OpenAI DevDays Full Recap: GPT-6.1 Sol, Decisions API, and Dots Explained

OpenAI dropped a wave of updates at the latest DevDays — from the new GPT-6.1 Sol model to a Decisions API that takes aim at Jeff, plus the enterprise assistant Dots and an upgraded Codex CLI. This article is based on a YouTuber's hands-on walkthrough of the event, helping you quickly identify what's already available, what's still in preview, and what's locked behind specific subscription tiers.

GPT-6.1 Sol: A Cheaper Model That Disappoints in Practice

GPT-6.1 Sol was the most talked-about model announcement at DevDays. OpenAI's pitch is straightforward: lower cost than Astra, with comparable performance. On paper, the official benchmarks mostly back that up — on the Deep Software Engineering benchmark, GPT-6.1 Sol even outperforms both GPT-6 So and GPT-6 Astra, while largely matching them on other benchmarks at significantly lower cost. By that measure, it qualifies as a frontier model.

But benchmark scores don't tell the whole story. The YouTuber used a more practical evaluation method he calls the "yapping frequency" test. He asked a deliberately misleading question in his YouTube channel management software: how to modify the data model to separate emails from sponsor labels? The correct answer is actually "no data model changes needed — this is purely a UI change."

The real test wasn't whether the model got it right, but how it responded when the user kept pushing back ("Are you sure that's the clean solution?"). The results were disappointing: GPT-6.1 Sol quickly caved under pressure, repeatedly second-guessing itself with sycophantic responses like "You're right, my earlier suggestion..."

By contrast, Anthropic's Opus 5.5 held its ground under the same follow-up pressure. The reviewer was blunt: as long as Opus 5.5 is available, he won't be touching GPT-6.1 Sol, calling it a "serious step backward" in interaction style. This subjective finding serves as an important reminder: a model's "character" and resilience under pushback may matter more to real-world usability than any benchmark score.

Decisions API: A Fast Decision Engine to Rival Jeff

In the reviewer's view, the Decisions API is the most exciting announcement from the entire event. It's OpenAI's answer to Jeff — a System 1-style model that doesn't follow the traditional next-token prediction path, but instead directly predicts probabilities across a set of possible outcomes.

In other words, it doesn't generate text. It assigns probabilities to a set of candidate answers. This makes it particularly well-suited for classification, scoring, and binary (yes/no) judgments, with a clear speed advantage. Crucially, the Decisions API also supports image input, opening up exciting possibilities for computer vision and robotics applications.

The API is currently in limited preview, with a broader rollout planned in the near future. The reviewer called it the feature he's most excited about and most likely to cover in a dedicated deep-dive video.

---

The Jeff referenced here is a fast decision-making model previously released by Anthropic, also positioned as a System 1-style intuitive reasoner — mimicking the kind of immediate human response that doesn't require deep deliberation. In contrast, System 2 models focus on slow, step-by-step chain reasoning (like the o1 and o3 series). Traditional LLMs use next-token prediction to generate text word by word, running a softmax over the entire vocabulary at each step, which introduces latency. The Decisions API bypasses text generation entirely and outputs a probability distribution over candidates — essentially working like a classifier — giving it a meaningful edge in latency and throughput. This type of model has strong potential for content moderation, intent recognition, A/B scoring, and other scenarios that don't require open-ended text output.

Dots: OpenAI's Managed AI Assistant

Dots can be thought of as a hosted, cloud-managed version of a Hermes agent or OpenClaw^Ad — an AI assistant that runs 24/7 and can proactively notify you about things. It connects with messaging platforms like Telegram, Slack, and WhatsApp, and integrates with your ChatGPT account to run ChatGPT threads and interface with various services on your behalf.

Its core pitch is "managed" and "more secure," targeting enterprise use cases. That said, it's currently only available to Pro and Business Premium subscribers — standard ChatGPT Plus users can't access it yet. The reviewer's take: unless Dots demonstrates unique value beyond what existing open-source solutions offer, its appeal will remain largely limited to business users.

Developer Tool Upgrades: Codex CLI and Agents API

Codex received two meaningful upgrades this time around. First, Codex can now run workloads in the cloud — no need to keep your machine on or consume local resources. It clones repos, manages secrets, and gets work done entirely in a cloud environment, and this is available across all subscription tiers.

