OpenAI pitches GPT-5.6 in AWS Kiro as the cheaper coding agent
OpenAI's Aug 24 note claims up to 82% lower cost per task in Kiro's spec-driven flow — and MCP's new roadmap swaps pasted API keys for real agent identity.
Copy markdown
82% cheaper per task — but not a new model
OpenAI's August 24 post says GPT-5.6 Terra ran successful tasks in AWS Kiro at “roughly 82% cost reduction” on Terminal-Bench 2.1, pitching more useful work per token for coding agents. The catch: Sol, Terra and Luna have been selectable in Kiro since July 14, so this is a price-performance push, not a fresh release.
Pick the tier: Luna runs at ~a quarter the cost
In Kiro the three tiers bill at 2.4x (Sol), 1.2x (Terra) and 0.6x (Luna) credit multipliers, each with a 272K context window. Kiro rates Luna above Claude Opus 4.8 on its Coding Agent Index (74.6 vs 72.5) at roughly a quarter the cost — worth a test if a coding agent is eating your credits.
MCP's new roadmap retires the pasted API key
The Model Context Protocol's Aug 22 roadmap prioritizes agent identity — DPoP and workload federation to replace long-lived API keys — plus Streamable-HTTP-over-stdio so local and remote servers share one transport, progressive tool discovery so huge catalogs stop degrading models, and webhooks instead of polling. If you ship MCP servers, plan around it now.