September 22, 2026
Agent-powered research got cheaper: GPT-6 Astra cut Parallel's costs by 50%
On September 22, Parallel reported that GPT-6 Astra halved research time and code cost in its test. The agent gathered 6 labor-market metrics for 4 states across 6 months, navigated several sites, and consolidated the data into one report. It achieved the same quality with fewer research calls and tokens.
Parallel says GPT-6 Astra formulates search queries more accurately and can delegate parts of the research to sub-agents in parallel. A single research run now completes faster and costs half as much.
**The model is available through the API now.** In the Responses API, specify `model: "gpt-6-astra"` and choose `reasoning.effort` from `low` to `max`; tool calling works only through the Responses API. A million input tokens cost $10, output tokens cost $50, and cached input costs $1; Batch and Flex cut the standard rate by 50%.
GPT-6 Astra has a context window of 1 050 000 tokens and can generate up to 128 000 tokens per response. In Parallel's independent ranking, the model scored 70.3 on the Search Intelligence Score, spent $301 per 1 000 tasks, and took 34.7 seconds per task; GPT-5.6 Sol cost $217 there and took 52.9 seconds.
Requests longer than 272 000 input tokens are billed at a higher rate.
Source
