Same Score for Just $1.33! GPT‑5.6 Makes AI Agents a Steal

Hi, it's me, Shiichan! Today I've got a builder-focused guide from OpenAI News, all about making AI agents faster and cheaper.

OpenAI News openai.com

What was announced?

OpenAI News published "The builder's guide to GPT‑5.6." With the GPT‑5.6 model family making frontier-level agent performance dramatically more affordable, the guide walks through how startups are using smarter model selection and new API controls to build faster, more capable agents at a fraction of the cost.

The story so far

Historically, tackling long-horizon tasks meant reaching for the flagship model at the highest reasoning effort available. That was largely because only those heavyweight models were capable enough to handle long context and tool calling reliably.

What changes

That's changed with the 5.6 family. With more test-time compute, the smaller Luna and Terra models can now perform similarly to GPT‑5.4 and GPT‑5.5 while costing significantly less.

Take BrowseComp, a search benchmark that tests a model's ability to dig up obscure facts. Three months ago, GPT‑5.5 (Extra High) scored 84.36% at a total cost of $33.27. At launch, GPT‑5.6 Luna (Extra High) delivered essentially the same score — 84.04% — for just $1.33, and prices have dropped further since.

These smaller models are a great fit for high-volume workloads, latency-sensitive interactions, and repeated steps inside agentic workflows. A legal-tech startup parsing handwritten memos before agentic analysis, for example, doesn't need a frontier model for the whole pipeline anymore — it can hand the extraction step to Terra or Luna and save significantly on cost.

Dive Deep

Three new primitives landed in the Responses API to make agents run more efficiently:

• Reuse work already performed: reasoning can now persist across model turns, and native compaction compresses long-running conversations, so the model keeps coherence over long task horizons without reconstructing context from scratch
• Native multi-agent orchestration: multiple agents can run across parallel workstreams to finish complex tasks faster. A primary agent delegates work to subagents, the subagents pursue their objectives in parallel, and then hand results back to the primary agent for final synthesis. This is also how the "ultra" capability setting in ChatGPT works under the hood
• Programmatic tool calling: the model can write JavaScript itself to orchestrate tools, run independent calls in parallel, and process the outputs outside the context window — so work like fetching 100 filings, filtering by date, and spotting relevant transactions never has to pass through the model's reasoning at all, leaving it free to focus on judgment calls

Combined, the effect is dramatic: on ARC-AGI-3, GPT‑5.6 Sol scored 13.3% with the standard harness, but jumped to 38.3% after enabling retained reasoning and compaction — while using roughly 6x fewer output tokens. Same model, nearly triple the score.

Multi-agent behavior is also steerable: telling the model when it should spawn subagents helps make sure the extra token spend only happens when it actually pays off.

There's also a quieter win in prompt caching: cache TTL across the whole model family has been extended to a minimum of 30 minutes, and cache breakpoints can now be set deterministically within the context window. Keeping a consistent prompt_cache_key makes it more likely requests land on the same inference engine, cutting latency too.

Wrap-up

• OpenAI News published a builder's guide to getting the most out of GPT‑5.6 for agents
• Smaller models Luna and Terra now get close to flagship-level performance at a fraction of the cost (BrowseComp: $33.27 → $1.33 for nearly the same score)
• The Responses API gained persisted reasoning + native compaction, native multi-agent orchestration, and programmatic tool calling
• On ARC-AGI-3, retained reasoning and compaction alone took the score from 13.3% to 38.3%, using about 6x fewer output tokens
• Prompt cache TTL is now a minimum of 30 minutes across the family
• A solid read for developers and startups who want smarter, cheaper agents