Author: editorial staff airassvet.ru. This review was prepared on September 29, 2026, the day of the conference, using primary OpenAI sources — the API changelog, the official announcement thread on the developer forum and the documentation on developers.openai.com — and cross-checked against reports from industry publications linked in the text. We did not test the new products ourselves: all test results and prices are taken from OpenAI and the cited sources. The editorial team has no commercial relationship with OpenAI or the publications mentioned, and there are no affiliate links in the text.
Short answer. OpenAI DevDay 2026 took place on September 29, 2026, in San Francisco; the company introduced more than 20 products and features. The main ones: the GPT-6.1 Sol model, which comes close to the flagship GPT-6 Astra for coding and computer use at one-fifth the price ($2 and $10 per million input and output tokens); Dots — always-on agents inside ChatGPT with their own cloud computer; the shared workspace ChatGPT Space; the high-speed mode Ultrafast (up to 8x faster for six times the price); computer use in the Agents API and the new Decisions API; cloud Codex and the ChatGPT Pro 500 plan for $500 per month. Important note for companies in Russia: Russia is not included in the list of countries where OpenAI officially provides its services.
Below are all the announcements in order: what was shown, how much it costs, who it’s available to and when, where the sources differ, and what is worth trying for business use. This overview is intended for executives, CTOs, and product teams building processes around language models and AI agents. We do not cover the legal, tax, or sanctions aspects of working with OpenAI — for those, you need a lawyer’s advice.
Contents
- What OpenAI DevDay 2026 is
- All DevDay 2026 announcements in one table
- GPT-6.1 Sol: almost Astra for one-fifth the price
- Ultrafast: speed at six times the price
- Dots: persistent agents in ChatGPT
- ChatGPT Space, Pages, and team collaboration
- Agents API and Decisions API
- Codex: cloud, voice, review, and security
- Plugins, Sign in with ChatGPT, and Marketplace
- Private Intelligence: ZDR with Private Safety Processing
- New ChatGPT plans: Pro 500 and stripped-down Pro 200
- What was left out: risks and caveats
- What all this means for business in Russia
- What to choose for your use case: quick reference
- How we prepared this review and its limitations
- Frequently asked questions
- How the airassvet.ru team helps implement agents inspired by DevDay 2026
- Conclusion
What OpenAI DevDay 2026 is
OpenAI DevDay is OpenAI’s annual conference for developers. In 2026, it took place on September 29 at Fort Mason Center in San Francisco; Sam Altman’s keynote began at 10:00 a.m. Pacific Time, or 8:00 p.m. Moscow time, and was streamed free for everyone (event page, OpenAI announcement). According to Altman, as reported by Engadget, ChatGPT now has 1.2 billion weekly users.
OpenAI DevDay 2026 is the conference where OpenAI announced more than 20 launches for developers and business: the new GPT-6.1 Sol model, Dots agents, the ChatGPT Space workspace, the Ultrafast speed mode, an expanded Agents API, cloud Codex, and new ChatGPT plans. CNBC summarized the overall message of the conference this way: CNBC: OpenAI is moving AI beyond the chatbot — toward software that works independently on the user’s behalf.
To understand the announcements, it’s important to remember that they were preceded by a very busy September. According to the API changelog and our earlier analyses:
| Date | Event |
|---|---|
| September 3, 2026 | Launch rollout begins for the flagship GPT-6 Astra (our analysis) |
| September 10, 2026 | Agents API enters public beta |
| September 22, 2026 | GPT-6 Sol and GPT-6 Luna are released (our analysis) |
| September 25, 2026 | A bug in image encoding in Sol and Luna is fixed; it had degraded visual tasks, including computer use |
| September 28, 2026 | OpenAI canceled the planned release of GPT-6.1 Astra because the model did not meet its safety standards (Decrypt) |
| September 29, 2026 | DevDay 2026: GPT-6.1 Sol, Dots, Ultrafast, computer use in Agents API, and other announcements |
The first practical takeaway is this: GPT-6.1 Sol arrived just one week after GPT-6 Sol. If you just finished testing GPT-6 Sol, you’ll need to recheck your results against the new version.