The Codex CLI itself has also been polished, with a cleaner interface and improved usability. New additions include the ability to turn repeatable instructions into "skills," along with commands like codex resume, codex --image, sub-agents, and slash commands for permission management. A small but handy improvement: selecting text in a response now automatically copies it.

That said, the reviewer admitted he's still not happy with the Codex CLI's UX — it lacks voice input and an undo option, and he personally prefers Cloud Code overall. The CLI is a powerful execution framework, but still has room to grow in ease of use.

Another major update for developers: the Agents API now supports computer use. When working with agents in Python code, you can now instruct them to directly "use a browser" to complete tasks. The reviewer demonstrated two approaches: one where OpenAI handles execution in the cloud (you see console output and need to manually authorize), and one using Playwright to run locally on your machine (opens a real browser, runs faster).

Both examples performed the same task — visiting Wikipedia and retrieving the page title and URL — providing a clear, practical illustration of the difference between cloud-hosted and local execution paths.

---

Playwright, referenced above, is Microsoft's open-source browser automation framework supporting Chromium, Firefox, and WebKit, commonly used for end-to-end testing and web scraping. Compared to OpenAI's cloud-hosted approach, a local Playwright setup has key advantages: the browser runs directly on the user's machine, allowing access to local session state, bypassing certain anti-scraping mechanisms, and avoiding the need to send page content to a remote server — resulting in lower latency and reduced privacy exposure. The tradeoff is that users need to install and maintain the local environment themselves. MCP (Model Context Protocol) is an open protocol proposed by Anthropic to standardize communication interfaces between AI models and external tools or data sources. OpenAI's support for building ChatGPT plugins via local MCP servers signals that both companies' tool ecosystems are converging toward a shared protocol layer.

Other Notable Updates

Beyond the highlights above, the event covered a range of features for different audiences:

• Ultrafast: A new premium speed tier with models running up to 8x faster at 6x the price. Worth the math for latency-sensitive use cases.
• Private Intelligence: An enterprise offering for organizations handling sensitive data, emphasizing privacy and security — zero data retention, secure processing, and no use for model training. Requires a separate agreement with OpenAI rather than a simple in-account toggle.
• ChatGPT Space: A collaborative workspace for higher-tier subscribers that goes beyond agents, coding, and conversation to generate pages, slides, documents, and other "artifacts" — think a hybrid of Canva and a collaborative doc editor, with a shared knowledge base.
• Plugin Extensions: Developers can build their own ChatGPT plugins using TypeScript or Python SDKs. The reviewer demonstrated creating a to-do app plugin with a UI via a local MCP server (exposed publicly through ngrok) and importing it into the ChatGPT sidebar.
• Sign in with ChatGPT: Now available on platforms like Devin and Notion, allowing users to leverage their ChatGPT subscription credits directly within those apps without purchasing separate API credits.
• ChatGPT Security Cloud: An interface for defensive cybersecurity use cases.

---

Sign in with ChatGPT works similarly to "Sign in with Google" as an OAuth-style authorization flow, but with a key difference: once a user authorizes, third-party apps (like Devin or Notion) can directly tap into the API quota and model capabilities included in that user's ChatGPT subscription — without the app vendor needing to purchase their own OpenAI API credits. For developers, this means a lower acquisition cost with no need to build a separate billing system. For users, it means subscription benefits are reusable across platforms. If this scales broadly, it transforms a ChatGPT subscription from a single product into a universal pass for AI capabilities — a strategically significant move in OpenAI's platform ambitions.

Takeaways

The signal from this DevDays is clear: OpenAI is pushing hard to bring its capabilities deeper into developer tooling and enterprise products. The Agents API's computer use support, cloud-based Codex, and the plugin ecosystem are all expanding the boundaries of what developers can build. The Decisions API represents a focused bet on the fast-decision niche.

From a user experience standpoint, however, the new GPT-6.1 Sol model's sycophantic behavior under pressure drew sharp criticism from hands-on testers — a reminder that trustworthiness and resilience under pushback are becoming more important differentiators than benchmark scores. For everyday users, it's worth focusing on what's already available: Agents API, cloud Codex, and plugin extensions. Decisions API and Dots are worth keeping an eye on as they roll out more broadly.