Tech • AI • Robotics • Game
OpenAI unveils its “AI employees”: the era of autonomous agents has begun
9.4/10TL;DR
OpenAI is cutting model prices overall, but charging a steep premium for faster access to scarce computing capacity and for keeping more work inside its own agent-driven workspace.
KEY POINTS
Two products: workspace and compute
OpenAI is increasingly selling both the place where AI work happens and the machine time needed to run it. For companies, that creates two separate decisions: whether to let work, context and collaboration live inside ChatGPT, and how much to pay for the compute that powers agents, coding tools and overnight tasks.
A wide pricing ladder across models
The flagship GPT-6 Astra is positioned as the premium model. GPT-6.1 Sol is described as close to Astra on coding, computer use and professional tasks at about one-sixth of the price, while GPT-6 Luna is roughly 100 times cheaper than Astra. These gaps apply to token pricing, the units used to process and generate text.
Dots turn assistants into standing roles
Dots are permanent agents with their own cloud computers and access to user-approved applications. Rather than answering one-off prompts, they can accumulate working memory, retrieve documents, summarize client exchanges and keep a file moving while staff are away. That makes them less like a chatbot and more like a semi-persistent role inside an organization.
Memory and control remain sensitive issues
A first Dot is included in some plans, but users cannot inspect or edit each memory item individually. Disconnecting an app does not automatically erase what the agent has retained; clearing that memory requires deleting the Dot. For businesses, that raises a governance issue: client history, pricing logic, promises and decisions are safer when preserved in company-controlled documents that another tool can later reuse.
Guardrails reduce risk, not eliminate it
OpenAI offers action rules, activity logs and monitoring that can stop a Dot, and some functions are deliberately constrained, such as read-only background search. Even so, agents can still encounter malicious instructions inside documents and be diverted. Human review remains necessary for commitments such as prices, delivery dates and contract terms.
Space expands OpenAI’s role in office software
With Space, ChatGPT, Dots and humans can work together through shared pages, teams and recurring tasks. Integrations such as Figma and Photoshop inside ChatGPT push more of the workflow into OpenAI’s environment, reflecting a broader shift in which model-centered systems absorb functions once split across multiple software products.
A powerful distribution channel for software vendors
OpenAI says ChatGPT now reaches about 1.2 billion weekly users. Through Sign in with ChatGPT, paid subscribers can use partner apps with their existing subscriptions and spending caps, while enterprise buyers can route committed OpenAI budget through a Marketplace. That gives publishers access to distribution, but also makes customer access more dependent on OpenAI’s placement and recommendations.
Services firms face a new competitive threat
Internal tests for specialized enterprise Dots have covered purchasing, invoicing, marketing, support and contracts. When OpenAI helps define agent responsibilities during pilots, it moves beyond model supply and into parts of integration work. That weakens firms whose offer is limited to deploying an agent, while increasing the value of domain knowledge, client trust and quality assurance.
The premium is mainly for speed
The biggest surcharge is tied to latency, not raw intelligence. Ultra Fast, an accelerated version of Astra, is marketed as up to 8 times faster in Codex and costs about 6 times the standard Astra API rate and 30 times Sol. Pricing cited was $60 versus $2 per million input tokens, and $300 versus $10 for output. That premium makes sense when saved time unblocks teams or enables use cases that would otherwise be impractical.
Usage caps point to compute rationing
The Pro 500 plan includes Ultra Fast, while Pro 200 has reopened at $200 after a suspension caused by demand, but with lower allocation for new subscribers. Tasks launched by Dots in Codex also consume the user’s allowance. Combined with repeated resets of usage limits, these moves suggest a strategy for rationing scarce compute while converting product announcements into usage spikes.
European access remains uneven
A major limitation remains the exclusion of many European Pro users from GPTs, while Business, Premium and Enterprise tiers retain access. The result is an uneven rollout of agent features that may leave freelancers and smaller operators at a disadvantage relative to larger account holders.
CONCLUSION
The expensive part of OpenAI’s new lineup is not simply smarter models, but faster compute and tighter control over where AI-assisted work is performed. For businesses, the safest approach is to keep core records and client knowledge in their own systems while treating premium speed as worth paying for only when delay is more costly than computation.
Ask a question