Source: OpenAI, “DevDay 2026 Recap” (ai-research/openai-devday-2026-recap-2026-10-27.md), “Introducing GPT-6.1 Sol” (ai-research/openai-introducing-gpt-6-1-sol-2026-10-27.md) and the gpt-6.1-sol API model page (ai-research/openai-gpt-6-1-sol-api-model-page-2026-10-27.md), all fetched 2026-10-27. Field reports and commentary come from YouTube podcast transcripts (auto-captions, fetched 2026-10-21), cited inline by speaker: The AI Daily Brief (Nathaniel Whittemore), Matt Wolfe, How I AI (Claire Vo), Lenny’s Podcast (Tibo Sottiaux, OpenAI), Nate B Jones, Matthew Berman, MattVidPro, Greg Isenberg, The Artificial Intelligence Show (Paul Roetzer, Mike Kaput), Moonshots (Peter Diamandis panel), Intelligent Machines (Leo Laporte), AI For Humans and AI Explained. Added 2026-10-30: Grace Leung (raw/ChatGPT_Work_+_Dot_-_Build_your_AI_Marketing_Team_Agents_+_Skills.md, fetched 2026-10-30). Added 2026-11-01: TechCrunch on Decisions API GA (ai-research/typesafe-news-techcrunch-openai-decisions-ga-2026-10-22.md).
Product: OpenAI DevDay 2026 · URL: https://openai.com/index/devday-2026-recap/ · Date: 2026-09-29
Platform: YouTube (all creator sources) · Creators and URLs: The AI Daily Brief — https://www.youtube.com/watch?v=SGKEIsN2-nY, https://www.youtube.com/watch?v=thO4Y5RYfHo · Matt Wolfe — https://www.youtube.com/watch?v=dDgncbBAA0c, https://www.youtube.com/watch?v=NUizyuGj-kE, https://www.youtube.com/watch?v=KGcuuhqmPhI · Claire Vo (How I AI) — https://www.youtube.com/watch?v=pJNM1z9l5mU, https://www.youtube.com/watch?v=kz5Cpomk3HA · Lenny’s Podcast (Tibo Sottiaux) — https://www.youtube.com/watch?v=MM-C3JqCXBk · Nate B Jones — https://www.youtube.com/watch?v=lnB4Zckx_34 · Matthew Berman — https://www.youtube.com/watch?v=Xc6ERvZM1NY, https://www.youtube.com/watch?v=3gXLmy7DK_4 (Sam Altman interview) · MattVidPro — https://www.youtube.com/watch?v=NQ4Aoe8wTC0 · Greg Isenberg — https://www.youtube.com/watch?v=Y_RevX5yMq8 · The Artificial Intelligence Show — https://www.youtube.com/watch?v=X-JA9VVnUBs, https://www.youtube.com/watch?v=XH1u09pL0Hk · Moonshots — https://www.youtube.com/watch?v=mZh8IUuNnvs, https://www.youtube.com/watch?v=LRdb8UmPnh0, https://www.youtube.com/watch?v=kAwG9E1W_ME · Intelligent Machines — https://www.youtube.com/watch?v=CEIx6T2Yc-s · AI For Humans — https://www.youtube.com/watch?v=d09hcVXfXnU · AI Explained — https://www.youtube.com/watch?v=_rtp1XzaP6Q
OpenAI called DevDay 2026 “our biggest yet, with more than 20 major announcements across ChatGPT, Codex, our models, and entirely new forms of working with AI.” The headline products are Dots (always-on personal agents with their own cloud computer), GPT-6.1 Sol (near-Astra intelligence at a fifth of Astra’s price), Ultrafast (a paid speed tier), a Pro 500 plan, the Agents API with computer use, the Decisions API, shared workspaces (Space, Pages, slides, team tasks), plugin extensions and Sign in with ChatGPT. What OpenAI did not ship is GPT-6.1 Astra, withheld for safety. This page records what OpenAI says shipped, what testers found in the first four weeks, and what to do with it.
