GPT-5.6 Sol, Terra & Luna Preview — Pricing, Tiers, and What Developers Should Know
OpenAI previewed GPT-5.6 on June 26, 2026 with three tiers — Sol (flagship), Terra (balanced), and Luna (fast). Official API pricing, caching rules, benchmarks, safety rollout, and how to prepare your stack.
Published 2026-06-27
On June 26, 2026, OpenAI (~500M monthly visits, SimilarWeb estimate) began a limited preview of the GPT-5.6 family: Sol (flagship agentic model), Terra (balanced everyday work), and Luna (fast, lowest cost). Official API list pricing is $5 / $30, $2.50 / $15, and $1 / $6 per 1M input/output tokens respectively, with 90% cached-input discounts and new 30-minute prompt cache rules. Sol adds max reasoning effort and an ultra multi-subagent mode. Access is initially restricted to trusted API/Codex partners before broader ChatGPT and API availability in the coming weeks.
If you need OpenAI-compatible access today while preview slots are limited, compare live routes on the APIMaster marketplace and verify what you actually receive with the free model tester.
What OpenAI Announced
OpenAI is introducing a new naming scheme with GPT-5.6:
- The number (5.6) marks the generation.
- Sol, Terra, and Luna are durable capability tiers that can evolve on their own release cadence.
During the preview, all three models are aimed at different cost/latency points on the same generation:
| Tier | Positioning | Input (per 1M tokens) | Cached input | Output (per 1M tokens) |
|---|---|---|---|---|
| Sol | Flagship for ambitious agentic work | $5.00 | $0.50 | $30.00 |
| Terra | Balanced model for everyday work | $2.50 | $0.25 | $15.00 |
| Luna | Fast, affordable high-volume work | $1.00 | $0.10 | $6.00 |
OpenAI states that Terra matches GPT-5.5-level performance at roughly half the cost, while Luna delivers strong capability at the lowest price point in the family.
Sol: Flagship Capabilities Developers Care About
GPT-5.6 Sol is positioned as OpenAI’s strongest model yet for long-horizon, tool-using workflows.
Highlights from the announcement:
- Terminal-Bench 2.1 — new state of the art on command-line agent tasks (planning, iteration, multi-tool coordination).
- GeneBench v1 — improved biology/genomics analysis vs GPT-5.5 with fewer tokens.
- Cybersecurity — competitive on ExploitBench² and ExploitGym³ with strong efficiency; OpenAI emphasizes defensive use cases (patching, review, education) over offensive automation.
maxreasoning effort — gives Sol more time to reason deeply on hard problems.ultramode — coordinates subagents for complex work beyond a single agent loop.
For production coding stacks (Codex, Cursor, custom agents), Sol is the tier to watch when your bottleneck is quality and multi-step reliability, not raw tokens per second.
Terra and Luna: When to Pick Which Tier
Use this decision frame:
| If your workload is… | Start with… | Why |
|---|---|---|
| Multi-file refactors, long agent runs, security research, deep reasoning | Sol | Highest capability ceiling; new max / ultra modes |
| Daily chat, RAG, moderate coding, most SaaS features | Terra | ~GPT-5.5 class at 2× lower official price vs Sol |
| Classification, extraction, high-QPS chat, bulk summarization | Luna | Lowest $/token; best for volume |
Rule of thumb: Luna for throughput, Terra for default production, Sol for hard agentic tasks where failure cost exceeds token cost.
Pricing Details That Change Your Bill
Beyond the headline $/1M rates, GPT-5.6 introduces more predictable prompt caching:
- Explicit cache breakpoints in the API.
- Minimum cache lifetime: 30 minutes on GPT-5.6 and later models.
- Cache writes billed at 1.25× uncached input rate.
- Cache reads keep the familiar 90% discount on cached input.
For agent loops that reuse system prompts, tool schemas, or long context prefixes, Terra and Luna can look much cheaper in practice than headline input pricing — but only if your integration actually hits the cache.
Example: on Sol, uncached input is $5/M; cached input is $0.50/M — a 10× reduction on repeated prefix tokens.
Preview Access: What “Limited” Means Today
OpenAI is not opening GPT-5.6 to everyone on day one.
Current preview constraints (as stated in the announcement):
- Initial access through API and Codex for a select group of trusted partners.
- Coordination with the U.S. government on a phased rollout tied to cyber-capability safeguards.
- Broader availability in ChatGPT, Codex, and the API planned for the coming weeks.
- Cerebras deployment of Sol at up to 750 tokens/sec targeted for July 2026, starting with limited customers.
OpenAI explicitly says it does not want permanent government gatekeeping on releases — the preview is framed as a short-term step while cyber EO frameworks mature.
Practical implication for developers: if your product roadmap assumes GPT-5.6 Sol this week, plan for fallback to GPT-5.5 / Terra-class routes until your provider confirms preview access.
Safety Stack: Why You May See Refusals or Delays
GPT-5.6 ships with OpenAI’s most layered safety stack to date, especially around cyber and biology dual-use requests.
Expect during preview:
- Hard refusals on prohibited cyber assistance, including jailbreak attempts.
