GPT-5.6 Luna logo

GPT-5.6 Luna

OpenAI · Released Jun 2026

pending

OpenAI's budget GPT-5.6 tier — TerminalBench 2.1 at 82.5% for $1/$6 per 1M tokens, undercutting Claude Haiku 4.5 and Gemini Flash on capability-per-dollar. The caveat: Luna is government-gated to ~20 trusted partners, with no independent verification or GA date. Best for teams building model-routing architectures who can escalate cheap failures to Terra or Sol.

Is it right for you?

Good for

  • Best capability-per-dollar in the GPT-5.6 family: TerminalBench 2.1 at 82.5% is only 0.9 points behind GPT-5.5 (83.4%) at 1/5 Sol's cost. For high-volume agentic workflows, Luna is the economics winner.
  • Speed and throughput: positioned as the fastest GPT-5.6 model, optimized for latency-sensitive products, agent swarms, customer support, and real-time classification pipelines.
  • Routine automation and extraction: classification, summarization, formatting, routing, draft generation, and background automation where cost-per-request and throughput dominate over raw reasoning depth.
  • Model-routing architectures: the bottom tier of a Sol > Terra > Luna escalation ladder. Use Luna by default, escalate to Terra or Sol when Luna's eval score drops below threshold on a specific task.

Not good for

  • Government-gated with no public access: same limited preview restriction. Luna's value proposition is clear on paper, but no independent verification exists.
  • CTF score of 85.19% is the lowest in the family — solid but not dominant. For security-critical workflows, Terra or Sol are the appropriate tier.
  • Lacks the safety depth of Sol and Terra: no activation classifiers on Luna (per system card). For regulated cyber and bio workflows, use Terra or Sol with full monitoring stack.
  • At $1/$6, Luna is competitive but not the cheapest model on the market — GLM-5.2 ($1.40/$4.40) and DeepSeek V4 Pro ($0.44/$0.87) offer lower output pricing for similar capability tiers.

How it performs by task

High-volume agentic coding (budget)

Very Good

82.5% on TerminalBench 2.1 is only 0.9 behind GPT-5.5. Remarkable for a budget tier. Route routine coding tasks to Luna, escalate complex ones.

Classification and extraction

Excellent

The ideal model for cheap, high-throughput tasks. Classification, entity extraction, sentiment, routing, and formatting at $1/$6 is the economics win of the GPT-5.6 family.

Summarization and drafting

Very Good

Fast and affordable for document summarization, email drafting, report generation, and content workflows where speed matters more than extreme precision.

Cybersecurity

Good

CTF 85.19% is solid for a budget model but trails Terra (91.84%) and Sol (96.7%). No activation classifiers. Not the right tier for security-critical work.

Hard reasoning and deep research

Fair

Budget tier with no max or ultra reasoning modes. For hard problems, route to Sol. For everyday reasoning, Terra. Luna is for volume, not depth.

Pricing

Input

$1 / 1M tokens

Output

$6 / 1M tokens

Context

Fastest and most cost-efficient GPT-5.6 tier.

View full pricing

Benchmarks

BenchmarkScoreSource
TerminalBench 2.182.5% Source
Internal CTF (capture-the-flag)85.19% Source

Verdict history

Jul 16, 2026
copy edit to brevity standard, detail preserved in children — Density rewrite: compressed 176w/9s to ~70w/3s per the 55–90w verdict_summary standard. Benchmark scores, pricing detail, audience fit, and per-task ratings remain in strengths[], benchmarks[], task_strengths[], and verdict_audiences[].

Verification log

  • Pricing— No changes

    Automated agent

  • Pricing— No changes

    Automated agent

  • Status— Updated

    Automated agent

    Summary re-dated to Jul 7 2026 and made explicit about what is awaited: stays pending pending GA + independent benchmarks; still government-gated (~20 orgs); family trending toward release (~Jul 9 GA per prediction markets, unconfirmed). Verdict unchanged (pending).

  • Pricing— No changes

    Automated agent

  • Pricing— No changes

    Automated agent

    Pricing unchanged: $1/$6 per 1M tokens. Government-gated. Confirmed by OpenAI blog.

  • Profile— No changes

    Imported at launch

  • Pricing— No changes

    Imported at launch

How we evaluate