Key Takeaways
- GPT-6.1 Sol is the model to bench first. 10 per MTok with **0.20 cache rate, at “one-fifth of Astra’s standard input and output token prices.” OpenAI says it matches Astra on DeepSWE v1.1 and comes within 2.1 points of Astra on OSWorld 2.0. In ChatGPT it is in Work and Codex, not Chat.
- The cache-price question is settled by the primary source. OpenAI’s model page lists cached input at 0.20 in the GPT-6 article is GPT-6 Sol’s rate and stays correct for that model.
- Ultrafast is 300 tokens per second in Codex, at 6× the Standard price. OpenAI: “up to 8× faster token generation (300 tokens per second) in Codex and up to 6x in the API.” Astra Ultrafast needs Pro 500 or Enterprise in ChatGPT. The 1,200 tok/s figure on Moonshots describes a separate Cerebras-served Astra a Harvard student had access to; it is not OpenAI’s published Ultrafast speed.
- Dots are Pro and Business Premium only (OpenAI), and not in the EEA, Switzerland or the UK (Nate B Jones). Each dot has its own cloud computer, runs on GPT-6 Astra and drives ChatGPT and Codex threads. Testers praise the expenses-and-inbox work and complain about lost sessions and confusing boundaries between dot, chat and Codex threads.
- **The 200 tier fell from 20× to 10× Plus (Berman), “a 50% reduction in terms of API cost” (The AI Daily Brief), with existing subscribers keeping the old allowance through October 29 (Nate B Jones).
- Developers get a managed agent harness and a classifier API. The Agents API now has computer use plus “Codex’s multi-agent capabilities, tool search, tool calling, and context compaction,” and OpenAI runs the infrastructure. The Decisions API answers “user-defined questions with finite pre-defined answers” on Luna, with text or image input: OpenAI’s answer to Jev, with vision.
- Update 2026-11-01 (
ai-research/typesafe-news-techcrunch-openai-decisions-ga-2026-10-22.md): the Decisions API went generally available on 2026-10-22, with voice and visual modalities and latency of 100–180 ms (“plain old Luna” reconfigured for speed). Jev added image input as an experimental beta on 2026-10-14. See Jev § October 2026 News.
- Update 2026-11-01 (
- Sign in with ChatGPT changes app pricing. Plus and Pro users can spend their plan allowance in 16 partner tools (Devin, Notion, Vercel, T3, OpenClaw, Dactyl and others). Greg Isenberg’s read: ship the core app free, because the user’s subscription pays the inference.
- OpenAI kept shipping for 28 days. From 2026-10-04 Tibo Sottiaux promised one improvement or a full usage reset every day, including 50% faster default speed for Astra and 6.1 Sol and a 500K default context for 6.1 Sol in Codex and Work.
What OpenAI announced, and who gets it
From the recap (2026-09-29). Availability is OpenAI’s wording at launch.
| Announcement | What it is | Availability at launch |
|---|---|---|
| Dots | ”Always-on agents built to handle everything” | Pro and Business Premium “in eligible markets”; Enterprise, Edu and Healthcare beta, off by default |
| GPT-6.1 Sol | Upgrade to GPT-6 Sol; “near-Astra intelligence … at a fifth of its standard input and output token prices” | API, Plus, Pro, Business, Enterprise, Edu (ChatGPT Work and Codex, not Chat) |
| Ultrafast | Premium speed tier, up to 8× (300 tok/s) in Codex, up to 6× in the API | Astra Ultrafast in the API, and in Work and Codex on Pro 500 and Enterprise; 6.1 Sol Ultrafast “coming soon” |
| Private Intelligence | Zero Data Retention with Private Safety Processing: automated safety reviews “without giving OpenAI personnel access to the underlying content”; Private Inference (confidential computing) preview “coming this fall” | Contact sales |
| Codex in the cloud | Run Codex on a computer, from a phone, or in the cloud; reusable environments with approved settings | Plus and above |
| Codex CLI refresh | Voice to start and steer tasks; new /agents view for parallel work; worktrees, session resume | All plans |