All DevDay 2026 announcements in one table
The summary table was compiled from the OpenAI announcements thread and the API changelog. Where the sources differ, that’s noted in the relevant section below.
| Announcement | What it is | Status as of 09/29/2026 | For whom |
|---|---|---|---|
| GPT-6.1 Sol | A model for coding, computer use, and professional tasks | Available in the API, ChatGPT Work, and Codex | Developers, paid ChatGPT plans |
| Ultrafast | Speed mode: up to 8x faster in Codex and up to 6x faster in the API | For GPT-6 Astra — now, for Sol — soon | API, Pro 500, Enterprise |
| Dots | Persistent agents in ChatGPT with their own cloud computer | Gradual rollout | Pro and Business Premium |
| Specialist Dots | Agents with assigned responsibilities within an organization | Preview for enterprise customers | Enterprise |
| ChatGPT Space | A shared workspace for teams and agents | Launched | Teams |
| Pages | Editable documents for people and agents | Launched | Teams |
| Collaborative slides | Presentations in Space | "In the coming weeks" | Teams |
| Teams and Team Tasks | Shared work and recurring tasks | Launched | Teams |
| ChatGPT in Slack and Microsoft Teams | Replies to @ChatGPT mentions | Launched | Teams |
| Meetings plugin | Meeting notes and tasks | Beta on macOS | Pro and Business |
| Agents API + computer use | Agent hosting, memory, tools, multi-agent support, and workflow through software interfaces | Public beta | Developers |
| Decisions API | Fast classification and routing to Luna models | Limited preview | Developers |
| Codex in the cloud | Tasks keep running with the laptop closed, managed from any device | Launched | Paid plans |
| Codex CLI | Voice control, view /agents, improved worktree |
Launched | Developers |
| Code Review | Summaries, diffs, questions about changes, automated cloud reviews | Launched | Developers |
| Codex Security Cloud | Scheduled vulnerability scans across repositories, fix preparation | Launched | Developers, security |
| Plugin extensions | Side panels, interactive panels, file preview | Launched | Plugin developers |
| MCP Events | Connected apps trigger automations | Launched | Developers |
| Sign in with ChatGPT | Sign in through ChatGPT and pay for requests from the user’s subscription | 16 partners at launch | Product developers |
| OpenAI Marketplace | Purchasing partner products against a corporate contract | Launched | Enterprise customers |
| Amazon Bedrock Managed Agents | Agents on OpenAI inside AWS | Launched | AWS customers |
| Private Intelligence | ZDR with Private Safety Processing; Private Inference preview | ZDR+PSP — now, Private Inference — in the fall | Enterprise customers |
| Pro 500 and Pro 200 changes | New $500 tier, reduced Pro 200 limits starting October 30 | Launched | ChatGPT subscribers |
| Platform speed improvements | Time to first token in the API down 45%, tool calls more than 30% faster | Announced | All API customers |
Next, let’s break down each block in detail.
GPT-6.1 Sol: almost Astra for one-fifth the price
GPT-6.1 Sol is OpenAI’s updated mid-tier model for advanced coding, computer use, and professional tasks. According to OpenAI, it nearly matches GPT-6 Astra while standard token pricing is five times lower. In the API, it is called gpt-6.1-sol and has been available since September 29, 2026 (API changelog).
GPT-6.1 Sol pricing
According to the API changelog and the comparison table from Simon Willison’s livestream:
| Price per 1M tokens | GPT-6.1 Sol | GPT-6 Astra | How many times cheaper Sol is |
|---|---|---|---|
| Input | $2.00 | $10.00 | 5x |
| Cached input | $0.10 | $1.00 | 10x |
| Cache write | $2.50 | no data in our sources | — |
| Output | $10.00 | $50.00 | 5x |
GPT-6.1 Sol’s base input and output prices are the same as GPT-6 Sol. The cached input changed: it costs $0.10 versus $0.20 for GPT-6 Sol, meaning it is twice as cheap and 95% cheaper than regular input (TNW). For agents, that is the key number: at each step, the agent resends the model a long context — instructions, history, tool results — and most of it repeats.
