OpenAI DevDay 2026: GPT-6.1 Sol API Pricing, Decisions API Preview, and Computer Use Access

Bottom line: OpenAI’s DevDay made gpt-6.1-sol available in the API at $2 input, $0.10 cached input, and $10 output per million tokens. The Decisions API is still a limited preview, and computer use depends on your plan rather than being enabled on every key.

The announcement page has twenty items. The call you can make today has a model id and a price. OpenAI’s Sept. 29 DevDay recap introduces GPT-6.1 Sol as near-Astra intelligence at a fifth of Astra’s standard token prices, and it introduces a Decisions API that is in limited preview with a broad release “in the coming days.” The Sol product page is the one with the numbers a bill will actually use. If you wire the preview first because it was higher on the recap, you will be waiting on a form while the model id was already live.

We wrote the last Sol price cut, which is not this model, in Opus 5.5 reset the API prices, and Sol halved them. We wrote compaction as a knob you choose, on a different vendor, in Claude API context compaction. We wrote an install that is two commands and a second step people skip to, in OpenShell’s install is two commands, Sentry is not. DevDay is the same shape. The recap is a menu. The first call is not the first heading.

The model id is gpt-6.1-sol, and Chat is not the surface

The Sol page says GPT-6.1 Sol is available starting the day of that post to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and in Codex. It is not yet available in Chat. Developers reach it in the API as gpt-6.1-sol. Standard prices are $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. Cached input is 95% less than standard input on that sheet, and 50% less than GPT-6 Sol’s cached input. The recap’s “fifth of Astra” line is the comparison.

The recap says GPT-6.1 Sol is available to all API, Plus, Pro, Business, Enterprise, and Edu users. The product page adds the split that matters if you are clicking around the consumer app: Work and Codex yes, Chat no. If a teammate says they cannot see it, ask which surface. “ChatGPT” is three products this week.

On OSWorld 2.0’s offline set, partial reward, release v2026.08.08, GPT-6.1 Sol beats GPT-6 Sol by seven percentage points at maximum reasoning effort, at less than half the cost. It comes within 2.1 percentage points of Astra at maximum reasoning, at roughly one-seventh the cost per task. A vendor score is a hypothesis. $2 and $10 are a line item.

Near is their adjective. The id in the docs is Sol’s. If a workflow was pinned to an Astra snapshot, this page is not a migration guide. It does not list a parameter that means “be Astra.” It lists a cheaper model that they say lands close on agentic coding, computer use, and professional work. Run your task. Do not rename the model in a comment and call it done.

Decisions API is a preview, so it is not step one

The recap’s description of Decisions API is short, and the shortness is the instruction. It focuses Luna on a set of user-defined questions with finite, pre-defined answers. You supply context as text or images. You get answers you can use to classify content, route a request, or pick an agent’s next action. It is in limited preview, with a broad release planned in the coming days. The recap does not print a price. It does not print a request body.

That is a classifier with a closed label set, sold as a separate API rather than as a prompt you hope obeys a schema. The design claim is real-time decisions, not open-ended writing. If your job is “which of these four queues,” this is the shape they are pointing at. If your job is “write the reply,” you are on the wrong heading, and Sol is the model id from the previous section. “Coming days” is not a date. A limited preview is a waitlist with better stationery.

When the broad release shows up, the thing to check first is still the label set. Finite answers are the product. If the docs that land later let the model free-text the decision, the recap’s claim has moved, and you should notice. Until those docs exist in a form you can call without a sales conversation, leave it off the critical path. We have watched “limited preview” sit there longer than a sprint.

Luna is the model name they attach to this API. It is not, on these pages, the same string as gpt-6.1-sol. Do not point a Decisions call at Sol because both were announced on Tuesday. Different jobs. If you only remember one id from this piece, remember gpt-6.1-sol, because that is the one with a price and a generally available API line.

Computer use is a plan gate, not a checkbox on every key

The Agents API now supports computer use, the recap says, so developers can build agents that interact with software to complete tasks. The same paragraph says it brings Codex’s multi-agent capabilities, tool search, tool calling, and context compaction into your application, and that OpenAI runs the infrastructure. The availability line is the one people will skip. It is available through the API, and in Codex and ChatGPT Work, on Pro 500 and Enterprise. The docs link they print is the computer-use tool guide: developers.openai.com/api/docs/guides/agents-api/tools/computer-use.