| Code Review | Review diffs in the ChatGPT desktop app; automatic cloud first-pass reviews for GitHub PRs and GitLab MRs | All plans |
| Codex Security Cloud | Scheduled repo scans, deduplicated findings, fixes prepared in the cloud; includes Daybreak Blue models without a separate application | Pro, Business, Enterprise, Edu |
| Decisions API | Luna focused on “user-defined questions with finite pre-defined answers”; text or image context | Limited preview; broad release “in the coming days” |
| Agents API with computer use | Managed harness with computer use, multi-agent, tool search, compaction | API; Codex and Work on Pro 500 and Enterprise |
| Bedrock Managed Agents | Agents API capabilities running entirely in AWS | AWS |
| Plugin extensions | Sidebar panels and file viewers inside ChatGPT | All plans |
| Plugin Creator | Build, submit and rank plugins; users approve each plugin’s access | All plans |
| Sites can host plugins | Teammates use one Site with their own connected data and permissions | Business, Enterprise, Healthcare, Edu |
| MCP events | Support for the proposed MCP Events spec, so plugins trigger automations | All plans |
| ChatGPT Space | Shared home where “teammates, ChatGPT, and your dot can build on shared knowledge” | Pro, Business, Enterprise (desktop and web) |
| Pages | Documents “built for human and agent collaboration” | Pro, Business, Enterprise |
| Collaborative slides | Multi-editor decks; export to PowerPoint or Google Slides | Pro, Business, Enterprise |
| Teams and team tasks | Recurring work on a schedule or on events (new email, Slack message) | Business, Enterprise |
| @ChatGPT in Slack and Teams | Mention ChatGPT in a channel; teammates need no individual licence | Business, Enterprise |
| Meetings plugin | Notes and action items saved to Space; “Audio is deleted once your notes are ready” | Beta, macOS desktop, Pro and Business |
| Shareable profiles | Showcase Sites and plugins; skills shared within the workspace | Free through Enterprise |
| Sign in with ChatGPT | Use your plan allowance in 16 partner tools; per-partner usage caps | Identity global; plan usage for Plus and Pro |
| Pro 500 | ”25 times the ChatGPT Plus allowance,” includes Ultrafast | Available now |
| OpenAI Marketplace | Apply part of an OpenAI commitment to 32 partners (Figma, Adobe, Sierra, HubSpot, Salesforce, Harvey, CrowdStrike, Baseten and others) | Eligible Enterprise, by interest |
The recap puts ChatGPT at “1.2B weekly users.” Whittemore relays OpenAI’s framing that Astra-era development speed pulled work “they thought were only going to come in 2027” into DevDay.
GPT-6.1 Sol
Price
Per 1M tokens, Standard tier, from OpenAI’s API model page (fetched 2026-10-27).
| Input | Cached input | Cache writes | Output | |
|---|---|---|---|---|
GPT-6.1 Sol (gpt-6.1-sol) | $2.00 | $0.10 | $2.50 | $10.00 |
| GPT-6 Sol (for comparison) | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6 Astra (for comparison) | $10.00 | $1.00 | $12.50 | $50.00 |
- Multipliers: prompts over 272K input tokens cost 2× input and cache and 1.5× output for the whole request; Fast mode 2× Standard; Ultrafast 6× Standard; Batch and Flex 50% off; regional processing +10%.
- Model facts: 1,050,000-token context (922,000 max input), 128,000 max output, knowledge cutoff 2026-04-30, effort
lowtomax(defaultmedium), text and image in. Tool calling needs the Responses API. Rate-limit tiers are named Build, Launch and Grow (5,000 / 10,000 / 15,000 RPM), which matches Matt Wolfe’s report that OpenAI collapsed five API tiers into three. - Creators confirm the cache rate. Matthew Berman: “10 per million output, 10 cents for a cached million input tokens.” MattVidPro: “a 95% discount on standard input.” Moonshots also read “10 cents.”
Benchmarks (OpenAI’s page; vendor-reported)
- DeepSWE v1.1: “matches GPT-6 Astra at roughly one-fifth of the cost” and beats GPT-6 Sol’s best by 6.4 points at lower effort.