Test results based on OpenAI data
OpenAI and TNW report the following results:
- DeepSWE v1.1 (complex software engineering tasks in real codebases): GPT-6.1 Sol matched GPT-6 Astra at roughly one-fifth the cost and beat GPT-6 Sol’s best result by 6.4 percentage points. According to Vellum, that is 75.2% at high reasoning. The figure is consistent with OpenAI’s September 22 data: 68.8% for GPT-6 Sol plus 6.4 percentage points equals 75.2%.
- AutomationBench (business processes): 2.2 points higher than Anthropic’s Claude Opus 5.5 at about one-third the cost.
- OSWorld 2.0 (computer use through the interface): within 2.1 points of Astra at about one-seventh the cost per task.
- Terminal-Bench Science 0.1: an average of $5.47 per task versus $23.21 for Claude Opus 5.5.
These are the vendor’s results on its own benchmarks and settings. They show the direction — a cheaper model has moved close to the flagship — but not how often your agent will succeed. You need to test it on your own tasks.
Where GPT-6.1 Sol is available
According to TNW:
- in API — under the name
gpt-6.1-sol, with multi-agent mode support in beta via the Responses API; - in ChatGPT Work and Codex — for Plus, Pro, Business, Enterprise, and Edu plans;
- in the main ChatGPT chat — not yet;
- in Ultrafast mode — later; OpenAI promises up to 8x faster generation in Codex.
On the OpenAI forum, Sol is marked as “coming soon,” but the API changelog records the release of gpt-6.1-sol on September 29. We consider the changelog a more accurate source; “soon” likely refers to the Ultrafast mode for Sol.
What an agent task costs: the calculation
Let’s take a hypothetical agent task: across all steps, it sent the model 200K input tokens, of which 150K were repeating cached context, and received 20K output tokens. We are not including the cost of cache writes or tools in the calculation.
| Item | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|
| 50K standard input tokens | $0.100 | $0.500 |
| 150K cached input tokens | $0.015 | $0.150 |
| 20K output tokens | $0.200 | $1.000 |
| Total per task | $0.315 | $1.650 |
| Same in Ultrafast mode (×6) | not available yet | $9.900 |
On tokens, Sol is about 5.2 times cheaper than Astra. But you need to calculate the cost of the accepted result: if Astra solves the task on the first try and Sol on the third, the difference narrows. So the first step is to run 20–50 real tasks from your workflow on both models and compare the share of accepted answers and the total cost. accepted result: if Astra solves the task on the first try and Sol on the third, the difference narrows. So the first step is to run 20–50 real tasks from your workflow on both models and compare the share of accepted answers and the total cost.
Ultrafast: speed at six times the price
Ultrafast is OpenAI’s new service tier that speeds up token generation up to 8x in Codex and up to 6x in the API, but costs 6x more than the standard rate. In Codex, that’s up to 300 tokens per second (The Decoder); for the future version on Sol, BGR cites 350 tokens per second.
What we know from the Ultrafast documentation:
- enabled with the
service_tier: "ultrafast"parameter with thegpt-6-astramodel; - works only with global processing and data storage in the U.S.; EU and other regional endpoints are not supported;
- for agents, OpenAI recommends WebSocket so you don’t lose time to network latency between tool calls;
- limits: 500K tokens per minute for tiers 1–3, 1M for tier 4, and 5M for tier 5.
In ChatGPT and Codex, the mode is available on Pro 500 and Enterprise plans.
When Ultrafast makes sense: when a person is waiting for a response in real time — a voice assistant, live customer chat, interactive code editing — or when a long agent chain is constrained by generation speed. When it doesn’t make sense: overnight batch jobs, reports, background agents. There, the 6x price gets you nothing. For budgeting, the simple rule is: Ultrafast does not make tokens cheaper; it reduces wait time, so it’s worth paying for only on steps where waiting costs money.
