TL;DR: OpenAI previewed GPT-5.6 on June 26, 2026 as three tiers — Sol ($5/$30 per million tokens), Terra ($2.50/$15), and Luna ($1/$6). Terra is the interesting one: GPT-5.5-level quality at half the price. But all three are gated to ~20 partner orgs by a government directive, so you can’t wire them into Cursor or Codex today.

	Sol	Terra	Luna
Role	Frontier flagship	Balanced daily driver	Fast + cheap
Price / 1M tokens	$5 in / $30 out	$2.50 in / $15 out	$1 in / $6 out
Best for	Hard agentic + security work	Most Cursor/Codex sessions	Autocomplete, cheap edits
The catch	Same sticker as GPT-5.5, marginal gain	Can’t use it yet (gated)	Unproven on real coding

Honest take: Terra is the tier that changes your bill — GPT-5.5 quality at half the token cost. But it’s locked behind a federal preview, so don’t restructure anything yet. Keep your flat-rate Cursor Pro or Codex-via-Plus setup and revisit when GA lands.

What OpenAI actually shipped on June 26

GPT-5.6 is a naming reset as much as a model release. The number (5.6) marks the generation. The names — Sol, Terra, Luna — mark durable capability tiers that OpenAI says will each advance on their own cadence. Instead of picking gpt-5.6-mini versus gpt-5.6, you pick a tier by what the task needs: raw intelligence, balance, or speed and cost.

Verified against OpenAI’s preview announcement and help-center pricing page (June 26, 2026), the three tiers are priced per million tokens:

Tier	Input / 1M	Output / 1M	Positioning (OpenAI’s words)
Sol	$5.00	$30.00	Most demanding reasoning, coding, agentic, and security tasks
Terra	$2.50	$15.00	Competitive with GPT-5.5, roughly 2× cheaper
Luna	$1.00	$6.00	Strong capability at OpenAI’s lowest cost

Sol also adds two reasoning controls: a max reasoning effort setting for the hardest problems, and an Ultra Mode that spawns subagents — OpenAI’s own take on parallel agentic work, similar in spirit to Claude Code’s dynamic workflows.

The part the press releases bury: you can’t use it

Here’s the catch that reframes the entire release. GPT-5.6 is a limited preview, available through the API and Codex to roughly 20 trusted partner organizations — not to the general public, not in ChatGPT, and not in your Cursor model picker.

The reason is regulatory, not technical. A June 2, 2026 executive order directs federal agencies to benchmark and assess frontier AI models for safety before broad release. At the U.S. government’s request, OpenAI is staggering GPT-5.6 out through a small partner set first, with general availability planned for “the coming weeks.”

If that pattern sounds familiar, it should. Anthropic’s Fable 5 and Mythos 5 were pulled entirely under a separate export-control directive in mid-June — the subject of our government-ban resilience piece. Two frontier labs, two government interventions, three weeks apart. The lesson for anyone building on cloud models is the same: the model you architect around can vanish or fail to arrive on someone else’s schedule.

So this article is not “how to switch Cursor to Sol today.” You can’t. It’s about what the pricing tells you, what to plan for, and what to do in the meantime.

What Terra means for your bill

Start with the tier that matters. Terra at $2.50/$15 is the story. OpenAI positions it as quality-competitive with GPT-5.5 while costing roughly half as much.

That claim checks out on the numbers. GPT-5.5, verified today, runs $5.00 input / $30.00 output per million tokens at standard rates (with a cached-input rate of $0.50/M and a 1.05M-token context window). Terra’s $2.50/$15 is exactly half the sticker — and, not coincidentally, the same per-token rate you’d get from GPT-5.5 on Batch or Flex processing. The difference is you get that price on the interactive, low-latency path.

