benchrStat hero shows the flagship (Sol). Terra and Luna prices are in the table below. The context window is genuinely unpublished — third-party "1.5M" figures are unconfirmed and we don't repeat them as fact.
The three-row price table
| Model | Input | Output | Cached input (implied) |
|---|---|---|---|
| Sol (flagship) | $5.00 | $30.00 | ~$0.50 |
| Terra (balanced) | $2.50 | $15.00 | ~$0.25 |
| Luna (cheapest) | $1.00 | $6.00 | ~$0.10 |
Input and output figures are announced rates from OpenAI's preview post and help center. The cached-input column is implied, not published: OpenAI says cache reads get the standard 90% cached-input discount, so we've applied that to the uncached input rate to show roughly what a cache hit would cost. No explicit cached dollar figure exists yet. Crucially, none of these rows appears on the live developers.openai.com pricing page — re-read on June 28, 2026, it still lists GPT-5.5 and GPT-5.4 only.
Each tier against its real anchor
The cleanest way to sanity-check announced prices is to line each model up against an existing OpenAI model at a price you can verify on the live pricing page.
Sol = the GPT-5.5 sticker. Sol's $5/$30 is identical to GPT-5.5 at $5/$30. So the flagship of the new series carries no headline price increase over the current flagship — you'd be paying the same per token for whatever capability gain Sol delivers. For external reference at that tier, Claude Opus 4.8 is $5/$25 (cheaper output) and Gemini 3.1 Pro is $2/$12 (cheaper on both sides).
Terra = the GPT-5.4 sticker. Terra's $2.50/$15 matches GPT-5.4. OpenAI frames Terra as GPT-5.5-class capability at roughly half the cost — that's a claim about quality per dollar, and the price is simply the GPT-5.4 rate. If the capability claim holds, Terra is the interesting one: GPT-5.5-grade output at the GPT-5.4 price.
Luna = between GPT-5 and GPT-5 Mini. Luna's announced rate undercuts GPT-5 at $1.25/$10 and sits just above GPT-5 Mini at $0.25/$2 (the table above has Luna's own rate). It is not a budget model in the Mini sense; it's a cheaper-than-GPT-5 fast tier. For an external floor, DeepSeek V4-Pro is $0.435/$0.87 — far below Luna, the usual reminder that OpenAI's cheapest is not the market's cheapest.
The caching policy, and what's missing from it
OpenAI describes the same two-part caching it uses elsewhere: cache reads get the 90% cached-input discount, and cache writes bill at 1.25x the uncached input rate. Apply the read discount and Sol's cached input is roughly $0.50/1M, Terra's roughly $0.25, Luna's roughly $0.10 — the implied figures in the table. The gap is that OpenAI has not printed an explicit cached dollar rate for any of the three. That's fine for a back-of-envelope forecast and not fine for a contract. If your workload reuses a large system prompt across many calls, the cache discount is the single biggest lever on your bill — and right now you're estimating it, not reading it.
Cost scenarios
These are forecasts at the announced rates, not invoices — you can't run this workload yet. Take a steady agent at 30M input + 10M output per month. On Sol that's $150 + $300 = $450/month; with a 90% cache hit on the input (implied ~$0.50/1M) it falls to about $15 + $300 = $315/month, since output is untouched by caching and dominates the bill. The same volume on Terra is $75 + $150 = $225/month, and on Luna $30 + $60 = $90/month. The output rate is the lever across all three: at $30, $15 and $6 per 1M, the model you pick moves the bill far more than caching does on output-heavy work.
Use-case fit
Best for (when it ships): teams already standardized on the OpenAI stack who want the new series' reasoning modes; cost-sensitive workloads that can drop from a flagship to Terra at the GPT-5.4 price for GPT-5.5-class quality (OpenAI's claim, unverified by us); and budget agent traffic that wants a cheaper-than-GPT-5 fast tier in Luna.
Skip if: you need to provision capacity now — GPT-5.6 is a ~20-partner government-gated preview you can't buy into. Skip it too if you need contractually firm pricing or published context limits, because neither exists yet. Today, GPT-5.5 at the same $5/$30 sticker is the model you can actually call.
Decision checklist
Confirm the price is real before you commit budget: do not treat GPT-5.6 as procurable until its rows appear on developers.openai.com/api/docs/pricing and the models are callable on your account. As of June 28, 2026 they are not.
If Terra's "GPT-5.5-class at half the cost" claim matters to your economics, plan to benchmark it the day it goes GA — that claim is OpenAI's, attributed, and not yet independently verified. The price ($2.50/$15) is just the GPT-5.4 rate; the value depends entirely on whether the quality claim holds on your tasks.