Dots: persistent agents in ChatGPT
Dots are always-on AI agents inside ChatGPT powered by the GPT-6 Astra model, each with its own cloud computer and browser, access to more than 4,000 apps through plugins, and memory of your preferences. This is the main consumer announcement at DevDay: according to OpenAI, Dots run 24/7, learn from feedback, and can act proactively (9to5Mac, BGR).
What Dots can do
According to BGR, Engadget, Simon Willison and OpenAI announcement threads:
- work in ChatGPT apps and on the web, in Slack and Microsoft Teams, and by voice call; SMS messaging is promised later;
- connect to more than 4,000 apps through plugins;
- with your permission, they gain access to the laptop;
- in their free time, they conduct “proactive research” using read-only tools;
- they communicate by voice through GPT-Live technology;
- Dots for teams will come later.
OpenAI showed a preview for enterprise customers Specialist Dots — agents with specific responsibilities within an organization. They are expected to be managed through Microsoft Agent 365, meaning through Microsoft’s familiar access management tools.
Who can access Dots and how much they cost
Here the sources disagree, and it’s worth knowing before you buy a subscription:
| Source | What is said about access |
|---|---|
| BGR | Pro 200 and Business Premium; not on the $100 Pro plan |
| Engadget | One Dot free for Pro and Business Premium, additional ones for a fee |
| Simon Willison | Pro and Enterprise |
| CNBC | Pro and Business Premium in supported countries |
What all sources agree on: there is no free or Plus access, and the rollout is gradual. According to BGR, the first Dot is included with the subscription and does not use up any limits. The price of additional Dots is not specified in our sources. For comparison, BGR notes that Meta’s competing agent Muse is available on the free tier and on the $20 and $100 plans.
How to control the agent
The main control mechanism is Custom Rules: for each action, you can require approval, block it, or allow it without asking questions (BGR). For business use, this is the minimum place to start: anything that sends emails on behalf of the company, spends money, or changes data in CRM should require approval. For more on how to set levels of agent autonomy, see our article on managing AI agent autonomy.
What went wrong
First impressions were mixed:
- the live voice demo failed on stage, and slow responses were noted (Engadget, Simon Willison);
- you can’t create a Dot from a phone — the app sends you to a computer (Simon Willison);
- Dan Shipper from Every , who tested Dots before launch, says the agent handled email and Slack filtering well and warned about calendar conflicts, but the version was “too rough”: permission issues, missing messages, browser connection failures. His advice is to wait a week or two while OpenAI fixes the minor bugs.
ChatGPT Space, Pages, and team collaboration
ChatGPT Space is a persistent shared workspace where the team, ChatGPT, and Dot agents work together with project files and context. CNBC describes it as a mix of Slack, Notion, and Google Drive with built-in agents.
What’s included in the team collaboration package (OpenAI announcements thread):
- Pages — a new document type that people and agents edit together; supports charts, images, checklists, dashboards, and interactive tools;
- collaborative slides — promised “in the coming weeks”;
- Teams and Team Tasks — shared work and recurring task assignments;
- ChatGPT in Slack and Microsoft Teams — responds to @ChatGPT mentions and uses connected tools directly in the conversation;
- Meetings plugin — meeting notes and a task list; macOS beta for Pro and Business;
- public profiles — a page that brings together your Sites and plugins.
Shipper from Every wrote his review directly in Space and notes that the built-in documents work noticeably faster than editing through a Google Docs agent. He sees the main drawback as the switching cost: most teams already live in Google and Notion. For business, the data question matters too: everything that lands in Space is stored with OpenAI.
Agents API and Decisions API
Agents API: now with computer use
Agents API is OpenAI’s managed service for running long-lived agents: OpenAI handles session orchestration, context compression, and recovery after failures, while the agent runs in an OpenAI sandbox or in the compute environment chosen by the developer. It entered public beta on September 10, 2026 (API changelog); it supports memory, tools, and coordination across multiple agents.