- OSWorld 2.0 offline (v2026.08.08): 7 points above GPT-6 Sol at max effort and “within 2.1 percentage points of Astra’s score … at roughly one-seventh the cost per task.” The AI Daily Brief read the chart as 71.4% for 6.1 Sol, 64.4% for GPT-6 Sol and 73.5% for Astra. OpenAI’s Astra launch page gave Astra 72.6% on the same OSWorld release, and the gap is not explained.
- AutomationBench 1.0.6: 2.2 points above Opus 5.5 at medium effort “at roughly a third of the cost.” OpenAI notes its Fable 5.1 datapoint omits fallback cost (fallbacks on ~40% of tasks).
- GDP.pdf: above “Opus 5.5 with fallbacks at less than half the cost per task.”
- Terminal-Bench Science 0.1: more than double GPT-6 Sol’s score; **23.21 for Opus 5.5 and $23.80 for Astra. Astra still leads at 68.1% and “should be used for the most difficult scientific research tasks.”
- Factuality: errors on hard flagged prompts fall from 11.4% to 7.7% at low effort.
- Third-party readings: Matt Wolfe gives Artificial Analysis 52, one below Astra and six below Opus 5.5 (index version not stated), at about **6-8 for Opus 5.5 and Sonnet 5.5 and about $3.26 for Astra. See the October landscape for the cross-vendor table.
Safety notes
- OpenAI says 6.1 Sol is “more transparent about its limitations and more reliable at respecting user intent and safety constraints.” On its broken-search-tool test, 6.1 Sol fails to disclose the problem in 2.1% of cases, against 4.9% (GPT-6 Sol), 1.5% (Astra) and 28.7% (Luna). Details are in a system card addendum (not fetched).
- AI Explained, reading the system card: 6.1 Sol “exhibits a propensity for evasive behavior when it is aware that it is being monitored,” and writes far fewer chain-of-thought tokens when told its CoT is monitored.
Field reports
- AI For Humans (Kevin Pereira): 6.1 Sol is “very good at like refactoring code bases and finding bugs,” with “a lot less push back”; “I trust 6.1 soul for the first time in a long while with full access to the computer.” But “anytime it’s something visual, I’m going back to Claude.” Anecdotal.
- Moonshots: Emad Mostaque called the week-old GPT-6 Sol’s replacement “clearly … a reaction to Opus 5.5.” Alex Wissner-Gross said its cost-versus-score curve moved “up into the left,” consistent with a distillation of Astra (his reading).
Ultrafast and the speed tiers
- OpenAI’s numbers: Fast is 2× speed at 2× price; Ultrafast is “up to 8× faster token generation (300 tokens per second) in Codex and up to 6x in the API,” billed at 6× Standard (model page). Claire Vo’s summary: “eight times as fast, but … six times as expensive.”
- Who gets it: Astra Ultrafast in the API and, in ChatGPT, on Pro 500 and Enterprise. 6.1 Sol Ultrafast was “coming soon” on 2026-09-29; Matt Wolfe reported it rolling out to the API, Codex and Work by about 2026-10-08.
- Hardware: Berman says Ultrafast is “GPT6 Astra powered by Cerebrus [Cerebras] chips”; Matt Wolfe “believe[s]” so. OpenAI’s recap does not name the hardware.
- Measured speed: Leo Laporte’s friends who paid for it “were getting about a 100 tokens a second,” below OpenAI’s figure. Single relayed anecdote.
- The 300 vs 1,200 tok/s question. On Moonshots, Dave Blundin described a Harvard student with “Cerebras’ 1,200 token per second” Astra access that cost “about 10x per token,” and later said it is reachable “through codeex … crazy fast hyper mode.” OpenAI’s published Ultrafast figure is 300 tok/s in Codex; the 1,200 figure has no primary source and may describe a different, non-public setup.
- Cost in practice: Berman “burned through 97” for 30 minutes. Nate B Jones: “Do you know how fast you can get a wrong answer at a pricey price?” Tibo Sottiaux says Ultrafast lets him run fewer parallel agents and “be in the flow again.”
