Tool brief · July 1, 2026
GPT-5.6 Sol, Terra, Luna: which tier survives your next agent sprint
The tool
OpenAI GPT-5.6 (Sol, Terra, Luna preview)
What it is
GPT-5.6 is a three-tier model family in limited preview. OpenAI is beginning a limited preview of GPT-5.6 Sol, our flagship and most capable model, alongside GPT-5.6 Terra, a strong lower-cost option, and GPT-5.6 Luna, our fastest and most cost-efficient model. The pitch to developers: Sol is built for frontier reasoning and long-horizon agentic work; Terra is a balanced everyday model with GPT-5.5-competitive performance at 2x lower cost; and Luna is the fastest, most affordable member of the family.
Access is gated. OpenAI is starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly.
The next-work-session test
You're building a research-agent loop: plan → tool-call → critique → retry. Today it runs on GPT-5.5 and eats budget on the critique step. The question isn't "is Sol smarter?" — it's which node in your graph gets swapped.
Concrete swap: keep GPT-5.5 as the planner, route only the long-horizon critique/verify node to a 5.6 tier, and re-run your eval harness. If Terra is genuinely competitive with 5.5 at half the price, it becomes the default worker node and Sol stays reserved for the hard-stuck retries. That's the sprint: one router change, one eval diff, one cost delta.
Pricing
Verified from OpenAI's own preview page: GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. OpenAI also states 5.6 introduces more predictable prompt caching including explicit cache controls — worth benchmarking against your current cache-hit rate before you commit routing changes.
Context: VentureBeat notes that OpenAI's cheapest option is overall a mid-priced model, and still more expensive than the frontier-level GLM-5.2. So Luna is "affordable" only relative to the OpenAI ladder.
What we'd actually use it for
Narrower than the marketing. Two honest uses:
Terra as a 5.5 drop-in on the hot path. If OpenAI's "2x cheaper, competitive with 5.5" claim survives your evals, this is a config change with a real P&L line attached.
Sol on the retry lane only. Reserve it for tasks where your current agent gives up or loops. The OpenAI safety hub is candid: GPT-5.6 Sol and Terra can find vulnerabilities and pieces of exploits, but in cybersecurity testing they were unable to carry out autonomous, end-to-end attacks against hardened targets. Read that as: still a stronger link in a loop, still not a full agent by itself.
Luna is interesting for high-volume classifiers and routers inside the agent, but at $1/$6 it's not cheap enough to displace embeddings-based routing.
Limits
You may not be able to call it. If gpt-5.6-sol / -terra / -luna appear you're in the preview and can call them via the Responses API or Codex. If you get a model_not_found or permission error when you call one, your account isn't in the preview yet. Plan your sprint around GPT-5.5 as fallback.
Benchmarks are vendor-run. Third-party analysts have already flagged this: if detected cheating rates are unusually high on agentic task suites, any benchmark that runs in an interactive environment may overstate real-world capability until evaluation harnesses adapt. Don't route production traffic on Terminal-Bench numbers. Run your own harness.
Governance obligations may attach. VentureBeat notes even the cheaper Terra and Luna tiers may carry new governance obligations for companies using them in security, life sciences or other sensitive workflows. Read the preview system card before you point a customer-facing agent at Sol.
Broad availability is soft-dated. OpenAI plans to make GPT-5.6 Sol, Terra, and Luna generally available in the coming week — a plan, not a commitment. Don't build a customer demo that assumes GA on a fixed date.
Try it if
- You have preview access and a working eval harness — the whole point is the A/B on your own tasks.
- Your agent loop's cost is dominated by a single reasoning-heavy node you could isolate and route separately.
- You want the explicit prompt-cache controls for a system-prompt-heavy agent (worth its own benchmark run).
- You're already on the Responses API and Codex and can flip a model ID without refactoring.
Skip it if
- You don't have preview access — build against 5.5 and leave a model-ID env var. Migrating later is a one-line change.
- Your workload is embedding-heavy or classification-heavy; no 5.6 tier is priced for that.
- You operate in regulated domains and can't absorb whatever preview-partner governance obligations attach.
- Your current bottleneck is tool-call reliability or eval coverage, not raw model capability. A new tier won't fix a leaky agent loop.
Primary source: the OpenAI preview announcement and the help center overview.
Source: openai.com