OpenAI's DevDay dumped 25 launches in one day

OpenAI's DevDay was on September 29th. I spent the week watching people comment on one launch at a time, each post about something different, and at no point was it clear how many things had actually been announced. So I went to OpenAI's own recap page and counted.

There are 25 announcements. In one day.

What actually matters if you write code

Setting aside the chat and collaboration products (Space, Pages, collaborative slides, profiles, marketplace, ChatGPT in Slack and Teams), this is what's left:

• GPT-6.1 Sol: new model, focused on agentic coding and computer use with fewer tokens per task.
• Ultrafast: a speed tier, up to 8× faster generation in Codex and 6× in the API.
• Codex in the cloud: remote development environments, reachable from any device.
• A refreshed Codex CLI, with voice-steered tasks and a better /agents view.
• Code Review in the desktop app, for reviewing GitHub and GitLab changes inside ChatGPT.
• Codex Security Cloud: repository scanning and infrastructure hardening.
• Agents API with Computer use and the Decisions API (real-time decisions over questions with predefined answers).
• MCP events for plugin automations.
• Bedrock Managed Agents, if you're on AWS.

That's already nine items that require reading documentation. The other sixteen compete for the same attention on the same day.

The math on "one fifth the price"

The claim that travelled furthest was about GPT-6.1 Sol: near-Astra intelligence at one fifth the price. I went to check it on the model's page.

On the price list it holds up exactly: Sol costs $2 input / $0.10 cached / $10 output per million tokens, against $10 / $1 / $50 for GPT-6 Astra. Five times, on every axis.

Except list price isn't what you pay. On that same page, OpenAI publishes average cost per task on Terminal-Bench Science 0.1: $5.47 for Sol against $23.80 for Astra (and $23.21 for Opus 5.5). Divide it out: 4.35×, not 5×. The gap is the cheaper model spending more tokens to get to the same place. It's the same input/output asymmetry that already showed up here when I wrote about indentation and tokens.

Second detail: "matches Astra" applies to one benchmark. DeepSWE v1.1 is the only place OpenAI writes matches. On OSWorld 2.0, Sol lands 2.1 points behind; on GDP.pdf it "approaches"; and on AutomationBench the published comparison is against Opus 5.5 (+2.2 points), not against Astra.

My verdict

Nothing here is a lie. The 1/5 is real as a list price, the 4.35× is real as cost per task, and the model genuinely is much cheaper than Astra.

The problem is the format. Twenty-five announcements on the same day guarantee that none of them gets evaluated: it becomes a list you scroll, not a set of tools you test. It took me twenty minutes to check two claims on one page; nobody is doing that twenty-five times. And when the headline is the only reading anyone has time for, the headline is what sticks.

If you only look at one thing on that list, look at cost per task for the model you already use. That's the number that shows up on your bill.