What OpenAI announced
DevDay 2026 ran in San Francisco on Tuesday 29 September and produced more than a dozen announcements. Most of them are aimed at ChatGPT users. Five matter directly to people who build on the API or run coding agents in production:
- GPT-6.1 Sol. A successor to GPT-6 Sol, which shipped only a week earlier. List price is unchanged at $2 per million input tokens and $10 per million output tokens, which OpenAI frames as one-fifth of GPT-6 Astra's standard rates. Cached input halves from $0.20 to $0.10 per million tokens. It is available in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu customers.
- Dots. Always-on agents running on GPT-6 Astra. Each Dot gets its own cloud computer and browser, can connect to more than 4,000 apps through ChatGPT's plugin ecosystem, and can be addressed from ChatGPT, Slack or Microsoft Teams.
- Codex in the cloud. Reusable cloud environments: each Codex task launched from a published environment gets its own virtual machine and keeps working while your own machine sleeps. Alongside it, Codex Security Cloud scans GitHub repositories and prepares fixes.
- Ultrafast. A premium speed tier. Up to 8x faster token generation in Codex and up to 6x in the API, reaching as much as 300 tokens per second — at six times the standard rate. Live now for GPT-6 Astra; a GPT-6.1 Sol version is promised soon.
- Decisions API. A limited-preview endpoint, built on a version of GPT-6 Luna, that picks from answers you define rather than generating text.
If you followed the GPT-6 family launch, you already know the base prices. We covered them in the Sol and Luna price cut and in Astra's real price and its 272K context cliff. This piece is about what is new on top of those numbers.
GPT-6.1 Sol: the saving is in the cache, not the sticker
Most coverage will lead with "near-Astra performance at one-fifth of the price". That is OpenAI's positioning, and it is worth unpacking carefully. The one-fifth comparison is against Astra's standard rates of $10 in and $50 out. It is not a price cut against the model GPT-6.1 Sol succeeds — GPT-6 Sol was already at $2 and $10.
On capability, OpenAI's own figures, as reported by VentureBeat and Implicator, put GPT-6.1 Sol within about two percentage points of Astra on the OSWorld 2.0 computer-use benchmark. Those are vendor numbers on a vendor-chosen configuration, and different outlets quoted different benchmark slices and reasoning-effort settings. Treat "close to Astra" as a hypothesis to test on your own tasks, not a procurement fact.
The change that will actually move invoices is cached input. Here is how it plays out on an agent loop, where the same system prompt, tool definitions and accumulated context get resent on every step.
| Rate (per million tokens) | GPT-6 Sol | GPT-6.1 Sol | GPT-6 Astra (standard) |
|---|---|---|---|
| Input | $2.00 | $2.00 | $10.00 |
| Cached input | $0.20 | $0.10 | — |
| Output | $10.00 | $10.00 | $50.00 |
An illustrative worked example, with assumed volumes. Take an agent task that carries a 100,000-token stable prefix across 50 steps. That is 5 million input tokens of prefix per task. Uncached, at $2 per million, the prefix alone costs $10.00 per task. Cached on GPT-6 Sol, it cost $1.00. Cached on GPT-6.1 Sol, it costs $0.50. The volumes are assumptions chosen to show the arithmetic, not a measurement of any real workload — but the shape holds: the further your agents lean on caching, the more of the bill this change touches.
The saving only lands if your cache actually hits. Keep the system prompt and tool schemas byte-for-byte stable at the front of the request, append new context at the end, and log the cached-token count the API returns on every call. If you have never looked at your cache hit rate, our guide to prompt caching across Claude, GPT and Gemini walks through the layout rules.
The routing implication is the more interesting one. If GPT-6.1 Sol really does sit close to Astra on the tasks you care about, the escalation path from Sol to Astra in your cascade should fire less often — and every avoided escalation saves four-fifths of the token cost of that step. Re-run your router evaluation before you change any thresholds; the method is in model routing and cascades.