- Paused generations when classifiers flag high-risk output for secondary review.
- Account-level review when patterns look like persistent misuse vs legitimate security work.
- Occasional false positives on legitimate code review, CTF prep, or defensive testing.
OpenAI reports 700k+ A100-equivalent GPU hours of automated red-teaming plus ongoing human expert testing. For enterprise buyers, they mention future privacy-preserving detection and customer-operated safety controls.
If you operate an AI gateway or resell model access, factor higher tail latency and non-deterministic blocks into SLAs during the preview window.
GPT-5.6 vs GPT-5.5: What Actually Changed
| Dimension | GPT-5.5 (prior gen) | GPT-5.6 family |
|---|---|---|
| Tier naming | Single generation label | Sol / Terra / Luna tiers within 5.6 |
| Cost structure | Prior list pricing | Terra ~half Sol; Luna lowest |
| Agent modes | Standard reasoning efforts | Sol adds max + ultra subagents |
| Caching | Prior behavior | 30-min min cache, explicit breakpoints, 1.25× write premium |
| Cyber evals | Strong | Sol/Terra/Luna all improve; Sol most capable |
| Release | GA paths | Phased preview with stronger safeguards |
How to Buy GPT-5.6 on APIMaster
OpenAI’s preview is invite-only for many developers. On APIMaster.ai, you can buy OpenAI-compatible API access the same way you already use GPT-5.5 — pay as you go from $1, no subscription, OpenAI SDK compatible.
Step-by-step
- Register — free account, email or OAuth.
- Top up from $1 — wallet in the console; supports PayPal, card, and other methods listed at checkout (often easier than binding a card directly on platform.openai.com).
- Open the marketplace — search for GPT-5.6, Sol, Terra, or Luna as routes go live. Compare live $/1M token prices across channels (often below official list during rollout).
- Create an API key in the console → copy once.
- Point your SDK at APIMaster’s base URL and the model slug shown on the marketplace card:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_APIMASTER_KEY",
base_url="https://apimaster.ai/v1",
)
resp = client.chat.completions.create(
model="gpt-5.6-terra", # use the exact slug from marketplace when listed
messages=[{"role": "user", "content": "Reply with exactly: ok"}],
max_tokens=16,
)
print(resp.choices[0].message.content)
- Verify before you scale — run the same key through the free model fingerprint tester. During preview scarcity, some relays silently downgrade model tier; APIMaster’s detection helps you confirm you’re getting what you paid for.
Tier picking on APIMaster: use Luna routes for bulk/low-risk calls, Terra as your default production model, and Sol when agent quality matters more than $/token.
Not listed yet? Top up anyway — Sol/Terra/Luna channels are added to the marketplace as upstream partners enable them; wallet balance carries over automatically.
Follow APIMaster for GPT-5.6 Launch Updates
GPT-5.6 is rolling out in phases. The fastest way to know when Sol / Terra / Luna hit the marketplace — and at what price — is to stay on APIMaster’s channels:
- Telegram group — model launch alerts, pricing changes, and support (t.me/apimasterai, ~community channel)
- Contact / community page — Telegram QR and feedback form
- Blog — deep dives like this one when major models ship
- Changelog — product and routing updates
When OpenAI widens preview access, we post which APIMaster channels pass fingerprint verification so you can switch base_url without rewriting your app.
Create your account → · Browse marketplace → · Join Telegram →
FAQ
What are GPT-5.6 Sol, Terra, and Luna? They are three capability tiers in OpenAI’s GPT-5.6 generation: Sol is the flagship agentic model, Terra is a balanced everyday tier (~GPT-5.5 class at lower cost), and Luna is the fast, lowest-cost tier for high-volume work.
How much does GPT-5.6 cost on the official API? List pricing per 1M tokens is Sol $5 input / $30 output, Terra $2.50 / $15, Luna $1 / $6, with cached input at 10% of uncached input price (90% discount). Cache writes cost 1.25× uncached input; cache entries last at least 30 minutes.
Is GPT-5.6 Sol available to everyone today? No. OpenAI started with a limited preview for trusted API/Codex partners in June 2026 and plans broader availability in the coming weeks. Check your OpenAI organization or provider dashboard for enablement.
What is GPT-5.6 Sol ultra mode?
ultra goes beyond a single agent by orchestrating subagents to accelerate complex tasks. It targets workloads where parallel exploration beats one long serial chain.
How does GPT-5.6 compare to GPT-5.5 for developers? Terra is positioned as competitive with GPT-5.5 at about half the official cost. Sol exceeds GPT-5.5 on agentic, biology, and cyber benchmarks OpenAI published. Luna trades peak capability for lowest $/token.
Will third-party API gateways have GPT-5.6 on day one? Usually not uniformly. Preview scarcity and provider-specific enablement mean aggregated gateways may roll out Sol/Terra/Luna at different times. Always verify the actual model with behavioral fingerprinting, not just the model string in the request.
Where can I read the primary source? OpenAI’s announcement: Previewing GPT-5.6 Sol (June 26, 2026). System card and expanded evals are promised at broader release.