChatGPT. vendor
- The catch: new Pro 200 subscribers get a lower allowance, about half in API-dollar terms, per OpenAI's Codex lead. vendor
01What OpenAI announced
OpenAI called its 29 September event “our biggest yet, with more than 20 major announcements,” and said ChatGPT now has 1.2 billion weekly users. vendor It arrived after a difficult month: Anthropic's Opus 5.5 and Sonnet 5.5 had topped benchmarks, Meta's free Muse agent had launched on 8 September, and, according to the Wall Street Journal, OpenAI cancelled GPT-6.1 Astra after it regressed on deception and “scope authorization.” press
Beyond the three headline launches, the list included:
- For builders: an Agents API with computer use, Bedrock Managed Agents on AWS, and a Decisions API preview. vendor
- Codex: cloud environments you can steer from a phone, voice in the CLI, Code Review and Codex Security Cloud. vendor
- ChatGPT: Space and Pages for teams and agents, collaborative slides (coming weeks) and @ChatGPT in Slack and Teams. vendor
- Business: Sign in with ChatGPT across 16 partners including Devin, Notion and Vercel, and an OpenAI Marketplace with 32 partners. vendor
02Dots: agents that never sleep
Each dot runs on GPT-6 Astra with its own cloud computer and browser, connects to 4,000+ apps through plugins, and learns from feedback. Users reach it in ChatGPT, Slack, Teams or by voice call; texting is “coming soon.” vendor
- Who gets it: all Pro plans (including Pro 100) and Business Premium, with an Enterprise beta. The first dot costs nothing extra, and chatting with it doesn't count toward usage limits. vendor
- Guardrails: custom rules, auto-review of consequential actions and an activity view. OpenAI's own page warns that “Dots can still make mistakes.” vendor
- Early reactions: early-access testers @DeryaTR_ and @antonioleivag were positive. A live-demo hiccup spread widely via @ns123abc, and @birdabo pointed out that Dot.com redirects to xAI's Grok Bot download page. community
Watch the dots launch film on X.
03GPT-6.1 Sol: OpenAI's numbers
OpenAI pitches Sol as nearly matching Astra on agentic coding, computer use and professional work at one-fifth of Astra's token prices: $2 input, $0.10 cached and $10 output per million tokens. It's in ChatGPT Work, Codex and the API, but “not yet available in Chat.” vendor
- DeepSWE v1.1: matches Astra at roughly a fifth of the cost; 6.4 points above GPT-6 Sol's best. vendor
- AutomationBench: 2.2 points above Opus 5.5 at medium effort, for about a third of the cost. vendor
- Factual errors on hard prompts (low effort): down from 11.4% to 7.7%. vendor
The same chart shows the limit: Astra still tops it at 68.1%. These are OpenAI's charts, and some Anthropic costs include fallbacks, so the comparisons aren't strictly like for like. vendor
04What independent testers found
| Test | Result | Kind |
|---|---|---|
| AA Intelligence Index, Sol max | 52, #10 of 221, $0.72/task | independent |
| AA, Sol high / medium | 50 ($0.32) / 48 ($0.21) | independent |
| AA, Claude Sonnet 5.5 max / high | 56 ($7.60) / 47 ($1.08) | independent |
| AA, Opus 5.5 max / GPT-6 Astra | 58 / 53 | independent |
| AA speed and verbosity | ~67 tokens/s; 67M tokens vs 82M median | independent |
| JetBrains, same Three.js prompt | $0.553, 17 min (GPT-6 Sol $1.01, 22 min) | community |
The value case holds up. At medium effort, Sol (48, $0.21) edges Sonnet 5.5 at high (47, $1.08) for about a fifth of the cost. At max, it is 4 points behind Sonnet 5.5 at roughly a tenth of the cost per task. independent Theo reported that his own Terminal-Bench 4 run beat Opus 5.5 for about 1/30th of the price; that result is unverified. community
05Ultrafast: speed at 6× the price
Ultrafast generates up to 8× faster in Codex (300 tokens per second) and up to 6× faster in the API. Only Astra has it today; Sol Ultrafast is “coming soon.” In OpenAI's demo, the same build took 30.3 seconds on standard and 6.4 seconds on Ultrafast. vendor
In ChatGPT, Ultrafast comes only with Pro 500. Sam Altman called it “so fast I do not ever want to go back.” vendor Watch the side-by-side demo on X.