New at DevDay — computer use. According to the changelog, agents now complete tasks in a browser hosted by OpenAI, while access approval for websites and account sign-in remain on your app’s side. Put simply, the agent can work with a program through its interface if it doesn’t have an API. This is a direct competitor to traditional RPA for web systems.
OpenAI also released the Codex harness on GitHub — the same agent-management logic that runs inside Codex. This gives teams more transparency and control over agent behavior (Decrypt).
For Russia and for sensitive data, the key question here is where the agent runs: in an OpenAI sandbox or in your environment. An agent with computer use sees everything visible on the screen, so its access rights need to be limited just as strictly as an employee’s. We wrote about real attacks on agents through tools in our analysis of a Hugging Face hack by an AI agent.
Decisions API: a decision in 150 milliseconds
Decisions API is a service for narrow, repetitive decisions: it uses the GPT-6 Luna model to classify inputs, route requests, or choose an action from a predefined list, and responds in about 150 ms. According to The Decoder, that's roughly 10 times faster than using Luna through the standard API. The response comes with a confidence score; The New Stack says the launch is a response to TypeSafe's Jev product.
Status: limited preview; pricing was not disclosed in our sources. Typical use cases for this kind of API include identifying a support request topic and sending it to the right queue, selecting a response template, or flagging a suspicious transaction for human review. The key limitation is built into the definition: the model chooses from your options rather than generating free text, which reduces the risk of made-up answers but requires you to define all allowed outcomes in advance.
Codex: cloud, voice, review, and security
Codex, OpenAI's coding agent, got the most updates (OpenAI announcements track, The Decoder):
- Codex in the cloud. Tasks keep running even when the laptop is closed; you can launch and monitor them from a phone or another device. On stage, OpenAI showed an app being rewritten in Rust (Engadget).
- Reusable cloud environments. Teams share configurations with approved settings and permissions. Available on Plus, Pro, Business, Enterprise, Edu, and Healthcare plans.
- New CLI. Two-way voice control, a /agents view for monitoring agents, improved prompt editing, session resumption, and Git worktree support.
/agents. - Code Review. In the desktop app, there's now a dedicated review interface: change summaries across multiple projects, diffs, questions about the changes, and integration with GitHub and GitLab. Automated reviews can run in the cloud.
- Codex Security Cloud. Scans entire GitHub repositories on demand or on a schedule, investigates findings, removes duplicates, prepares fixes, and watches for new commits. Users get access to Daybreak Blue models — OpenAI's security-focused layer for cybersecurity — without a separate application.
A shipper from Every puts the shift like this: manually launching separate Codex sessions “already feels archaic.” For an engineering leader, that means the question changes from “how do I make a programmer faster?” to “which tasks can I hand over to an agent end to end, and how do I verify the result?” We covered this transition in our article on the previous Codex update.
Plugins, Sign in with ChatGPT, and the Marketplace
The second big DevDay theme was distribution: OpenAI is turning ChatGPT into a platform where third-party products find users and get paid.
- Plugin extensions — side panels, interactive panels, and file browsing inside ChatGPT and Codex; Figma and Adobe demonstrated integrations.
- Plugin recommendations in conversation. ChatGPT suggests the right plugin on its own as the dialog unfolds (Engadget).
- Simplified moderation. Status tracking for reviews, clearer feedback, and simple updates to already published plugins.
- Sites with plugins. Each user connects their own data and permissions. We wrote about Sites themselves in a separate guide.
- MCP Events. Apps connected through MCP can now trigger automations on their own — for example, a new CRM event can start an agent.
- Sign in with ChatGPT. A user signs into your product with a ChatGPT account and pays for AI requests through their subscription. At launch there are 16 partners, including Notion, Vercel, and Devin (BGR). To connect, you need to request a client ID from OpenAI (documentation).