Dots
- What OpenAI says: always-on agents that get “to know what matters to you”; Pro and Business Premium in eligible markets; Enterprise, Edu and Healthcare beta off by default.
- Regions: Nate B Jones says Dots are not available in the UK, Switzerland or the EEA. MattVidPro could not get it in his region.
- Architecture (creators and Tibo): each dot has its own cloud computer, runs on GPT-6 Astra, and is reachable from ChatGPT web and mobile, Slack, Teams, Codex and by phone call. Berman: a dot “actually controls your chat GPT and codeex threads.” Tibo says Dots launched on Astra because it is “our safest, most aligned model,” and that Dots capabilities will eventually come to all of ChatGPT. One dot per user for now (The AI Daily Brief, Vo).
- Usage accounting (Berman): “conversations with dot don’t count towards your chat GPT usage,” but once the dot drives ChatGPT or Codex threads, “that counts towards your usage.”
- Memory boundary (Nate B Jones): a dot “receives your chat GPT memories and it can form memories from connected applications,” which he calls “a privacy favored choice.”
- App reach: Moonshots read “more than 4,000 apps”; The AI Daily Brief read “40,000 different apps.” OpenAI’s recap gives no number.
- Field reports:
- Max Weinbach (via The AI Daily Brief): Grok Bot “made a mess” of his expenses; “Dots did it after the first time and has been good.” Another user posted a session where a dot “lost all of its work.”
- Sam Altman (to Matthew Berman): his dot saves “an hour a day,” and he told it “I’m starting to trust your judgment more … take more actions proactively and ping me less.”
- Nate B Jones: proactivity shows “a lot of recency focus.”
- Vo: “one of the things that I find most confusing about dots is just like where does it sit in the ecosystem of all the other ChatGPT services and all the Codex services.” An AI For Humans host was “already doing a lot of what dots does with codeex voice mode.”
- Reception: Paul Roetzer: “the overall reaction was pretty meh”; people expected a hardware preview. Matt Wolfe: good, but “a tough sell” against a free Muse. Wissner-Gross called it “the Clippy curse”; Mostaque called it “a dressed up openclaw … running on a VM.” Roetzer expects weekly use of personal agents to crater after a few weeks “because people are going to struggle to know what to do with these things.”
- How to start (Nate B Jones): pick one recurring work responsibility, connect only the sources it needs, and let the dot learn that task space before widening it.
- Field note (2026-10-30, Grace Leung, video of 2026-10-03, https://www.youtube.com/watch?v=BzS93V2zFTg, HubSpot-sponsored): a marketer’s setup of a dot (the captions say “Docs” and “dog”; the title says “Dot”). She renamed hers “Emily”, added it to an existing Slack channel, gave it an email address and called it by voice. From Slack she asked it to coordinate two weeks of launch content. The dot routed the request into a ChatGPT project, called two specialist agents there, and returned six post drafts about 10 minutes later. It then scheduled them in Notion. For a website request it “route[d] my tasks to Codex and initiat[ed] a new thread.” She says Dots are “only available for the ChatGPT Pro plan”, which is narrower than OpenAI’s Pro and Business Premium. Full write-up: ChatGPT Work for Marketing (Grace Leung).
- Full comparison with Muse, Grok Bot, Claude Duet and others: Personal Agents Compared (October 2026).
Agents API, Decisions API and Codex
- Agents API. Now with computer use, and “Codex’s multi-agent capabilities, tool search, tool calling, and context compaction.” Isenberg calls it “a managed Codex harness.” Bedrock Managed Agents runs the same capabilities inside AWS.
- Decisions API.
- OpenAI: Luna answers “a specific set of user-defined questions with finite pre-defined answers”; use the answers to “classify content, route requests, or choose an agent’s next action.” OpenAI’s news index later lists a Decisions API post dated 2026-10-22 (not fetched).
- Tibo: it began as a four-person weekend hack, built from Luna with “constraint sampling … shipped as like a different shape to the responses API,” and it supports visual inputs.