Pro 500 is not “Pro.” If the seat is Plus, or Pro without the 500 in the name, the recap does not put computer use there. Enterprise is on the list. A startup key on the default tier is not. Read the plan before you write the integration ticket.

The idea, summarize a long run so the next step fits, is not new this week. What is new is OpenAI saying the Agents API now carries Codex’s version of it, along with tool search and multi-agent, on their infrastructure. The recap does not include one. If you already compact on another vendor, do not assume the field names survived the flight. Read the guide they linked. Then decide whether you want their loop or yours.

Computer use also means the agent is clicking something. That is a production question, not a demo question. Which desktop, whose credentials, what the sandbox is allowed to open. The recap says OpenAI runs the underlying infrastructure so your team can focus on the application. That sentence is a convenience and a trust boundary. Their machines are in the path. If your security review cannot accept that, the feature is not “on” just because the docs page loads. Bedrock Managed Agents, also in the recap, is the other door: OpenAI worked with Amazon so agents built on the Agents API can run entirely in AWS and talk to AWS resources. That is a different trust story. It is not the default of the computer-use paragraph. Do not conflate them in the design doc.

Ultrafast is a speed tier with two clocks

Ultrafast, in the recap, is a premium speed tier. Up to 8× faster token generation in Codex, which they quantify as 300 tokens per second, and up to 6× in the API. GPT-6 Astra Ultrafast is available in the API, and in ChatGPT Work and Codex, on Pro 500 and Enterprise. GPT-6.1 Sol Ultrafast is “coming soon” in that recap. The Sol page says “in the coming days,” and limits the 8× claim to Codex compared with Sol’s own standard speed.

Astra Ultrafast is the one they say is on today, and only on those plans. Sol at standard speed is the one with the $2 / $10 price. If the ticket says “Sol, but fast,” you are writing a ticket for a tier that the product page has not dated. Ship Sol at the posted price. Add Ultrafast when the dashboard actually offers it for that model id, and remeasure. An 8× claim against their own standard is not an 8× claim against your current p95. Your p95 includes your tools, your retrieval, and the queue. Theirs is token generation.

300 tokens per second is a generation rate, not an end-to-end request rate. A computer-use loop that waits on a screenshot will not hit 300 just because the text channel can. They should not.

What can wait

Dots are always-on agents, available on Pro and Business Premium in eligible markets. Enterprise, Edu, and Healthcare can try a beta when a workspace admin enables it. It is off by default. That default is the correct setting until you have read what “always working on your behalf” means for your data.

Private Inference is a preview “this fall,” behind a contact form, in the Private Intelligence section. Zero Data Retention with Private Safety Processing is a separate docs link in that same blurb. Fall is not this week’s setup task. If a procurement thread is asking for confidential computing, send them the contact form and do not promise the preview is generally available.

Codex CLI did get a change you can try without a new contract. Voice to start and steer tasks, an /agents view for delegating and tracking, better prompt editing, session resume, worktrees, and a cleaner terminal. Available to all plans, the recap says. That is a client update. It does not change the model id. Update the CLI if you live in it. Do not tell yourself the CLI refresh is the Sol migration. They shipped on the same day. They are not the same commit.

Point a non-production Codex or API call at gpt-6.1-sol and read the invoice against $2, $0.10 cached, and $10. Confirm you are in Work or Codex or the API, not in Chat, if the picker is empty. Leave Decisions API on the preview list until it has a public price and a request you can copy from docs. Open the computer-use guide only if the seat is Pro 500 or Enterprise, and only after someone has named the sandbox. Ignore Ultrafast for Sol until the control exists. Leave Dots off. The other seventeen announcements can wait until one of them is blocking a task you already have.

What this means for you

A sensible rollout order is: switch eligible workloads to gpt-6.1-sol first, apply for the Decisions API preview separately if you need it, and check your plan’s computer-use eligibility before building features that depend on it.