OpenAI DevDay 2026: Dots, GPT-6.1 Sol and What You Can Use

Practical launch guide · Official sources checked October 1, 2026

OpenAI DevDay 2026 changes how you delegate work more than it changes the answer to “which chatbot should I buy?” Dots adds a persistent assistant, Work handles deliverables, Space organizes shared material, and GPT-6.1 Sol adds another model choice. Those are different layers, with different access rules. Buying a higher plan before identifying the layer you need can solve the wrong problem.

The official event date is September 29, 2026. This is HiseHub’s source-based interpretation, not a hands-on review of every feature. Availability was checked October 1; rollout and workspace settings can still change what your account shows.

October 10 maintenance: the Ultrafast section now includes GPT-6.1 Sol, model-specific inference residency and separate API pricing. The original October 1 review of Dots, Work, Space and other launch features is unchanged. HiseHub has not run a speed or account-access test.

What was announced—and which problem does each part solve?

OpenAI’s DevDay roundup also covers reusable cloud coding environments, Security Cloud, selected Sign in with ChatGPT integrations and MCP Events. Do not treat an announcement list as proof that every feature has reached every account.

For a small business, the useful question is not “how many new features can I enable?” It is “where does my current workflow lose information, stop while my computer sleeps, or require repeated manual handoffs?” Choose one of those bottlenecks. A new assistant cannot compensate for an unclear deliverable or a missing source of truth.

Dots: always-on cloud work does not mean always-on access to your laptop

The most important distinction is between a Dot’s cloud computer and a connected personal computer. Cloud work can continue with your devices off. Work that needs your own computer requires that device online with the ChatGPT app open. Messaging a Dot does not automatically connect your local files or plugins. See OpenAI’s Dots documentation.

At launch, rollout covers adult Pro $100/$200/$500 users outside the EEA, UK and Switzerland, plus Business Premium and Enterprise worldwide. Enterprise access needs admin enablement. Setup starts on desktop; a supported mobile app can then continue the same Dot. Missing access is therefore not necessarily a login error or an insufficient credit balance.

Consider a simple weekly competitor-price brief. A cloud-only task that reads public pages and prepares a draft can be a sensible first experiment. A task that must read a spreadsheet saved only on your laptop has a different dependency. If the laptop is unavailable, an assistant should report that missing input—not confidently fill the gap with an invented price.

Our recommendation is to give the assistant a small, explicit contract:

• Use named sources, not an unrestricted search for whatever looks relevant.
• Return a dated draft, source links, changes and unresolved gaps.
• Keep a record of the previous run so it can distinguish a real change from a repeated item.
• Stop before publishing, emailing, purchasing or changing production data unless you have deliberately authorized that step.

This is workflow design advice, not an assurance that Dots will never make a mistake. An always-on system increases the importance of limits because it can repeat an error without you watching each run.

Work across devices: continuation and execution are different

OpenAI’s Work guide distinguishes cloud coordination from local execution. Eligible new tasks can continue across desktop, mobile and web when sync is enabled. Local resources still depend on the connected computer; existing tasks do not all become local-access cloud tasks retroactively.

That distinction matters when you leave your desk. Being able to send the next instruction from a phone does not prove that the agent can read a laptop-only folder at that moment. The conversation can be portable while its inputs remain device-bound.

Before a long task, write down three things: the output location, the source location and the approval boundary. For example, a draft saved in a shared cloud workspace, based only on approved cloud documents, is easier to continue remotely than a task that depends on a spreadsheet and browser session on a sleeping desktop.

For a website owner, a useful first task is an editorial audit: give Work five published URLs, ask it to identify dated claims and broken internal links, and require a reviewable checklist. It has a clear stopping point and does not need permission to overwrite the site. Add implementation only after the findings have been checked.

GPT-6.1 Sol: check the product surface before the subscription

GPT-6.1 Sol is listed for Plus, Pro, Business, Enterprise and Edu in supported Work and Codex clients. Enterprise and Edu need an admin to enable it; Free and Go are excluded at launch. It is not an ordinary Chat model. OpenAI’s model availability guide is the source for these distinctions.

If you cannot find it, check Work versus Chat, your client, your selected workspace and its permitted models before paying for another plan. An API model listing does not override a managed workspace’s policy.

For developers, the API identity is gpt-6.1-sol. OpenAI documents a 1,050,000-token context window, a 922,000-token maximum input and 128,000 maximum output. Use Responses for tool-calling workflows; the documented Chat Completions support does not include tool calling. See the API model specification.

A larger context window is capacity, not an instruction to upload everything. A focused set of files with explicit acceptance checks is often a more useful experiment than a huge repository dump with “make this better.” Record where the model misses requirements and how much review the result needs.

What actually changed in the API price?

