OpenAI released two new developer APIs this week that show how artificial intelligence companies are rethinking pricing as their products shift from single questions and answers toward long-running, autonomous agents.
On September 11, OpenAI launched an Agents API in public beta, according to The Decoder. The API lets developers build cloud-based agents that run for extended periods, execute code, process files, delegate tasks to sub-agents, and call tools in parallel, using the same infrastructure that powers OpenAI's Codex and ChatGPT products, with support for the Model Context Protocol, custom functions and built-in tools such as web search. Developers can host the agents on OpenAI's own sandboxes or through partners including Cloudflare, Vercel, and Oracle. There are no extra fees for the service and billing is based solely on token usage, according to The Decoder.
A day earlier, on September 10, OpenAI introduced GPT-Live-1, a speech API priced at $0.05 per minute that lets apps listen and talk at the same time, The Decoder reported. The company said the model cuts turn-taking latency to 0.8 seconds, down from 1.4 seconds in its previous real-time model, and named Yelp as an early adopter using the API for phone-based reservations. The Decoder also reported gains on OpenAI's own measures, with full-duplex interactivity at 80.1% against 45.4% for the earlier GPT-Realtime-2.1 model, and tool-calling accuracy at 87%, up from 60%.
Rivals are pricing around cached context, not raw tokens
The launches follow a similar move by Anthropic, which cut the price of cache reads for its Claude Fable 5.1 model by 75%, from $1.00 to $0.25 per million tokens, according to VentureBeat. Anthropic said the change lowers effective costs by roughly 25% for typical workloads and by as much as 45% for "highly agentic workloads," since agents repeatedly re-read the same system instructions, tool definitions, and conversation history rather than issuing one-off queries. Standard input and output token prices for Fable 5.1 were unchanged from Fable 5, at $10 and $50 per million tokens respectively.
Two different levers, one direction
Taken together, the three launches point to companies competing less on headline per-token rates and more on how usage is measured in the first place. OpenAI's Agents API keeps a single token-based meter running regardless of how long an agent operates, while its GPT-Live-1 charges by the minute for continuous voice sessions. Anthropic, meanwhile, is discounting the specific mechanic — cached context — that determines much of what an autonomous coding or research agent actually costs to run.
None of the three companies has published independent, audited comparisons of real-world task costs under the new pricing. Developers building agents that run for hours or days will be watching how usage bills, not just list prices, change as adoption grows across the industry in the coming months.