Put it in dollars against a realistic coding session. A typical agentic Cursor or Codex run — say 50K input tokens (files, context, tool results) and 8K output tokens (edits, explanations) — costs:

Model	Input cost	Output cost	Per session	~500 sessions/mo
GPT-5.5	$0.25	$0.24	$0.49	~$245
GPT-5.6 Terra	$0.125	$0.12	$0.245	~$123
GPT-5.6 Sol	$0.25	$0.24	$0.49	~$245
GPT-5.6 Luna	$0.05	$0.048	$0.098	~$49

If Terra genuinely holds GPT-5.5 quality, a heavy API user cuts their model spend in half by switching one string. That’s a bigger practical win than Sol’s benchmark bragging rights, because most day-to-day coding doesn’t need frontier reasoning — it needs “good enough, fast, and cheap.”

Sol: a marginal gain at flagship price

Sol is priced identically to GPT-5.5 ($5/$30). So the only reason to move from GPT-5.5 to Sol is if Sol is meaningfully better — and here the honesty gets harder, because the benchmarks are preview claims you can’t reproduce.

OpenAI’s reported numbers put single-agent Sol at 88.8% on Terminal-Bench 2.1, with Sol Ultra (subagents) at 91.9% — the latter a claimed new state of the art. For context, our Terminal-Bench 2.1 leaderboard analysis put the top usable models in the low-to-mid 80s (GPT-5.5/Codex CLI at 83.4%, Opus 4.8/Claude Code at 82.7%), with the Anthropic Mythos-class model that briefly topped 88% then getting suspended.

Two caveats keep this from being a buy signal:

1. You can’t verify it. The model is gated to 20 orgs. No independent harness has re-run these numbers on a uniform setup, and native-harness scores routinely run 5–7 points above uniform-harness scores. Treat 88.8% and 91.9% as vendor preview figures until tbench.ai confirms them.
2. A SWE-Bench Pro score for Sol hasn’t been published. Terminal-Bench rewards sequential shell tasks; SWE-Bench rewards one-shot repository fixes. A tool can lead one and trail the other.

At equal price to GPT-5.5, a few benchmark points — on a benchmark you can’t reproduce — is not a reason to re-plumb your stack.

The Codex gotcha most people miss

Here’s a real trap worth its own section, because it trips up anyone reasoning about “GPT-5.6 pricing for Codex.”

Codex CLI billing and API billing are two different meters. If you run the Codex CLI signed in with a ChatGPT Plus, Pro, or Business subscription — the setup covered in our Codex CLI review — you are not paying per token. You’re drawing on your plan’s included allowance. The $2.50/$15 Terra rate is an API price. It hits you only if Codex (or Cursor) is calling the model through an API key you pay for directly.

So when GA arrives, the practical question splits:

• You use Codex via a ChatGPT plan → the token price barely matters directly. What matters is whether OpenAI raises plan prices or tightens allowances when GPT-5.6 becomes the default backend. Watch the plan terms, not the per-token table.
• You use Cursor or Codex with your own API key (BYOK) → the token price is your bill, dollar for dollar. Terra’s 2× cut is real money for you.

Mixing these up is how people either over-worry about pricing that doesn’t touch them or under-budget for the meter that does. Check which path you’re on before you do any cost math. In Cursor, that’s Settings → Models (managed vs. your own key); in Codex, it’s whether you signed in with a subscription or exported an OPENAI_API_KEY.

Where GPT-5.6 sits against the rest of the market

Pricing only means something in context. Here’s Terra and Luna against the coding backends developers actually reach for in mid-2026, by API sticker (input/output per 1M tokens):

Model	Input / 1M	Output / 1M	Notes
GPT-5.6 Sol	$5.00	$30.00	Preview only; = GPT-5.5 price
GPT-5.6 Terra	$2.50	$15.00	Preview only; ~GPT-5.5 quality, half price
GPT-5.6 Luna	$1.00	$6.00	Preview only; unproven on coding
GPT-5.5	$5.00	$30.00	GA, powers Codex + Copilot today
Claude Fable 5	$10.00	$50.00	Suspended under export directive
GLM-5.2	~$1.40	~$4.40	Open-weight, MIT, self-hostable
Codestral 2	$0.30	(low)	Apache 2.0, self-hostable

Terra lands in the same neighborhood as GLM-5.2 on input cost but with far pricier output. If your workloads are output-heavy (lots of generated code and explanation), an open-weight model like GLM-5.2 as a Cursor/Cline backend or Codestral 2 stays cheaper — and can’t be pulled by a government order, because it runs on your hardware. That resilience angle is exactly why the OpenCode + Ollama local stack keeps earning its place as a fallback.

