TL;DR (July 2026): GPT-5.6's pricing is out, and the naming is now official (OpenAI's own preview post uses "Sol"): Sol at $5 / $30 per 1M input/output tokens - the same list price as GPT-5.5 - Terra at $2.50 / $15, and Luna at $1 / $6, a new cheap production tier. The community verdict is near-unanimous: Luna is the story, not Sol. A frontier-family model at $6 output undercuts the mid-tier of six months ago and lands in eval queues next to the open-weight Chinese workhorses. The two catches: access is still a limited, government-gated preview (roughly 20 vetted partners, US-only), so for most teams these are forward prices, not payable ones; and the skeptics have a receipt-backed worry - GPT-5.5's output price had already doubled from $15 to $30, so "Sol holds the line at $30" may just mean the next frontier bracket quietly moves to $60. Price the tier structure, not the press release.
When we covered GPT-5.6's government-gated rollout in late June, the pricing picture was still community speculation. It has since firmed up: OpenAI's own "Previewing GPT-5.6 Sol" post made the tier naming official, and consistent list prices are circulating for all three tiers. So this is the pricing piece: what each tier costs, what it replaces, and the one argument about this price sheet worth having.
The three tiers, priced
| Tier | Input $/1M | Output $/1M | Positioning | Nearest prior anchor |
|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | Frontier - agentic coding, cyber, hardest tasks | GPT-5.5 ($5 / $30) - same list price |
| GPT-5.6 Terra | $2.50 | $15.00 | Mid tier - the balanced default | GPT-5.5's launch output price ($15) |
| GPT-5.6 Luna | $1.00 | $6.00 | High-volume production workhorse | No direct predecessor - new bracket |
| GPT-5.4 Pro (for scale) | $30.00 | $180.00 | Premium reasoning | - |
List prices as circulating in early July 2026; the preview is limited, so treat them as launch pricing subject to change at broad release. Caching economics carry over and improve: cached input reads are 90% cheaper on GPT-5.6-era models, with cache writes billed at 1.25x the input rate - which, as we covered, is exactly the discount structure that breaks naive cost tracking if you meter total input tokens at list price.
Why Luna is the actual headline
The most-echoed community take on the announcement was some variant of: Sol is a nice improvement, but Luna is the significant one, because of the price. That is the right read, for a reason this site has been hammering all year: the workhorse tier is where the volume lives. Almost nobody's bill is dominated by their hardest queries; it is dominated by the classification, extraction, summarization, and routine-coding calls that run millions of times. A credible frontier-family model at $1 / $6 reprices exactly that traffic.
- Against OpenAI's own stack: Luna's $6 output is 60% cheaper than Terra and 80% cheaper than Sol/5.5. If Luna passes your evals for workhorse tasks, routing them down from a $30-output model is a bigger bill cut than any negotiated discount you will ever get.
- Against the cheap open-weight tier: Luna does not win on price. DeepSeek V4 Flash lists around $0.14 / $0.28 - roughly 7x cheaper on input and 20x on output than Luna. What Luna buys is the same vendor, SLA, and compliance envelope you already have, at a price close enough that the convenience argument stops being embarrassing. Whether that is worth 7-20x is a cost-per-task measurement, not a debate.
The skeptics' receipt: the quiet tier-up
The pushback thread worth reading makes a structural argument, not a benchmark one. The framing "Sol costs the same as 5.5" sounds like price discipline - until you remember how 5.5 got to $30: its output price doubled relative to the $15 bracket the previous flagship occupied. The skeptic's model is a ratchet: each generation, the "frontier" bracket quietly steps up ($15 → $30 → $60?), the marketing compares the new model favorably against the inflated bracket ("2.5x cheaper than Pro!"), and the tier names reshuffle so no like-for-like comparison survives two generations.
You do not have to fully buy the ratchet theory to act on it. The defensive posture is the same either way:
- Track your blended $/task over time, per workload. Tier names are marketing; your cost per completed task is accounting. If it drifts up across model generations while list prices "hold," the ratchet is real for you.
- Never anchor on the vendor's comparison. "Cheaper than Pro" is a comparison against the most expensive thing on the menu. Compare against what you actually ran last quarter.
- Keep the exit priced. The reason the ratchet works is switching inertia. A router with a benchmarked fallback - Luna to Flash, Sol to an open-weight frontier - turns a price hike into a routing decision instead of a budget line.
The asterisk: you (probably) cannot buy it yet
All of this is forward pricing. GPT-5.6 remains in the limited, US-only preview created by the government's rollout intervention - roughly 20 vetted partners, access approved customer by customer. For everyone else, the practical guidance is what it was in the gating piece: do not re-platform on a preview. Put Luna in the eval queue for the day broad access lands, next to DeepSeek V4 Flash and the other workhorse candidates, and let the numbers pick. Prediction-market money has the broad release landing this summer; your migration plan should not care whether it is July or September.
The honest take
This is, on its face, good news: the most capable model family on the market grew a $1 / $6 tier, and cached-read pricing keeps improving. Real per-token costs for workhorse tasks are falling on every axis that matters. The only way to lose here is to consume the price sheet as narrative instead of as data - to "upgrade" to whatever tier the launch post frames as the deal, without measuring what your actual tasks cost on it. Tier names changed three times in eighteen months; the metric never changes. Meter every call per model, compute cost per completed task, and route accordingly. Models are temporary. The meter is permanent.