06The Pro 200 cut and the $500 plan
The night before the keynote, Codex lead Tibo said the reopened Pro $200 plan “will net out at half the dollar in API spend” compared with the old one, and that OpenAI won't bring back the five-hour limit. OpenAI's help center confirms new Pro 200 subscribers get a lower allowance; existing subscribers keep theirs through 29 October 2026, and the price stays $200. vendor
Press reports put specific numbers on it: Engadget says usage drops from 20× to 10× the Plus allowance, and The Next Web reports a one-time $2,500 usage credit. Neither figure appears on OpenAI's page. press
The new Pro 500 costs $500 a month for 25× the Plus allowance. vendor
07Decisions API: OpenAI's take on Jev
The Decisions API uses GPT-6 Luna to pick from answers a developer defines, using text or image context, to classify, route or choose an agent's next step. It's in limited preview, and pricing hasn't been announced. In OpenAI's demo of 10,000 routing requests, each took 150 ms versus 1.6 s. vendor Watch the demo on X.
Builders quickly called it a “Jev killer” (@theo, @birdabo, @mreflow), after TypeSafe's tool launched 15 September. Its founder later posted that it “seems to be luna with constrained decoding.” community
08Where it falls short
- Paywalls: dots need a paid plan while Muse is free, and Ultrafast needs Pro 500 or 6× API pricing. press vendor
- Availability: dots aren't offered in the EEA, Switzerland or the UK, per Fortune. press Specialist dots are limited to enterprise pilots. vendor
- Not the smartest: Sol trails Sonnet 5.5 and Opus 5.5 at max effort on Artificial Analysis. independent
- Churn: GPT-6 Sol was replaced after one week. vendor
- Mood: @LyraInTheFlesh called the day “a lot of let down.” community
09How it stacks up
| OpenAI launch | Rival | How they compare |
|---|---|---|
| Dots | Meta Muse, xAI Grok Bot | Same always-on idea. Dots bring Astra, 4,000+ apps and Slack/Teams, but Muse is free. |
| GPT-6.1 Sol | Claude Sonnet 5.5 | Same $2 / $10 list price. Sol is cheaper per task; Sonnet scores higher at max (56 vs 52). |
| Decisions API | TypeSafe Jev | Same idea; OpenAI's adds image inputs. Its price isn't announced. |
10Handle with care
- The Pro 200 figures (20× → 10×, the $2,500 credit, the 30 October date) are press-reported, not on OpenAI's page.
- The claim that OpenAI cut the stream's audio to hide the dots demo hiccup is unverified.
- “Sol beats Opus 5.5” holds only on some of OpenAI's charts and one creator's run; on Artificial Analysis it scores lower.
- GPT-6.1 Astra's cancellation is reported by the WSJ, not announced by OpenAI.
- Artificial Analysis's Sol numbers were only hours old on 29 September, and its Sonnet 5.5 scores came from a pre-release build with a bug; both may change.
11Sources
- vendor OpenAI: DevDay 2026 recap · Introducing dots · Introducing GPT-6.1 Sol
- vendor OpenAI API pricing · About ChatGPT Pro tiers
- vendor @OpenAI · @OpenAIDevs · @sama · @thsottiaux
- independent Artificial Analysis: GPT-6.1 Sol · Claude Sonnet 5.5
- press Engadget · The Next Web · Fortune · Gizmodo (WSJ)
- community @theo · @jetbrains · @DeryaTR_ · @mreflow · @ns123abc · @birdabo · @CompleteSkeptic · @LyraInTheFlesh