Dots: always-on agents, and the UK gap
Dots are OpenAI's bet that the agent should not live inside a chat window. A Dot runs on GPT-6 Astra, works from its own isolated cloud computer and browser, and pursues an assigned goal in the background. It can run recurring tasks, delegate to sub-agents and look for relevant new information on its own. OpenAI's examples include investigating bugs mentioned in Slack and drafting invoices for approval.
The control model is where builders should look first:
- Custom rules let you allow, block or require approval for specific actions.
- Background research is read-only on connected apps, according to OpenAI, so a Dot working unprompted cannot send messages or change content.
- Credentials: OpenAI says Dots sign in with saved passwords without exposing them to the model.
- Local access is opt-in. Access to your own files and machine is off by default.
Who can actually use it
Pro and Business Premium subscribers get their first Dot included at no extra charge, in eligible markets. For Enterprise, Edu and Healthcare workspaces, Dots are a beta that is off by default and must be enabled by an admin. And there is a regional catch that matters a great deal for half our readership: according to OpenAI's launch details, as reported by Mixed, Pro access to Dots excludes the European Economic Area, Switzerland and the UK at launch. Business Premium is not restricted the same way: OpenAI told the press that Business Premium users get Dots in all supported ChatGPT regions. OpenAI has not given a date for Pro in the excluded markets.
| Plan | Dots at launch | UK / EEA / Switzerland |
|---|---|---|
| Pro | First Dot included, in eligible markets | Excluded for now |
| Business Premium | First Dot included | Available in all supported regions |
| Enterprise, Edu, Healthcare | Beta, admin must enable | Check with your workspace admin |
In practice, an independent developer in London evaluating Dots on a personal Pro subscription cannot, while a colleague on a company Business Premium seat can. OpenAI has not published a full list of eligible markets, so a team in Bengaluru or Pune should confirm availability in its own account rather than assume it. If you are writing a cross-border pilot plan, do not let one office's experience stand in for the other's.
Build or buy an always-on agent?
Dots are a ChatGPT product, not a developer primitive. If you want an always-on agent inside your own product, the relevant launch is the computer-use capability OpenAI added to its Agents API on the same day, which lets agents operate an OpenAI-hosted browser. That fits the direction we flagged earlier in September, when OpenAI began pulling agents in-house ahead of the Agent Builder sunset. The decision now looks like this:
| Question | Lean towards Dots | Lean towards building your own |
|---|---|---|
| Who is the user? | Your own staff, inside ChatGPT, Slack or Teams | Your customers, inside your product |
| Where must data and logs live? | Comfortable with OpenAI's workspace controls | You need your own audit trail, retention rules or regional hosting |
| Which model? | GPT-6 Astra is fine | You route across vendors or need a cheaper model for most steps |
| Failure handling | Approval rules are enough | You need retries, checkpoints and durable state you control |
For the build path, the hard parts have not changed: durable execution so a long-running job survives a crash (see background agents and durable execution), and a sandbox with least-privilege network access (see sandboxing AI agents). A Dot that only reads by default is a sensible design. An agent you build should copy that default.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →Codex Cloud and Codex Security Cloud
Codex in the cloud runs on the same idea behind OpenAI's purchase of Ona: coding agents that keep working unattended. You publish a reusable environment — repository, dependencies, configuration — and each task launched from it gets its own virtual machine. Close the laptop and the task keeps going; reopen it from a phone or browser later. It is rolling out across Plus, Pro, Business, Enterprise, Edu and Healthcare.
According to SiliconANGLE, a few details are worth knowing before you plan around it: Plus subscribers get smaller machines with half the processors and memory of higher tiers; tasks can be recovered for up to seven days after last use; secrets can be injected through a proxy so the real values are not exposed in the environment; GitLab and self-hosted GitHub Enterprise Server are not yet supported; and computer and browser use do not work inside these cloud environments. That last point matters: Codex Cloud is a coding sandbox, not a general agent computer.