What to actually do right now

You have three sane moves while GPT-5.6 sits behind the preview wall:

Stay on flat-rate for interactive work. Cursor Pro ($20/mo, unlimited Auto) and Codex-via-ChatGPT-Plus insulate you from per-token pricing entirely. When Terra becomes the default, flat-rate plans absorb it — you get the upgrade without touching a pricing table. This is the same logic from our Copilot billing-change analysis: metered pricing can move against you, flat rates can’t until renewal.

If you’re BYOK, pre-write the Terra swap. The day GA lands, the change is one model string. Keep your Cursor OpenAI-compatible config and any Codex profile ready so you can A/B Terra against GPT-5.5 on your own tasks within an hour — not restructure your workflow from scratch.

Keep a model-agnostic escape hatch. Two government interventions in three weeks is a pattern, not a fluke. A local backend you fully control — an open-weight model on your own GPU — is the only setup no directive can disable. If you don’t have one wired up, our sister site’s local GPU and model guides at runaihome.com walk through the hardware, and aifoss.dev tracks the fully open-source tooling.

The pricing is genuinely good news — Terra at half of GPT-5.5 is the kind of cut that moves budgets. Just don’t rebuild anything around a model you can’t run yet.

FAQ

Can I use GPT-5.6 in Cursor or Codex today? No. As of July 2, 2026 it’s a limited preview available through the API and Codex to roughly 20 partner organizations, gated by a federal directive. General availability is planned for “the coming weeks,” but no date is confirmed.

Which tier should I use for coding when it’s available? Terra for the vast majority of work — it’s positioned at GPT-5.5 quality for half the token cost. Reserve Sol for genuinely hard agentic or security tasks where its extra reasoning earns its flagship price. Luna is worth testing for autocomplete and cheap edits, but its coding quality is unproven.

Is Sol actually better than GPT-5.5? On OpenAI’s preview benchmarks (88.8% single-agent, 91.9% Ultra on Terminal-Bench 2.1) it edges ahead, but those numbers are vendor claims that no independent harness has reproduced, and there’s no published SWE-Bench Pro score yet. At the same $5/$30 price, that’s not enough to switch on.

Does the $2.50/$15 Terra price affect my Codex CLI bill? Only if you run Codex with your own API key. If you’re signed in with a ChatGPT Plus, Pro, or Business plan, you draw on your plan’s allowance and per-token API prices don’t apply directly — watch your plan’s terms instead.

Why is GPT-5.6 gated when GPT-5.5 wasn’t? A June 2, 2026 executive order directs federal agencies to benchmark frontier models for safety before broad release. OpenAI is complying by starting with a small approved-partner preview. It mirrors the export-control action that pulled Anthropic’s Fable 5 and Mythos 5 in mid-June.

Sources

• Previewing GPT-5.6 Sol: a next-generation model — OpenAI
• A preview of GPT-5.6 Sol, Terra, and Luna — OpenAI Help Center
• OpenAI Previews GPT-5.6 With Sol, Terra, and Luna — MarkTechPost
• OpenAI unveils GPT-5.6 Sol, Terra and Luna — limited preview per US Gov — VentureBeat
• GPT-5.6 Sol, Terra, and Luna: OpenAI’s Next-Gen Model Family — DataCamp
• OpenAI API Pricing 2026: GPT-5.5 per-token table — Morph
• OpenAI API Pricing — official developer docs

Last updated July 2, 2026. GPT-5.6 is in limited preview and pricing/availability may change at general release; verify current state on OpenAI’s official pages before purchasing.