- OpenAI Marketplace. Enterprise customers spend their OpenAI contract on partners' products. Named companies include Adobe, Canva, Figma, Notion, Salesforce, Vercel, Zendesk (Simon Willison) and Palo Alto Networks (Decrypt); open models are available through Baseten.
- Amazon Bedrock Managed Agents. Agents based on OpenAI models run inside the customer's AWS environment.
Sign in with ChatGPT changes the economics of AI features in SaaS: if the user pays for requests through their own subscription, the developer does not need to build token costs into their pricing. The downside is dependence on the user's OpenAI account and its rules.
Private Intelligence: ZDR with Private Safety Processing
For enterprise customers, OpenAI has combined privacy features under the name Private Intelligence. ZDR with Private Safety Processing is already available: according to the ZDR with Private Safety Processingdocumentation , this is automatic API request safety screening without storing your data at OpenAI.How it works:
records are encrypted and stored in the customer's cloud storage — AWS S3, Azure Blob Storage, or Google Cloud Storage;
- checks are performed in a hardware-attested environment where OpenAI employees do not have access to decrypted data; records are stored for 30 days according to the customer's storage lifecycle rules; double encryption: an internal HPKE layer from OpenAI and an external layer using the customer's keys (Enterprise Key Management); OpenAI's key alone is not enough to decrypt a record;
- OpenAI keeps only an index with operational metadata and a link to the storage, without a copy of the contents;
- records are retained for 30 days under the client’s storage lifecycle rules;
- dual encryption: an internal HPKE layer from OpenAI and an external layer encrypted with the client’s keys (Enterprise Key Management); the OpenAI key alone is not enough to decrypt the record;
- OpenAI retains only an index with operational metadata and a link to the storage, without a copy of the content;
- Organizations already approved for Zero Data Retention configure PSP in the API console themselves.
Preview Private Inference is slated for this fall, with no exact date announced. We covered what Zero Data Retention is and why it is not enough on its own to protect data in an earlier article Zero Data Retention: What It Is and How to Protect Data in AI.
New ChatGPT tiers: Pro 500 and a pared-down Pro 200
According to TNW, BGR and OpenAI announcement threads:
| Tier | Monthly price | Limit relative to Plus | What changed |
|---|---|---|---|
| Pro 100 | $100 | 5× | No changes; Dots, according to BGR, are unavailable |
| Pro 200 | $200 | 20× → 10× starting October 30, 2026 | Reopened to new subscribers after a pause since September 10 |
| Pro 500 | $500 | 25× | New tier with access to Ultrafast |
Details on Pro 200: starting October 30, 2026, the limit in ChatGPT Work and Codex drops from 20 to 10 Plus limits, and GPT-6 Pro messages in chat drop from 200 to 100 per week. Current subscribers keep the old limits through October 29, then receive a one-time grant of $2,500 in usage credits, which expires on December 31. OpenAI also promised not to bring back the five-hour limit that Pro 200 used to have (TNW).
At the end of the conference, OpenAI also announced a global reset of usage limits; our sources do not provide details.
What was left offstage: risks and caveats
DevDay took place against the backdrop of developments worth keeping in mind when planning:
- GPT-6.1 Astra was pulled. The day before the conference, OpenAI scrapped the planned October release of GPT-6.1 Astra because the model did not meet its safety standards (Decrypt). All of the most autonomous DevDay products — Dots, Specialist Dots, Ultrafast — run on GPT-6 Astra, which, as we wrote, already has a Critical cybersecurity status in OpenAI's own classification (our analysis).
- Agent incidents. Decrypt notes that the conference followed a summer of breaches involving AI agents, including an incident with Hugging Face and government systems. An agent with its own computer and access to 4,000 apps is a major attack surface.
- The demos malfunctioned. The Dots voice demo on stage did not work (Engadget), and Willison notes that similar OpenAI initiatives have previously produced "mixed results."
- Vendor-provided numbers. All GPT-6.1 Sol benchmark results come from OpenAI; there was no independent verification at the time of publication.
