OpenAI Releases GPT-5.6 Builder Guide: Smallest Model Costs Drop 25x
Nashnova编辑部
OpenAI published its GPT-5.6 builder's guide, showing the family's smallest model, Luna, nearly matches the flagship from three months ago — at 25× lower cost — signaling a rapid collapse in the price floor for AI reasoning.
Cost down 25×: what actually happened?
Luna is the smallest model in the GPT-5.6 family. On the BrowseComp benchmark it scored 84.04%, just 0.32 percentage points below GPT-5.5's 84.36% three months earlier.
Total spend for the same test dropped from $33.27 to $1.33 — roughly a 25× compression.
This means → tasks that required the flagship three months ago can now be handled by the smallest model at near-zero cost. In plain terms = the cargo that needed a semi-truck can now ride on an electric tricycle.
Sol's score jumped on a hard benchmark — how?
ARC-AGI-3 — a benchmark that drops an AI into unseen 2D game scenarios and asks it to figure out the rules on its own — remains unsolved above 50% by any top model.
GPT-5.6 Sol initially scored just 7.8% on the official leaderboard. The reason: the standard test harness wiped the model's private chain-of-thought after every step and truncated context starting from the oldest content.
Researchers Ilan Bigio and Ted Sanders switched to the Responses API and enabled "reasoning retention" plus "compression." Sol's score jumped to 38.3%, while output tokens fell 6×.
This means → the leap did not come from a smarter model. The old harness kept erasing its "scratch notes"; the new setup simply let existing capability through.
Three cost-cutting deployment methods — where does each save?
Programmatic tool calls: the model writes code to batch-execute tool calls in a sandbox instead of round-tripping one at a time. Financial-research firm Rogo reported 21% fewer input tokens with no quality loss.
Native multi-agent architecture: the lead model splits tasks across sub-models running in parallel. ChatGPT's Ultra tier defaults to four agents simultaneously. Obvious co-founder Jon Bell said they submit six briefs at once; the model writes, builds, and discusses in parallel — stable throughout.
Prompt caching: cache lifetime extended to at least 30 minutes, with custom breakpoint support. Engineers at AI firm Ploy set a breakpoint on a 29,000-token shared prompt and cut uncached input by 28%.
Ultrafast mode: 14× faster — but is the trade-off really gone?
OpenAI released an Ultrafast mode preview the same day. GPT-5.6 Sol can reach 750 tokens per second — up to 14× faster — now live on the API.
The industry long assumed a trade-off: real-time speed meant switching to a smaller, less capable model. Ultrafast claims to deliver high speed while maintaining Sol-level intelligence.
This reflects OpenAI's push to compress both cost and latency at once — but whether the trade-off is truly eliminated awaits large-scale developer testing.
Content is for reference only, not financial advice.