OpenAI's Dev Day Was a Disappointment

Summary

The hosts argue that OpenAI’s Dev Day felt disappointing despite its many announcements. The event introduced several useful features, but none seemed to meet the audience’s main expectation: a major model release that could compete with Anthropic’s recent models. They also criticize OpenAI’s timing and presentation, saying that bundling many medium-sized announcements together made it difficult for individual features to get attention.

T3 Code and agent workflows

• Orchestrator V2 lets an agent delegate work to other agents and models from within T3 Code. For example, one agent can ask several other models to review a change, with their threads appearing in the interface instead of disappearing into separate command-line processes.
• The hosts describe using these capabilities to create review and verification loops: one model writes code, another reviews it, and the first revises the work until they agree or the author is confident enough to proceed. They also use “full send” and “full stop” workflows to have agents validate changes, then either merge them or return with a clear reason they cannot be completed.
• T3 Code’s newer UI features include rendering model-generated HTML and charts in threads, using an embedded browser so agents can screenshot and inspect their work, and hiding active threads in a compact “working” section. The latter is especially useful for people running many threads at once.
• The hosts say T3 Code’s programmatic PR watcher required substantial engineering. It uses multiple polling strategies and filters to check for relevant changes without wasting API limits or waking agents unnecessarily.
• They discuss a broader shift toward using Markdown files and agent-accessible tools as lightweight software. An instruction file can direct an agent to run commands, analyze data, or produce a report, sometimes replacing a dedicated app or service. Examples include agents auditing their tool-call history and diagnosing where time is spent.
• Their main caution is that models are still poor at ambiguous product decisions and other judgment calls when there is no clear way to verify success. Models can be strong at long, measurable tasks but may choose the wrong direction and keep pursuing it.

Model and API economics

A large portion of the episode examines why AI subscriptions, API prices, and actual compute costs are difficult to compare.

• The hosts say OpenAI’s reduction in subscription value came just before Dev Day and damaged the reception of the announcements. They estimate that the $200 ChatGPT plan now provides far less API-equivalent usage than some competing subscriptions. Their cited estimates are about $2,000 per month of usage for that plan versus roughly $12,000 for an Anthropic plan, though these are the hosts’ calculations rather than official figures.
• They argue that price per token is not the same as the cost of completing a task. Token efficiency, cache reads, model quality, and the number of agent steps all affect the final cost.
• In their account, OpenAI made its newer 6.1 Soul model unusually inexpensive, especially for cached input, making it attractive for high-volume work. However, they consider it a less capable general coding partner than Opus 5.5, which they say makes better choices and performs better on many tasks.
• They describe Ultra Fast as useful for quick, interactive tasks but costly for sustained development work. One host says it cost hundreds of dollars to review just two pull requests and warns that the $500 plan is not a good upgrade for most users. For developers who need more sustained usage, they suggest an Anthropic subscription may be a better fit.
• The hosts distinguish models that can reliably execute a clearly defined task from those that can independently choose the right task or make sound product decisions. They praise smaller open-weight models for staying on task in bounded work, such as screening a large collection of pull requests, while reserving autonomous, ambiguous work for a small number of leading proprietary models.
• They also discuss long-running agent tasks and long context windows. In their experience, Anthropic models are better at keeping long projects on track. They suspect this reflects both stronger decision-making and differences in how models handle very large contexts. High-context usage can also become expensive, particularly when repeated cached reads are billed at higher rates.

OpenAI Dev Day features and product critique

• The hosts note that Dev Day included roughly 30 announcements, including Dots, Spaces, MCP-related features, Codex Cloud, and a new CLI. Their criticism is less that every feature is bad than that useful features were buried among the headlines and were difficult to discover in the product.
• Dots is presented as a proactive assistant, but the hosts find its single-threaded stream of messages overwhelming. They say a useful assistant should prioritize information and choose the right channel or moment, rather than sending a long series of updates in one place.
• Spaces is described as a shared workspace for documents and other artifacts, where agents can work alongside people and teams. The hosts see potential in the feature but criticize its low visibility.
• More broadly, they say the ChatGPT desktop app and Codex experience can feel like a collection of features without a clear organizing vision. They contrast that with T3 Code, where a smaller group can make product decisions quickly, while acknowledging that this approach can also lead to features being shipped with too little review.
• They suggest OpenAI’s Dev Day presentation may have suffered because many teams were releasing work at once and some features were not fully polished. They would have preferred a more staggered rollout that gave each feature time to be noticed and improved.

Other products and notes

• The hosts discuss Grok in a Tesla, mainly as a personal-use example of voice interaction and handling work while the car drives. The conversation is enthusiastic but does not offer a structured product review.
• The episode includes a PostHog sponsor segment describing product analytics, session replay, error tracking, logging, MCP access, a dashboard agent, Slack integration, and an automated “self-driving” feature that can create pull requests.

Reviews, guides, and tutorials

There is no formal review, guide, or step-by-step tutorial. The episode does share practical workflows, particularly agent-to-agent code review, verification loops, using Markdown instructions to automate tasks, and managing large numbers of agent threads.

Main speakers and sources

• Theo and Ben, the two Nerd Snipe hosts, provide the commentary and product evaluations.
• Julius and other T3 Code contributors are discussed in connection with Orchestrator V2 and T3 Code development.
• PostHog is the episode sponsor.