- Regional restrictions. Dots launch only in supported countries, while Ultrafast is available only with U.S. processing or global processing, with no EU endpoints.
- Few details. OpenClaw Enterprise Harness was mentioned only briefly, pricing for the Decisions API and additional Dots was not disclosed, and the Private Inference date is unknown.
What all of this means for businesses in Russia
Let's start with the main limitation. In the list of supported countries by OpenAI there is no Russia, and OpenAI itself warns that access to its services from other countries or enabling such access may result in account suspension. That means building a mission-critical process in a Russian company directly on OpenAI products is risky, regardless of their quality. In addition, transferring personal data to a foreign provider falls under the requirements of Federal Law No. 152-FZ, and for certain industries, starting March 1, 2027, there may be a list of cases where only sovereign and national models are permitted (our analysis of Law 243-FZ).
But DevDay is also useful for those who will not use OpenAI directly: it shows where the entire market is headed, and these approaches carry over to other models.
- The agent as a permanent employee, not a chat window. Dots show the model everyone is moving toward: the agent works in the background, sees email, messaging apps, and CRM, and acts based on approval rules. This can be built on other models too, including ones deployed inside the company's own environment.
- Working through the interface instead of integration. Computer use in the Agents API lowers the barrier to automating systems without an API. For Russian companies, the analogous task is agent-based automation in 1C, ERP, and the browser.
- A low-cost model for most steps. The price gap between the flagship model and the midrange model is roughly fivefold, while quality is close on many tasks. The architecture of "a low-cost model by default, a premium model for complex steps" is becoming the norm regardless of provider. How to build a layer that lets you swap models without redesigning the workflow is covered in the article on a unified LLM API for business.
- Narrow use cases — with a separate fast model. Decisions API confirms that classification and routing are better handled by a small model with a fixed answer set than by a flagship model.
- Data control — as a separate product. Private Safety Processing with customer-owned keys shows that the question of "where the data is stored and who can see it" has become a product issue, not just a legal one. For Russian companies, the answer is often to deploy the model on their own servers.
Which option to choose for your task: a quick reference
The table below is an editorial selection framework based on OpenAI's stated capabilities and pricing. It does not replace testing on your own data.
| Task | Which DevDay announcements fit | Why | What you risk |
|---|---|---|---|
| Complex programming, high-volume agent workflows | GPT-6.1 Sol | Close to Astra at 5x lower cost | Test results are still only from OpenAI |
| The most complex tasks, where mistakes are costly | GPT-6 Astra | Flagship model | $10/$50 pricing, Critical cybersecurity status |
| Real-time response to a human | Ultrafast | Up to 6x faster in the API | 6x pricing, processing only in the US or globally |
| Classification, routing, list selection | Decisions API on Luna | About 150 ms, fixed responses | Limited preview, price unknown |
| Automating software without an API | Agents API with computer use | The agent works through the interface | Broad agent permissions, beta |
| Personal executive assistant | Dots | Runs in the background, 4,000+ apps | Early version, access only on higher-tier plans |
| Team collaboration with agents | ChatGPT Space | Documents and agents in one place | Data stays with OpenAI, migration from Google and Notion |
| Code and vulnerability review | Codex Code Review and Security Cloud | Scheduled scans, ready-made fixes | Agent access to repositories |
A practical sequence for a company already working with OpenAI from a supported country: first recheck current tasks on GPT-6.1 Sol — that is the fastest way to cut costs; then look at the Decisions API for classification workflows; try Dots and Space on noncritical processes in one to two weeks, once things stabilize.
How we prepared this review and its limitations
This review was prepared on September 29, 2026, the day of the conference. Workflow:
- We compiled the official list of announcements from the OpenAI thread on the developer forum.
- We verified dates, model names, and prices against the API changelog and the developers.openai.com documentation: Ultrafast, Private Safety Processing, Sign in with ChatGPT, and the list of supported countries.
