GPT-5.6 Luna: When the Cheapest Tier Is the Right Choice

Luna costs $1 input / $6 output per million tokens. That's one-fifth what Sol costs and less than half what Terra costs. And on the benchmark that arguably matters most for agent developers — Agents' Last Exam, which tests 55 professional domains across long-horizon workflows — Luna scores 50.3%. GPT-5.5 scores 46.9%.

So before defaulting to Terra because it's the "safe" middle tier: Luna outperforms your current production model at a fraction of the price. That's worth testing seriously before committing to a higher tier.

Luna's Actual Benchmark Performance

All scores sourced from OpenAI's official GPT-5.6 launch (July 9, 2026):

Luna exceeds GPT-5.5 on Agents' Last Exam, HealthBench Professional, and DeepSWE. It trails GPT-5.5 slightly on Coding Agent Index and BrowseComp, but by margins that most production use cases will not notice — while costing roughly one-quarter as much per token.

What Luna Is Optimized For

Luna is designed for tasks where speed and throughput matter more than peak accuracy. OpenAI built Luna to serve the high-volume, low-latency end of agent workflows: the calls that happen dozens or hundreds of times per session, where waiting for Sol-level reasoning would break the user experience.

Triage and routing

In a multi-agent architecture, a lightweight model often handles the initial classification of incoming requests — determining urgency, routing to the appropriate specialist agent, filtering duplicates. Luna is fast and cheap enough to run this layer continuously at scale without the cost adding up to Sol-level spend.

Monitoring and summarization

Long-running agent sessions generate a lot of intermediate output. Luna can monitor that output, generate running summaries, flag anomalies, and surface key results — acting as a cheap context-compression layer that keeps the primary agent's context window from overflowing.

High-volume content generation

For workflows that produce content at scale — hundreds of product descriptions, ad copy variants, email personalizations — Luna's cost profile changes what is economically viable. At $6 per million output tokens, producing 1,000 short-form outputs costs approximately $0.006 per output at typical lengths. Sol would cost approximately five times more.

Streaming responses with low latency

Luna is the fastest tier in the GPT-5.6 family. For applications where the user sees output as it is generated, Luna's response speed creates a noticeably better experience compared to the heavier reasoning overhead of Sol or even Terra.

Parallel agent coordination

In ultra mode, GPT-5.6 Sol coordinates four agents by default. If you are building a similar parallel architecture using the Responses API's multi-agent beta, Luna is the logical choice for the lower-level agents in the coordination hierarchy — fast, capable, and cheap enough to run in parallel without the cost compounding prohibitively.

Where Luna Falls Short

Luna has two clear limitations that should drive you toward Terra or Sol:

1. Long-context recall drops sharply. On OpenAI's MRCR v2 benchmark at 512K–1M context (8-needle retrieval), Luna scores 41.3% — compared to 72.5% for Terra and 73.8% for Sol. If your agent processes long documents, maintains extensive conversation history, or needs reliable recall across large inputs, Luna is the wrong choice.

2. Abstract reasoning gaps. On ARC-AGI-3, Luna scores 0.18% against Sol's 7.78%. For tasks that require genuinely novel reasoning rather than pattern application, Luna's lighter architecture shows.

Beyond those two areas, Luna's trade-offs are narrower than the price difference suggests.

Luna vs Terra: The Practical Decision

The gap between Luna and Terra is smaller than the gap between Terra and Sol on most benchmarks — yet the price difference is the same multiple. For many production workflows, the decision between Luna and Terra should be made empirically: run the same workload through both, compare output quality, and measure the actual difference before defaulting to Terra because it sits higher in the tier list.

Situations where Terra consistently wins:

• Tasks requiring reliable long-context recall (512K+ tokens)
• Workflows where output quality directly affects downstream user-facing products without human review
• Multi-step reasoning chains where small errors compound

Situations where Luna is the right default:

• Any task that previously ran on GPT-5.5 (Luna matches or exceeds GPT-5.5 on most benchmarks at lower cost)
• High-volume pipelines where output cost compounds across millions of tokens
• Latency-sensitive streaming applications
• Parallel agent coordination layers where the primary reasoning happens in a higher-tier model

Luna in a Mixed-Tier Agent Stack

The most practical architecture for complex agent workflows is not a single model choice — it is a tiered stack where each model tier handles the tasks it is best suited for.

A common pattern:

• Sol handles the hardest reasoning steps: initial problem decomposition, final synthesis, tasks where errors are expensive
• Terra handles the main execution path: the bulk of the reasoning, coding, and content generation
• Luna handles the volume work: monitoring, triage, summarization, routing, parallel coordination

AnyCap's capability layer sits across the full stack, providing image generation, web search, video production, and storage to whichever tier is running at each step. Luna calling AnyCap for a web search costs the same as Sol calling it — the capability cost is the same regardless of which model tier is doing the reasoning.

The Bottom Line on Luna

Luna is not a budget compromise. It is a purpose-built tier for the high-volume, low-latency, cost-sensitive layer of production agent workflows — and it starts from a benchmark baseline that exceeds GPT-5.5. For teams that spent the past year building on GPT-5.5, Luna deserves a direct evaluation before defaulting to Terra: in many cases, it delivers the same or better results at lower cost.

The use cases where Sol is clearly the better choice are real and important. But they are narrower than they appear at first glance.

All benchmark data sourced from OpenAI's official GPT-5.6 announcement (July 9, 2026). Pricing is per 1 million tokens at standard API rates.