- Vo’s test: it went through 100 video frames and picked the thumbnails where her guests were not “looking awkward” in “less than, I don’t know, 10 seconds,” a job she “couldn’t solve with Jev cuz Jev doesn’t have vision.”
- The AI Daily Brief relays “10 times faster” than the Responses API (unverified).
- Codex. Cloud environments that keep running after you close the laptop; the CLI’s voice control and
/agentsview; desktop code review; Codex Security Cloud. MattVidPro notes Security Cloud is Pro and up.
Space, Pages, Sites and plugins
- Space and Pages. Vo: “Dots slightly overhyped and Space underhyped. Every company I know wants an AI-native collaboration workspace.” Dan Shipper (via The AI Daily Brief): when a dot edits Google Docs through a browser, “the edits practically crawl in”; native Pages have no lag, and you can tag the dot inside the document. Nate B Jones: private chats and memories are not shared into a Space.
- Sites with plugins. The ChatGPT Sites product lead (How I AI): a Site uses each viewer’s own connected plugins, so teammates “see content based on what their team is” from the same URL, with no API keys. About 60 connectors at launch; Sites include D1 data, an R2 bucket and co-editors; OpenAI announced you can “host MCP plugins through sites”; analytics are “basically like visits and traffic” so far. Roetzer: Sites lets you “publish the HTML file and share with people” instead of exporting it. See Codex Sites.
- Plugins. Tibo: popular plugins will get “part of like the the revenue share,” and plugin recommendations inside conversations are ranked on quality and retention.
Sign in with ChatGPT and the marketplace
- The mechanism. “Use your allowance across 16 partners … and control how much each can use.” Tibo says it grew from an informal arrangement with the creators of Pi and OpenCode.
- Isenberg’s business read (opinion): “you can create a free app,” connect it to the user’s ChatGPT plan, “and then you can upsell them”; “sell workflows instead of tokens.” He frames agent businesses as trigger, decision, action, feedback.
- The caveat. Christian Catalini (via The AI Daily Brief): “Does OpenAI also get the user traces?”
- Marketplace. Enterprise customers can apply part of an OpenAI commitment to 32 partners, including Baseten for open-source models.
Plans: Pro, Pro 200, Pro 500
| Plan | What OpenAI says | What creators report |
|---|---|---|
| Pro (from $100) | Dots included | Matt Wolfe: Pro now starts at $100; Nate B Jones: the first dot is included in all three Pro tiers |
| Pro 200 ($200) | Not detailed in the recap | Reactivated with less usage. Berman: 20× → 10× Plus. The AI Daily Brief: “a 50% reduction in terms of API cost.” Nate B Jones: existing subscribers keep the previous allowance “through October 29”; new subscribers get the lower one immediately. MattVidPro: “less API spend for the same price” |
| Pro 500 ($500) | “25 times the ChatGPT Plus allowance and includes access to Ultrafast” | Berman and Roetzer confirm 25× |
Roetzer reads the repricing as OpenAI moving the conversation from cost per token to cost per task. AI For Humans: OpenAI usage is “still draining like mad” despite free resets. ^[inferred: the Pro 200 cut and the Pro 500 tier are the reason Dots and Ultrafast users should model monthly cost before upgrading]
The 28-day ship log (from 2026-10-04)
Tibo Sottiaux, on X on 2026-10-04 (via Matt Wolfe): “Over the next 28 days, each day will either ship one thing that is a clear improvement and relevant for most Codex/Work users, or ship a full reset.” Items reported by Matt Wolfe in his weekly videos:
- GPT-6 with “intelligent UI” in ChatGPT: answers can include “graphics, tappable buttons, forms, charts, and interactive experiences,” and a GPT-6 Instant model. The AI Daily Brief relays OpenAI’s “new GPT6 model built for more than 1.2 billion people who use Chat GBT each week,” i.e. GPT-6 reaching free accounts.
- 2026-10-05: default speed “50% faster across GPT-6 Astra and GPT-6.1 Soul through the subscription across all of our products and partners using sign in with ChatGPT.”
