OpenAI DevDay 2024: DOT AI Agents, New Pricing, Decisions API
OpenAI's DevDay announcements included new AI assistants, pricing tiers, and API functionalities, alongside some notable comparisons to existing technologies.
DOTs: Personal AI Agents
OpenAI introduced "DOTs," personal AI agents powered by GPT-6 Astra, available for $100 per month. These agents are designed to automate tasks and interact with digital environments. The concept originated from Claudebot, an AI personal assistant developed by Peter Steinberger, which gained rapid popularity on GitHub before he was hired by OpenAI.
In a keynote demonstration, Sam Altman showcased a DOT named Alfred, which successfully removed an old inventory API by tracing dependencies, updating integrations, running tests, and submitting pull requests. A live demo by Holly Lee, however, highlighted the challenges of real-time AI demonstrations.
An interesting side note is the domain ownership for "Dots." The domain Dots.com is owned by a defunct women's clothing brand, while Dot.com is owned by XAI and redirects to Grokbot's landing page, leading to some internet humor.
New Pricing and Models
OpenAI unveiled new subscription tiers and models:
- Pro 500 Plan: Priced at $500 per month, this plan offers 25 times the usage of ChatGPT Plus and access to "Ultraast," a new speed tier for Astra that processes 300 tokens per second.
- Ultraast API: For API users, Ultraast costs $60 per million tokens for input and $300 per million tokens for output.
- Reduced $200 Plan: The existing $200 per month plan saw its usage cut in half.
- GPT 6.1 Soul: Described as offering "near Astro Intelligence" at a fifth of the price, GPT 6.1 Soul costs $2 per million tokens for input and $10 per million tokens for output, matching the price of the recently launched GPT-6 Soul.
Decisions API
A new "Decisions API" was announced, designed to provide quick, specific answers from a predefined list. Instead of asking a large language model to generate a paragraph and then parsing JSON, users can provide a small "Luna" model with a fixed set of answers, and it will select one in milliseconds. This eliminates the need for parsing or retry loops, as the model cannot generate responses outside the given list. This functionality is similar to "Jev," a product from Typesafe AI, a company founded by an OpenAI researcher.
Other Announcements
OpenAI also revealed:
- Pages: A Google Docs clone.
- Space: A Notion clone.
- Slackbot: A Slack integration similar to Claude Tag, demonstrating extensive token usage.
- Sign in with ChatGPT: This feature allows users to log into third-party applications using their ChatGPT accounts, offloading token usage from the application provider to the user. It launched with 16 partners, including Devon, Notion, Vercel, and OpenClaw.
Fastino Labs: An Alternative for Decision Models
For developers who prefer to own the weights of their decision models, Fastino Labs offers an alternative to Jev. Fastino's Gliner 2.5 to Side is an open-weight model that runs locally and is up to eight times faster than Jev. It allows users to train their own Jev-like models and maintain ownership of the weights. The Gliner family of models has been open-source since 2024 and has over 50 million downloads.
Fastino also released a new model called Glide, which reportedly surpasses Jev in accuracy for intent routing, fact-checking, and hallucination checks. These models can be hosted anywhere, but Fastino's inference API provides a quick way to integrate them.
Takeaways
- OpenAI introduced DOTs, personal AI agents built on GPT‑6 Astra, priced at $100/month, capable of automating tasks and interacting with digital environments.
- In the keynote, Sam Altman demonstrated a DOT named Alfred that automatically removed an inventory API by tracing dependencies, updating code, running tests, and submitting a pull request.
- New subscription tiers include a $500 Pro 500 plan with 25× ChatGPT Plus usage and the Ultraast speed tier processing 300 tokens per second, while the existing $200 plan’s usage limits were halved.
- The Decisions API offers a lightweight “Luna” model that selects answers from a predefined list in milliseconds, eliminating parsing and retry loops, and is comparable to Typesafe AI’s Jev product.
- Fastino Labs released open‑weight models like Gliner 2.5 and Glide that run locally up to eight times faster than Jev and provide higher accuracy for intent routing, fact‑checking, and hallucination detection.
Frequently Asked Questions
What is the Decisions API and how does it differ from standard LLM responses?
The Decisions API is a lightweight service that uses a small “Luna” model to choose an answer from a predefined list in milliseconds, rather than generating free‑form text. Because the model can only output items on the list, it removes the need for JSON parsing, retry loops, and reduces latency compared with typical large language model completions.
How did the DOT named Alfred automate the removal of an inventory API during the DevDay demo?
Alfred, a DOT powered by GPT‑6 Astra, analyzed the existing inventory API, traced its dependencies, generated updated integration code, executed tests, and automatically created a pull request to delete the old API. The end‑to‑end automation demonstrated the agent’s ability to perform complex code‑base modifications without human intervention.
Who is Fireship on YouTube?
Fireship is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.