How GPT-5.6 fuses frontier intelligence with frontier efficiency

Source: OpenAI Date: 2026-07-29 URL: https://openai.com/index/gpt-5-6-frontier-intelligence-efficiency

Summary

An explainer post, not a new launch: GPT-5.6 (Sol / Terra / Luna) shipped July 9, and this piece walks the efficiency work behind it — ~20% lower serving cost from GPU kernel improvements, 15%+ token-generation gains from better speculative decoding, plus harness-level savings (context bloat, tool-call and repeated-work reduction). Terra matches GPT-5.5 intelligence at half the price; Luna is priced 80% below Sol.

Implications

Direct feed into the token-cost-as-operating-cost thread: this is OpenAI publishing its own cost-per-task math three weeks after launch, the same week Vibe’s cache-hit tracking, a token-saver skill reporting 95%+ token reuse, and Gemini Flash’s efficiency framing all point the same way — the frontier conversation has shifted from “can it?” to “what does it cost per task?” Also a closed-clock bookkeeping note: source URL 403’d on direct fetch (openai.com/index gate), and the date on this stub (07-29) is not the model’s actual launch date (07-09) — flag for any future dep-updates run that a “new” OpenAI index post is not automatically a new-model signal.