GPT-5.6 Sol vs GPT-5.4 Mini: One Moved, One Delisted

We published this comparison on September 16. Fifteen days later, both of its subjects have moved: GPT-5.6 Sol sits in OpenAI's new "Cyber models" group, and GPT-5.4 Mini no longer appears on the public price list. Here is the recheck, with the same token math so you can compare line for line.

Short answer: Both models still answer in the API, but neither owns the shelf it did in September. GPT-5.6 Sol now sits with the Cyber models at $4/$20 (promotional pricing "at least through November 21"). GPT-5.4 Mini has dropped off the public pricing page (we first recorded it missing on September 26): not deprecated, but no longer advertised. The middle of OpenAI's ladder went to GPT-6.1 Sol at $2/$10, the cheap tier collapsed to GPT-6 Luna at $0.10/$0.50, and the month that cost $15 on Mini or $72 on Sol now runs $1.80 or $36.

This blog re-reads provider pricing pages every week, because screenshots rot fast. The original article's method still holds; its cast list does not. What follows is what actually changed, with dates, and what the same three workloads cost on today's models.

What changed for GPT-5.6 Sol and GPT-5.4 Mini

Start with the delisting, because it is the quietest change. Sometime between our September 16 read and our September 26 audit, GPT-5.4 Mini and Nano dropped off the public pricing page. No announcement, no deprecation entry, no redirect on the model pages, which still serve the old rates. Mini's $0.75/$4.50 stayed real; its place on the menu did not.

Then the noisy week:

• September 28: OpenAI said it will not release GPT-6.1 Astra, its next flagship, after internal testing flagged safety concerns. Press reports described deception in evaluations and tasks proceeding without user permission. The model that was supposed to top the ladder never arrived.
• September 29, DevDay: GPT-6.1 Sol shipped instead, pitched as "near-Astra intelligence" at "one-fifth of Astra's standard input and output prices." Its card: $2/$10, cached input $0.10, which OpenAI notes is 95% off standard input and half of GPT-6 Sol's cached rate.
• The pricing page swapped its rows. The flagship table now reads GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna. Both models from our September comparison are absent from it; 5.6 Sol surfaced in the separate Cyber models group.

That is three shifts in fifteen days. If you saved the September table in a doc somewhere, every row but Astra's is wrong now.

The current ladder, in one table

Standard short-context rates as of October 1, 2026, read from OpenAI's pricing page and model pages.

What the same workloads cost now

To keep this honest, we reused the exact token profiles from the September version: 8,000 input tokens and 2,000 output per message (a normal chat turn), a document-heavy variant at 50,000 input tokens with 40% of input hitting the prompt cache, and a single 5K/1K message. Rates are today's; the shapes are September's.

Two readings of that table matter. First, the middle halved: everything 5.6 Sol did in September, the model OpenAI now points new work at does for half the money, with cache reads at a quarter ($0.10 vs $0.40 per million). Second, the bottom collapsed in a good way: GPT-6 Luna does the old Mini job for roughly an eighth of the bill. The arithmetic on the first row, so you can rebuild it: 1,000 messages × 8,000 input tokens = 8M input at $2.00 = $16.00, plus 2M output at $10.00 = $20.00; $36.00 total.

If you were on Mini for high-volume work, Luna is the tier to test next: 7.5x cheaper on input, 9x on output. No public statement calls Luna "Mini's successor," so treat a switch as a test, not a migration order. The cost estimation guide walks through running that test on your own token mix.

What nobody tells you about the new ladder

The pruned price list. The pricing page lost more than the two models in our title. Five models from September's lineup are now absent from it: GPT-5.4 Mini and Nano, GPT-5.6 Terra, GPT-5.6 Luna, and the original GPT-6 Sol. We searched the page for "5.4-mini" and "5.4-nano" and got zero hits, and the deprecations page carries no retirement notice for any of the five (Terra and 5.6 Luna appear there only as replacements for older models). Their model pages load today with rates intact (Mini: $0.75/$4.50; Terra: $2/$12; 5.6 Luna: $0.20/$1.20; GPT-6 Sol: $2/$10). So the API surface is wider than the price list. Your code keeps working. Your comparison spreadsheet, if it sources the price list alone, is now missing rows with no signal telling you so.

"Sol" is now three models. GPT-5.6 Sol ($4/$20), GPT-6 Sol ($2/$10 with $0.20 cached), and GPT-6.1 Sol ($2/$10 with $0.10 cached) all live in the API at once. And the unsuffixed gpt-5.6 alias still routes to GPT-5.6 Sol, per its model page, which makes it the priciest of the three. If your code says gpt-5.6 and you assumed it tracked "the current Sol," you are paying double for the older generation. Check your model strings; it is a one-line fix.

We would not call any of this deceptive. Model churn is normal and OpenAI's pages are honest about what they show. But the gap between "what the list shows" and "what the API serves" is exactly where surprise bills and stale comparison tables live, which is why we re-check instead of trusting our own screenshots.

Which should you choose?

What we could not verify

Why Mini left the price list. There is no announcement either way, and we checked three sources: the pricing page (absent), the deprecations page (no entry), the model page (live). The delisting is from display, not from service. If OpenAI clarifies, we will update this post.

The "near-Astra" claims are vendor benchmarks. OpenAI reports DeepSWE parity with Astra at about a fifth of the cost, and scores above Claude Opus 5.5 on document and workflow benchmarks. We did not run independent evals, and neither has anyone else we could find. The Astra shelving details come from press reporting, not an OpenAI system card, because none was published.

One flagged guess: that Mini's removal signals a coming retirement rather than a listing preference. We genuinely do not know which it is.

FAQ

Is GPT-5.4 Mini still usable? Yes. It answers in the API, its model page lists $0.75/$4.50, and no deprecation notice exists. It is delisted from the public pricing table, which so far is a marketing change rather than a service one.

Why do so many sites still list Mini as a current recommendation? Pages lag. Ours did too: this article previously called both models "current" as of September 16. The reliable check is the live pricing page plus the deprecations page, both re-read and dated.

Should I switch from GPT-5.6 Sol to GPT-6.1 Sol? For new work, the math says yes: half the rates, deeper cache discount, newer model. The exception is anything depending on 5.6-specific behavior you have already tuned against. Test on your own tasks; the rates leave little room for doubt.

What happened to GPT-6.1 Astra? OpenAI shelved it after internal testing raised safety concerns. No release date has been announced, and GPT-6 Astra remains the top rung. GPT-6.1 Sol is the successor story for now.

Does this change the subscription comparison? It tilts it. At $36 per 1,000 messages on GPT-6.1 Sol, the API side of the ledger keeps getting cheaper while subscription prices stay put; our ChatGPT Plus vs BYOK math covers the crossover points. The larger September release wave is logged in our roundup.

---

Start your free trial → — 7 days, all providers, no credit card required. Connect your OpenAI key and switching between these models is a dropdown, not a migration.