OpenAI DevDay 2026: Dots Agent Buggy as Sol Cuts API Costs

A sprawling DevDay with one unfinished centerpiece

OpenAI's DevDay 2026, held September 29, was less a model launch than a full product sweep. The company announced always-on agents called Dots, a cheaper GPT-6.1 Sol model, a cloud-hosted version of Codex, new agent-building APIs, and a collaborative ChatGPT workspace called Spaces.35 Early hands-on reports suggest the developer-facing pieces are the most solid part of the release. Dots, the consumer agent OpenAI put front and center, appears to have shipped before it was ready.

Dots: compelling idea, rough execution

Dots are agents built to keep working after the initial instruction. They run on GPT-6 Astra and get their own cloud computer and browser, so they can use connected tools and services while pursuing a goal.5

Every's review praised the ambition and was frank about the problems. The reviewer said Dots had become their main way of using ChatGPT, often chosen over starting a fresh Codex thread. They still could not recommend the tested version.2 They hit frequent permission errors and dropped messages, and their Dot often could not connect to the in-app browser.2 The bigger complaint was conceptual. Both the user and the agent got confused about where work should happen: on the Dot's own computer, in the main thread's browser, or in a separate thread the Dot had spun up and was orchestrating.2

Developers Digest's assessment lands in the same place. It frames Dots as a signal of where OpenAI is heading, not a stable surface to build on, and notes that Dots does not replace a developer API.3 Every warned that piling on this much at once risks losing the simplicity and coherence that made Codex good in the first place.2

Spaces drew a warmer response. Every called it an improvement and an easy drop-in replacement for Proof, a document editor it had previously relied on.2

Sol is the real developer story

For anyone building on the API, GPT-6.1 Sol is the most concrete takeaway. It is available as gpt-6.1-sol and supports the Responses API, tool calling, computer use, and MCP.3 OpenAI positions it as close to Astra in capability at a lower price, aimed at coding, computer use, and professional workloads.5

The pricing gap is large:4

• Fresh input and output: about one-fifth of Astra's rates.
• Cached input: about one-tenth of Astra's rate.
• Net effect: roughly 5.7x cheaper for comparable token usage, which matters most for agent workloads that reuse context heavily.

OpenAI's own benchmarks claim Sol matches Astra on DeepSWE v1.1 at about a fifth of the cost. On OSWorld 2.0, it says Sol comes within 2.1 points of Astra at maximum effort for roughly a seventh of the cost per task.4 These are vendor numbers, and Developers Digest notes that the open question is whether quality holds on real workloads.4

Alongside Sol, OpenAI opened a public beta of an Agents API for building managed Codex-style agents and launched Codex Cloud for hosted, asynchronous coding.3 Its suggested order of adoption is Sol and the Agents API first, with Codex Cloud where remote execution saves coordination time.3

Speed tiers, benchmarks, and pricing backlash

Several rumored items did ship, including an Ultrafast API tier and a new top-end Pro 500 plan, though Ultrafast access is gated.4 On Hacker News, one developer said the Decisions API had been overlooked and shared internal benchmarks. Running GPT-6 Luna in batches of 20 records gave about 1.6x the throughput of a model called Jev, at 1.2x the cost, with no accuracy loss on an 800-row dataset.1 Another commenter pushed back, pointing out that the API had only been announced, not released.1 The test still circulated, and Developers Digest cited it as something worth reproducing.4

The sharpest reaction was about pricing. One Hacker News commenter argued that the most important news was the subscription change. They said $200-plan users were losing half their usage, while a new $500 tier restored the old allowance plus a little more.1 That account comes from a single user and is not confirmed elsewhere. Still, alongside the confirmed Pro 500 tier,4 it suggests heavy users may face higher costs.

The read

DevDay 2026 shows OpenAI pulling more of the agent loop into its own platform: sessions, tools, execution environments, and background work.3 For developers, that shift comes with a real incentive in Sol's pricing. Dots shows the cost of moving fast. The idea is strong enough that a tester kept using it despite the bugs, but OpenAI's breadth is now outrunning its polish.2 The likely near-term outcome is that developers adopt Sol and the Agents API quickly, while Dots stays a preview of where OpenAI is going rather than a product to depend on today.