- We added details from reports by 10 outlets: CNBC, BGR, Engadget, 9to5Mac, TNW, The Decoder, The New Stack, Decrypt, Every, and Simon Willison's live stream.
- Where sources differ — on Dots access and the status of GPT-6.1 Sol — we showed the discrepancy instead of picking the convenient version.
- Our cost-per-task calculations are based on published prices and are labeled as a hypothetical example.
Limitations:
- the announcement pages on openai.com (recap, GPT-6.1 Sol, Dots, Agents API) were not accessible to our loading tool, so we verified their content through the changelog, the forum, and media coverage;
- we did not test the products ourselves; the test results are OpenAI's data;
- the prices for additional Dots and the Decisions API, the date of Private Inference, and details on limit resets were not known at the time of publication;
- availability of specific features depends on country, plan, and phased rollout.
Frequently Asked Questions
When did OpenAI DevDay 2026 take place?
OpenAI DevDay 2026 took place on September 29, 2026, at Fort Mason Center in San Francisco. Sam Altman's keynote began at 10:00 a.m. Pacific Time (8:00 p.m. Moscow time) and was streamed for free.
How much does GPT-6.1 Sol cost?
In the API, GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, $2.50 per million cached write tokens, and $10 per million output tokens. That is five times cheaper than GPT-6 Astra on input and output.
How is GPT-6.1 Sol different from GPT-6 Sol?
GPT-6.1 Sol was released a week after GPT-6 Sol with the same input and output pricing, but with cached input that is twice as cheap. According to OpenAI, on DeepSWE v1.1 it is 6.4 percentage points better than GPT-6 Sol and matched GPT-6 Astra.
What are Dots in ChatGPT?
Dots are always-on agents inside ChatGPT, powered by GPT-6 Astra. Each has its own cloud computer and browser, access to more than 4,000 apps through plugins, voice interaction, and rules for which actions require approval. Available on Pro and Business Premium plans in supported countries.
What is Ultrafast in OpenAI?
Ultrafast is a high-speed service tier: up to 8x faster generation in Codex and up to 6x faster in the API, at 6x the standard price. It currently works with GPT-6 Astra, and the version for GPT-6.1 Sol is promised later. In ChatGPT, it is available on Pro 500 and Enterprise.
Can you use the new DevDay 2026 features in Russia?
Russia is not on OpenAI's list of supported countries, and the company warns that accounts may be blocked when accessed from other countries. For critical processes, Russian companies are better off adapting DevDay approaches to models that are legally available, including those deployed on their own servers.
How the airassvet.ru team helps implement agents inspired by DevDay 2026
If after DevDay you want an agent that runs in the background with your systems, but you cannot or do not want to depend on OpenAI, the team airassvet.ru can:
- design and implement an AI agent connected to your communication channels and business systems, with rules for which actions it performs on its own and which require employee approval;
- set up agent-based automation for systems without a convenient API — 1C, ERP, and web interfaces — that is, do what OpenAI showed as computer use in the Agents API;
- deploy and fine-tune a language model on the company's servers (on-premise), so requests and data stay within your environment;
- audit the process, prepare the data, build an MVP, and test it before launch.
The first step is to choose one process you would like to hand over to an agent and document how it works today: what data and systems are involved, what constraints exist, and what criteria define a successful agent run. If you have such a process, discuss the project.
Conclusion
OpenAI DevDay 2026 cemented the shift from chatbot to agents that work on their own. The most practical announcement for teams already working with the API is GPT-6.1 Sol: results close to GPT-6 Astra at five times lower cost and with cached output that is twice as cheap. The most ambitious is Dots, always-on agents with their own computer, but the first version is rough and available only on higher-priced plans. For developers, the most important pieces are computer use in the Agents API, the Decisions API for fast narrow decisions, and cloud Codex. Russian companies should keep in mind that OpenAI does not officially operate in Russia and should transfer the approaches themselves — background agents, a cheap default model, data control — to available models. The right place to start is with one process and an honest comparison of the cost of the resulting output.