At Standard short-input rates, GPT-6.1 Sol is $2 per million uncached input tokens and $10 per million output tokens—the same base rates as GPT-6 Sol. Its cache-read rate is $0.10 rather than $0.20. Cache writes are $2.50. Above 272,000 input tokens, rates become $4 input, $0.20 cache reads and $15 output for the full request. These figures come from the OpenAI API rate card, checked October 1.

Worked example, not a performance measurement: one request with 100,000 input tokens, a verified 90% cache-read share and 2,000 total billable output tokens costs $0.049 at the stated GPT-6.1 Sol rates, versus $0.058 for GPT-6 Sol. That excludes the earlier cache-write cost, tools, premium processing and tax. A reused prompt is not automatically a cache hit.

A 300,000-input/2,000-output request with no cache reads costs $1.23, not the $0.62 you would get by incorrectly applying the short-input rate. Long-context billing changes the whole request; it is not a marginal charge on only the last 28,000 input tokens.

Load the cached example or load the long-input example in our calculator. These API amounts do not estimate the cost of a Work task in ChatGPT credits.

Ultrafast: check the model, region and billing route

As of October 10, Ultrafast supports GPT-6 Astra and GPT-6.1 Sol in Work and Codex on Pro $500 and eligible Enterprise/Edu plans. Enterprise access is off by default until a workspace owner enables it. Other self-serve plans cannot unlock it just by buying credits. See the current speed-mode requirements.

Inference residency is model-specific: GPT-6.1 Sol supports the United States and Europe (EEA + Switzerland), while Astra supports the United States only. This updates the Astra-only launch description; it does not establish access for your account. OpenAI’s October update documents Sol’s rollout.

For eligible ChatGPT-signed-in usage, Ultrafast consumes included limits at 8× Standard; purchased credits and Enterprise pay-as-you-go use 6×. OpenAI’s up-to-8× generation-speed claim is for Astra, not a guarantee for Sol or whole-task completion. With an API key, use API token pricing instead of those credit multipliers.

API example: choose the processing tier before comparing the bill

The API guide lists Ultrafast for all API users, with separate rate limits. It is not the same eligibility contract as a ChatGPT plan. These are GPT-6.1 Sol prices in USD per million tokens, at up to 272,000 input tokens, checked October 10 against the API rate card:

Hold usage constant: 100,000 input tokens with 90,000 verified cache reads and 2,000 billable output tokens gives $0.049 Standard, $0.098 Fast or $0.294 Ultrafast. Prior cache writes, tools, residency uplifts and tax are excluded. Above 272,000 input tokens, use the tier’s long-context rates for the whole request; do not multiply only the excess tokens. These are arithmetic examples, not measured speed or quality.

Our calculator’s manually verified rows use Standard rates; selecting Sol does not turn on a speed-tier multiplier. For a real comparison, keep reasoning effort and task requirements fixed, then record returned usage, accepted results, elapsed time and total spend. Website loading, tools, tests and approvals may dominate the wait even when generation is faster.

Space: organize source material before automating its use

Space provides editable pages and groups of related pages when enabled for your account. Sharing a whole space grants access to its pages; use a page’s own controls when you intend to share only that page. OpenAI’s Space setup guide explains the access boundary.

For a repeatable research workflow, maintain a short page defining the question, approved sources, date checked and unresolved decisions. Keep facts separate from draft recommendations. The goal is not another place to paste an AI response; it is a record that a later task or human reviewer can understand without guessing what you meant.

Which other DevDay announcements should developers investigate?

Three deserve a targeted follow-up, rather than a blanket migration. Reusable cloud coding environments can reduce repeated setup. Security Cloud is a research preview for GitHub-related security work, not a promise that a scan replaces a security review. Sign in with ChatGPT starts with selected integrations and eligibility rules; it is not a universal replacement for any application’s API key. The official roundup links their separate documentation.

If you operate a self-hosted interface, ask whether its adapter supports the particular authentication method and API features before switching users. A chat-compatible endpoint, an agent workspace and subscription-based authentication are not interchangeable integration contracts. Our Self-Hosted AI Tools directory can help shortlist interfaces, but does not certify every new provider feature.

A sensible first-week rollout

1. Choose one real task. A source-linked weekly brief or a small code change is better than “automate my business.”
2. Confirm access without buying first. Check the live surface, account, workspace and region in the Model Access Map and provider documentation.
3. Start read-only or with a reversible draft. Define required sources, output and a review point.
4. Record the actual outcome. Accepted result, missing facts, review minutes, elapsed time and billable usage are more useful than the number of agent steps.
5. Expand only after a repeatable success. Keep an exception path for missing inputs, outages and uncertain findings.

For refreshed launch facts, use the Model Release Tracker. For independent model evidence, use AI Model Comparison. For the bigger purchase decision, read AI subscription versus API costs. DevDay gives you more ways to organize work; it does not remove the need to choose the right access route, budget and review boundary.