The $1.33 Agent: A Field Guide to GPT-5.6 Small-Model Economics

$1.33. That’s what GPT-5.6 Luna spent to score 84.04% on BrowseComp, OpenAI’s benchmark for agents that dig through the web to answer difficult questions. GPT-5.5, the flagship until recently, scored 84.36% with reasoning cranked to maximum. The gap between those scores is rounding error. The gap between those bills is 25x.

On August 13, OpenAI published a builder’s guide for GPT-5.6, a collection of production lessons from startups already running the model family in live products. Most of it is standard API documentation. Buried in the benchmarks, though, is a shift that changes how you should budget any agent project: the expensive parts of agent workloads got dramatically cheaper, and the models responsible are the small ones.

The numbers first

Luna and Terra are GPT-5.6’s smaller models. Given enough test-time compute (letting the model think longer instead of making it bigger), they now rival GPT-5.4 and 5.5 on many long-horizon tasks, at a fraction of the price.

The guide’s own comparisons:

• BrowseComp. GPT-5.5 Extra High: 84.36% at $33.27 per run. GPT-5.6 Luna Extra High: 84.04% at $1.33. Same benchmark, nearly the same score, one-twenty-fifth the cost.
• Browser Use’s task suite. Browser Use, the team behind the open-source browser automation framework, ran Luna against 106 of their hardest web tasks. Luna completed 78% of them for about $14 total. A state-of-the-art model finished 80% for roughly $235.
• Extraction at Hypha. Luna kept 98% of GPT-5.5’s extraction accuracy at one-eighteenth the cost.
• Code exploration at PlayerZero. Switching the task to the smaller cut of the family cut inference costs 64% and response time 90%, while F1 actually improved by five points.

One more result deserves a closer look because it breaks a common assumption. On Agents’ Last Exam, GPT-5.6 Sol (the flagship) running at low reasoning effort outperformed GPT-5.5 running at high reasoning, with the same harness. Better output, less thinking. The newer generation is efficient enough that brute-force reasoning is no longer the default path to quality.

Read those together and the pattern is hard to miss: for a large share of agent work, the model you pick matters more than how hard you push any single model.

What this means for your stack

If you’re building agents and still sending every step to the flagship, you’re paying a 25x tax on the boring parts. Agent pipelines are mostly boring parts. Extraction, tagging, first-pass summarization, the hundred small tool calls between the two moments where judgment actually happens.

The routing pattern here is simple enough to implement in an afternoon:

• Route repetitive work to Terra or Luna. Anything with a clear right answer (extraction, classification, formatting) doesn’t need the flagship. Hypha’s numbers are the template: 98% of the accuracy, one-eighteenth the cost.
• Give browsing and long-horizon tasks to Luna with room to think. Luna Extra High is the configuration that matched GPT-5.5 on BrowseComp and completed 78% of Browser Use’s hardest tasks. Test-time compute is what closes the gap, so budget reasoning effort, not model size.
• Reserve Sol for judgment calls. Ambiguous decisions, or anything where a wrong answer is expensive. That’s a minority of steps in almost any pipeline. Pay flagship prices there and only there.
• Let small models try before the flagship decides. The Browser Use result points at another pattern: a cheap model completing 78% of tasks for $14 means you can run Luna first and escalate to the expensive model only on failure. At these prices, retrying with escalation costs less than starting big.

The plumbing matters too

Two of the guide’s less flashy sections are worth your time if you build on the Responses API.

Reasoning continuity lets a follow-up request pick up where the previous one stopped thinking, instead of paying for the reasoning again. For multi-step agents, that’s a direct cut on the most expensive line of the bill. Multi-agent orchestration and programmatic tool calling are new primitives trained end-to-end with the models; the orchestration one in particular matters if you’re implementing the routing pattern above, since it’s the mechanism for handing work between a big model and a small one.

Prompt caching also changed in ways that help requests land on warm inference engines. If you run agents with stable system prompts or repeated context, read that section before anything else in the guide. It’s free money for the exact workload agents generate.

What to do this week

1. Split your last month of token spend by task type. The repetitive categories (extraction, classification, formatting) are your candidates for Terra or Luna. Don’t estimate this from memory; pull the actual numbers.
2. Re-run one real task with Luna Extra High and compare cost and output. Use a task you already have ground truth for, so the comparison is honest.
3. Set up escalation, not replacement. Route to the small model first, escalate to Sol on failure or low confidence. You keep the quality floor and still capture most of the savings.
4. Check whether you can use reasoning continuity and the new prompt caching. Both cut costs without touching your prompts.

The benchmark race gets the headlines, and fair enough. But teams shipping agents in production are being sorted by a different metric now: cost per completed task. GPT-5.6 moved the frontier, sure. For your budget, the bigger news is that it moved the price of the work most of your pipeline actually does.

The guide is on OpenAI’s site. Worth reading in full, if only for the tables.