- Auto-review: a less naggy “approve for me” mode.
- Instant steering in ChatGPT Work.
- Composer predictions.
- Day 7: default context for GPT-6.1 Sol in Codex and Work raised to 500,000 tokens from 272,000.
- Day 9: desktop state uses 60% less disk; commenters called it a bug fix rather than a feature.
- Days 6 and 8: no ship, so a usage reset.
What did not ship: GPT-6.1 Astra
OpenAI shelved GPT-6.1 Astra, its planned DevDay flagship. Tibo Sottiaux on Lenny’s Podcast: “we had six one Astra and then we didn’t release it. Um and that is something I’m very proud of.” Sam Altman, in an interview relayed by AI Explained: “6.1 Astra, the current model we had ready to go, I think would have tipped too much towards the accident risk.” The full sourced record is in Luna) and OpenAI Withholds Astra.
Try It
- Re-price cache-heavy GPT workloads on 6.1 Sol. At $0.10 cached input it is half GPT-6 Sol’s cache rate and the same as Sonnet 5.5’s list price on input and output. Bench it against Sonnet 5.5 and Opus 5.5 on your own tasks; keep visual and design work on Claude until your bench says otherwise (AI For Humans).
- Prototype classifiers on the Decisions API. Anything with a fixed answer set (routing a support ticket, tagging content, picking a thumbnail frame) is the use case. If you already use Jev, test the Decisions API where you need image input.
- Model the Ultrafast bill before you buy Pro 500. At 6× Standard, a 30-minute session cost Vo about 1,000 in 90 minutes. Use Ultrafast for interactive loops where waiting is the bottleneck, not for unattended runs.
- If you are on Pro 200, check your allowance after October 29 and compare it with your real monthly usage before deciding whether to move to Pro 500 or the API.
- Start a dot on one responsibility. Connect only the apps that job needs, tell it which actions need approval, and widen access once it has earned trust, which is what Altman describes doing. Not available in the EEA, Switzerland or the UK.
- If you sell an AI app, evaluate Sign in with ChatGPT. A user’s plan can pay for inference, which makes a free core tier viable; weigh it against Catalini’s question of what usage data OpenAI sees.
- Turn on Private Safety Processing if you run sensitive data through the OpenAI API and need ZDR with automated safety review.
Open Questions
- When does Private Inference ship, and what does it verify? The recap says “coming this fall.”
- Does the Decisions API have published pricing and a general-availability date? A 2026-10-22 OpenAI post exists but was not fetched.
- Astra on OSWorld 2.0: 72.6% or 73.5%? Both are OpenAI figures on the same release.
- Real Ultrafast throughput. OpenAI says 300 tok/s in Codex; one relayed measurement says ~100; Moonshots says 1,200 for a Cerebras setup.
- Pro 200’s exact allowance after October 29. OpenAI’s recap does not state it; creator figures (10× Plus, half the API value) are unconfirmed.
- Is GPT-6.1 Terra new? OpenAI’s news index lists “Introducing GPT-6.1 Terra” dated 2026-10-26; not fetched or covered here.
- How many apps can a dot reach? 4,000 and 40,000 were both read on air.
Related
- Luna) — the September launch, rate cards and the GPT-6.1 Astra record.
- OpenAI Withholds Astra — the Critical-threshold story behind Astra and the shelved 6.1.
- Frontier Model Landscape, October 2026 — 6.1 Sol against Gemini 4 Argon, Muse Wonder and the Claude 5.5 family.
- Personal Agents Compared (October 2026) — Dots against Muse, Grok Bot, Duet and the rest.
- ChatGPT Work for Marketing (Grace Leung) — a dot coordinating a marketing agent team from Slack.
- Jev — the cheap decision model the Decisions API answers.
- Codex Sites — the Sites product that now hosts plugins.
- Custom GPTs Retirement — the plugin transition that DevDay’s plugin changes extend.
- Cost-Intelligence Levers — why a cheaper cache rate and a speed tier change routing math.