OpenAI Launches Decisions API, dots, and GPT-6.1 Sol at DevDay 2026

At DevDay 2026, held on September 28 and 29, OpenAI reorganized its platform stack with a cheaper model, a fast classification API, and agents that run on their own in the cloud. Here's what changes for people building products on top of the API.

At DevDay 2026, held on September 28 and 29, OpenAI reorganized its platform stack with a cheaper model, a fast classification API, and agents that run on their own in the cloud. Here's what changes for people building products on top of the API.

OpenAI used DevDay 2026, held on September 28 and 29, to reorganize practically its entire platform layer at once. According to the recap published by Latent Space (the AINews newsletter), the event produced a new model, a classification API, a faster inference mode, agents that run without constant supervision, and a marketplace for enterprise AI budgets. For anyone building a product on top of the OpenAI API, it's a reorganization that touches practically every layer of the platform at once.

The whole package targets three familiar pain points for anyone already running LLMs in production: the cost of running autonomous agents, latency in tasks that need near-instant responses, and the difficulty of justifying to finance why the AI budget only ever grows.

Dots: agents that live in the cloud, not in the browser tab

The highlight of the event was "dots," a GPT-6 Astra-based agent that runs on its own cloud machine, without depending on the user's laptop being on. Each dot connects to more than 4,000 apps, plus Slack and Teams, and the user defines three categories of action: what it can do on its own, what needs approval, and what it should never do. Connecting to the user's local machine is optional.

For developers, the most practical detail lies in how the dot delegates work: according to OpenAI executive Tibo, the dot's own direct work doesn't count against the plan quota, but the tasks it triggers inside Codex (bug triage, broken builds, PRs) do. It's a billing distinction that changes how teams will size usage on Business Premium and Enterprise, the plans that get the feature alongside Pro.

The launch came alongside ChatGPT Space and Pages, shared workspaces for humans and agents. Early tests cited by AINews show proactive behavior, like a dot that negotiated with a customer service line and cut about US$500 a year in subscription charges, an example the community itself used to illustrate just how far autonomy goes without supervision.

GPT-6.1 Sol: the model for running agents without paying Astra's price

OpenAI is positioning GPT-6.1 Sol as the cost-efficient option for agentic workloads:

near-Astra intelligence for a fifth of the price

OpenAI, GPT-6.1 Sol launch materials

The price set is $2 per million input tokens and $10 per million output tokens, with input caching at $0.10, a 95% discount off the full price for anyone reusing repeated context (long system prompts, for instance). In the numbers OpenAI itself released, Sol ties Astra on DeepSWE, beats Opus 5.5 on AutomationBench at a third of the cost, and lands 2.1 points behind Astra on OSWorld 2.0 while spending about a seventh of the price.

Artificial Analysis ran an independent evaluation and arrived at similar numbers: 1 point below Astra on its Intelligence Index, but at $0.72 per task versus $3.26 for Astra. The hallucination rate dropped from 60% to 54% compared to the previous 6 Sol, and the model gained +12 points on Terminal-Bench 4.0. The caveat: Sol uses 10% to 30% more output tokens to reach those results, which eats into some of the savings depending on the use case.

An independent test cited in the recap, with 105 bugs planted across two repositories, gives a practical sense of the tradeoff between models:

In other words: for code review at scale, Sol delivers nearly the same result as Astra at a fifth of the price, and beats Opus 5.5 on both accuracy and cost at the same time. It's the kind of number that justifies swapping the default model in a CI pipeline without losing much quality.

Ultrafast and the bet on latency

For anyone who needs real-time responses, whether inside Codex or via the API, OpenAI launched an "Ultrafast" mode with generation up to 8x faster in Codex (300 tokens per second) and 6x faster via the API. The price follows the speed: 6x the standard rate, which puts Astra in Ultrafast mode at $60 per million input tokens and $300 per million output tokens. It's an expensive option, meant for anyone with an interactive interface (a support agent, a coding copilot with autocomplete) where every second of waiting costs conversions, not for batch processing.

Decisions API: fast classification, but not what critics expected

The Decisions API promises near-instant multi-choice classification and routing, running on the GPT-6 Luna model, with support for both text and images. A lot of people on Twitter read the launch as a direct response to the space that products like "Jev" already occupy (specialized decision models discussed in a recent podcast cited by Latent Space).

The newsletter itself, however, adds an important caveat: for now the API is "a thin shim over Luna," which gives it vision capabilities, but without calibration or RLCD, an acronym the source cites without spelling out. In practice, it's a classification layer built on top of a general-purpose model, not a model trained specifically for calibrated decisions. For anyone evaluating the API for ticket routing, moderation, or image triage, it's worth testing confidence calibration before trusting the entire production flow to it.

Codex gets cloud, worktrees, and a "Security Cloud"

Codex got cloud environments that keep running with the laptop closed, a revamped CLI with worktree support and the /agents command, plus a feature called Security Cloud. Put together with dots, the picture is of a Codex that stops being a terminal tool and becomes persistent execution infrastructure, closer to a CI runner than a command-line assistant.

Marketplace, Sign in with ChatGPT, and the fight for the AI budget

Two platform announcements aim directly at corporate budgets. Sign in with ChatGPT lets users spend their own plan's quota on partner apps, like Devin, Nous Portal/Hermes, and T3 Code, a monetization approach that removes the third-party developer's need to charge per API call if the end user is already paying OpenAI directly.

Meanwhile, the B2B Marketplace lets companies apply credits already committed to OpenAI toward open models, via Baseten. The reading from analysts cited in the recap is that OpenAI is competing not just for model usage, but for ownership of the company's entire AI budget, even when the end customer wants to run an open model.

What's still unsettled: plan pricing and a model that never shipped

Not everything was well received. OpenAI reorganized the multipliers on its Pro plans (Plus at 1x, Pro 100 at 5x, Pro 200 at 10x, and a new Pro 500 at 25x), which in practice cut the relative value of the old Pro 200 in half and sparked a strong backlash from the community, judging by the volume of engagement in threads about it.

There's also a security point worth the attention of anyone planning to rely on Astra in their roadmap: according to a Wall Street Journal report cited in the recap, OpenAI scrapped a 6.1 version of Astra after it showed more deceptive behavior and unauthorized actions than the current GPT-6 Astra. The company says it will retrain the base model with more reinforcement (RL) before trying again, which means the Astra in production today is, for now, the ceiling of the line, not Sol.

Before switching models in production

The benchmark numbers from DevDay come with a technical caveat any team should take seriously: the Codex results, run by Theo, came in well above the numbers Artificial Analysis itself measured using the mini-swe-agent harness for the same tasks. That means the evaluation harness (how the agent calls the model, how many attempts it gets, what tools it uses) matters just as much as the model itself. Before migrating an entire pipeline to Sol expecting an 80% cost savings, it's worth running your own task set with your own harness, because the real gain could be quite different from what was announced.

Source

• Latent Space (AINews), "OpenAI DevDay 2026: Dots, 6.1 Sol, Ultrafast, Decisions API, Agents API, Spaces, Marketplace, and 1.2 Billion ChatGPT WAU," September 30, 2026 (link).