Updated
Updated · The New Stack · Sep 7
OpenAI Says GPT-6 Astra Cuts Task Costs Despite 2.5x Higher Token Prices
Updated
Updated · The New Stack · Sep 7

OpenAI Says GPT-6 Astra Cuts Task Costs Despite 2.5x Higher Token Prices

3 articles · Updated · The New Stack · Sep 7

Summary

  • $10-per-million input and $50-per-million output pricing for GPT-6 Astra is 2.5 times GPT-5.6 Sol's rates, but OpenAI says many developers can still lower total run costs by using less reasoning and fewer tokens.
  • OpenAI's benchmarks showed Astra-low beating Sol-high on intelligence score, 49 to 48, with first-token latency of 2.53 seconds versus 11.87 seconds; on Terminal-Bench 4.0, Astra scored 57.9% versus 37.3% at about 9% lower cost per task.
  • Shinpr's codebase test found Astra-medium finished analysis, implementation and review in 51 minutes for $25.67, versus Sol-high's 75 minutes and $31.79, largely because Astra needed 80 requests instead of 238.
  • ARC Prize showed the opposite pattern on agentic workloads: Astra's max reasoning scored 62.7% on ARC-AGI-3 and cost $26,098, below low reasoning's $38,166, because stronger reasoning reduced failed actions and retries.
  • Astra's new configuration_update feature lets apps raise or lower reasoning between turns without resetting the request, reinforcing OpenAI's argument that per-token pricing alone does not predict real deployment cost.

Insights

Could OpenAI's massive price hike for GPT-6 Astra actually be a secret weapon for cutting your AI development costs?
Why did GPT-6 Astra's most expensive reasoning setting fail to catch a simple coding bug that its cheaper version easily spotted?
Will the shift from token pricing to cost per resolved task completely upend how businesses budget for AI in 2026?