Codex Security Cloud scans whole GitHub repositories on demand or on a schedule, checks new commits, removes duplicate findings and prepares fixes. It includes OpenAI's Daybreak Blue cyber models without a separate Daybreak application. The Decoder reports availability for Pro, Business, Enterprise and Edu on desktop and web.
A scanner that prepares fixes is a scanner that writes code into your repository. Route its output through the same pull-request review as any human change, and do not grant it merge rights. For regulated codebases in UK financial services or Indian fintech, write down who reviews an AI-prepared security fix before the first one arrives — auditors will ask.
Ultrafast: when six times the price is worth it
Ultrafast is the easiest announcement to misuse. GPT-6 Astra Ultrafast in the API costs around $60 per million input tokens and $300 per million output tokens — six times standard. In Codex and ChatGPT Work it sits behind Enterprise and a new Pro 500 plan at $500 a month with 25 times Plus usage. If the same 6x multiplier applies to GPT-6.1 Sol Ultrafast when it ships, that would put fast Sol at $12 in and $60 out, above standard Astra.
Speed is worth paying for only when someone is waiting and the wait costs money. An illustrative task with 200,000 input tokens and 20,000 output tokens costs about $3 on standard Astra and about $18 on Ultrafast, before caching. That is the whole decision in one line: is the time saved on this task worth $15?
| Workload | Ultrafast? | Why |
|---|---|---|
| Live incident debugging with an engineer at the keyboard | Often yes | Engineer time and downtime dwarf token cost |
| Interactive pair-programming in Codex | Sometimes | Worth trialling; measure whether faster tokens change how much gets done |
| User-facing flow with a hard latency budget | Test first | Time to first token and network may dominate; see latency budgets |
| Overnight batch, background agents, Codex Cloud tasks | No | Nobody is waiting; you pay 6x for nothing |
The cleanest policy is to make Ultrafast an explicit, per-request opt-in tied to a named workload, with a budget alert, rather than a default anyone can flip on. Finance teams in London and Bengaluru will both ask the same question at month end, and "it felt snappier" is not an answer.
The Decisions API: a small launch that could replace glue code
The Decisions API does one thing: you pass context (text or image) and a closed list of options, and it returns the most likely option with a confidence score. OpenAI says it answers in about 150 milliseconds, against roughly 1.6 seconds for GPT-6 Luna doing the same job. It is in limited preview with broader availability promised shortly, and pricing has not been published.
That is a direct fit for the classifier steps in most agent stacks: intent routing, "should this escalate?" checks, and choosing the next tool from a fixed set. If you currently prompt a small model to answer with one word and then parse it, this is the API-native version. Two cautions. Until pricing is public, you cannot compare it with a fine-tuned small model you host yourself. And a confidence score is only useful once you have checked it against labelled data from your own traffic — calibrate before you put a threshold on it.
What builders in India and the UK should do this week
- Re-price your agent loops on GPT-6.1 Sol. Pull a week of logs, separate cached from uncached input, and recompute. The saving scales with how cache-friendly your prompts are.
- Re-test the Sol-to-Astra boundary. If GPT-6.1 Sol handles tasks you currently escalate, the cost difference per escalated step is large. Measure on your own evaluation set, not on OSWorld.
- Check Dots eligibility per office. A UK team on Pro cannot use Dots today; a Business Premium seat can. Do not run a pilot that only one region can reproduce.
- Keep Ultrafast off by default. Opt in per workload, with a spending alert.
- Plan for portability. Dots, Codex Cloud and the Agents API all pull state onto OpenAI's infrastructure. That is convenient, and it is also lock-in. Keep your prompts, tool schemas and evaluations in your own repository so you can move.
DevDay's message is that OpenAI wants to run the agent, not just the model. For many teams that will be the right trade. Make it deliberately, with your own numbers in hand.