OpenAI & AWS Tests Cut GPT-5.6 Agent Task Costs by 82%
新智元 · wechat · 2026-08-28
OpenAI and AWS released joint test results showing that by optimizing the environment and model, the cost for GPT-5.6 Terra to complete a task on the Kiro platform (Terminal-Bench 2.1) dropped by approximately 82%.
Optimization Logic
The 62% cost difference did not come from official price cuts (Terra dropped only 20%), but from efficiency gains:
- Reduced model Token output.
- Fewer tool calls.
- Avoided failed retries and circuitous paths.
Framework & Model Division
Tests show that billing costs are heavily influenced by the combination of "Agent Framework + Model". Kiro uses a Spec-driven process, breaking down requirements before handing them to the model, significantly saving tokens previously wasted on trial and error. Additionally, OpenAI suggests using Sol for planning and Luna for implementation in the pipeline to optimize cost-performance.
Competitive Landscape
The GPT-5.6 series (Sol/Terra/Luna) is now available on AWS Kiro, listed alongside Claude Fable. Kiro employs a hidden Chain of Thought strategy. This move allows OpenAI to compete with Anthropic on AWS's turf, shifting the evaluation standard from pure "accuracy" to a "balance of score and cost".
More from coding & agent
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- Eight non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15