ERNIE 5.5 Flash Preview vs GPT-5.6 Luna: The Rate Card It Is Measured Against Is Already Retired

ERNIE 5.5 Flash Preview and GPT-5.6 Luna are being named in the same breath this week, and the awkward part is the calendar. ERNIE 5.5 Flash Preview went live as a billed endpoint on Baidu's Qianfan platform on 10 November 2026, with the launch post following a day later. GPT-5.6 Luna shipped on 9 July 2026, and it has been flagged as deprecated on the Artificial Analysis tracker since 22 September 2026, when GPT-6 Luna arrived at roughly half the price. So the comparison a reader actually needs is not the one the two names suggest. It is ERNIE 5.5 Flash Preview against a rate card that its own vendor has already moved past — and, second, against the successor that replaced it. The successor is priced at $0.10 and $0.50 with the same measured index and sits on the same key as the model it replaced.

That does not make the head-to-head pointless. It makes it a different piece of work. There is a real price difference, a real context difference, and a genuine question about which of these two you would put in front of an unattended task. What there is not, on either side, is a shared scoreboard. Every number below is either Baidu's own run, Artificial Analysis' own run, or a published rate, and none of them can be subtracted from the other.

What happened to GPT-5.6 Luna, and why it matters here

GPT-5.6 Luna is a 1M-context model priced at $0.20 per million input tokens and $1.20 per million output. On Artificial Analysis it carries an Intelligence Index of 23.8, and the same record carries a deprecated flag with the supersession date of 22 September 2026. Its replacement, GPT-6 Luna, is listed at $0.10 input and $0.50 output on the same 1M context and scores 23.5 at xhigh effort and 18.5 at medium on the same index — statistically level with the model it replaced, at less than half the output rate.

Two things follow, and both cut against the obvious framing of this page. First, a deprecated SKU is not a dead one — it still answers requests, and plenty of code still points at it — but it is a card nobody is defending on price any more. Second, if your reason for comparing ERNIE 5.5 Flash Preview to a Luna is that Luna is the cheap fast tier you already know, the model you actually know is now GPT-6 Luna. Holding a two-day-old preview against a retired card flatters the preview on price and tells you nothing about the bill you would pay next quarter.

There is a third thing, less flattering to the comparison and worth stating early: GPT-5.6 Luna has an independent score. ERNIE 5.5 Flash Preview does not. As of 11 November there is no ERNIE 5.5 row anywhere on the Artificial Analysis leaderboard, so every ERNIE figure in this piece is Baidu's, run on Baidu's harnesses, judged by models Baidu chose.

Six dimensions, both sides on the same line

Read these as a map of where the two disagree, not as a tally. The price and context rows are hard numbers from published cards. The score rows are not the same kind of number as each other, which is the subject of the next section.

• Price — ERNIE 5.5 Flash Preview at $0.28 in and $1.13 out per million tokens vs GPT-5.6 Luna at $0.20 in and $1.20 out. ERNIE's input is 40% dearer; its output is about 6% cheaper.

• Successor price — ERNIE has no successor and one card vs GPT-6 Luna at $0.10 in and $0.50 out, which is under half ERNIE's output rate and a fifth of its input rate.

• Context window — 1,024,000 tokens vs 1,000,000. A rounding difference; neither is the deciding factor.

• Output ceiling — 921,600 tokens in a single call vs a 1M context with no comparable single-call output ceiling published in the same form. This is the one spec where ERNIE is genuinely in a different class.

• Independent score — no Artificial Analysis row vs an Intelligence Index of 23.8. One side of this line has a third party behind it and the other does not.

• Measured cost per task — unpublished by Baidu vs $0.33 per task on Artificial Analysis' suite at an average of 522 seconds per task, falling to $0.15 per task on GPT-6 Luna.

• Status — a preview Baidu calls an intermediate checkpoint from post-training vs a SKU marked deprecated with a named replacement already shipping.

Why the two headline numbers cannot be subtracted

The number people will want to put side by side is Terminal-Bench 4.0: 46.5 for ERNIE 5.5 Flash Preview, from Baidu's appendix, against 11.6 for GPT-5.6 Luna, from Artificial Analysis' own evaluation. A 35-point gap looks decisive. It is not a gap at all, because the two figures come from different harnesses, different scaffolding and different judges, and only one of them was produced by someone with no stake in the result.

Baidu is unusually forthcoming about its bench setup, which is what makes the mismatch legible. Its footnotes name the harnesses — Claude Code 2.1.x, mini-swe-agent 2.4.6, Codex 0.162.1 and something abbreviated DSH — and the judge models, Qwen3.8-Max, DeepSeek-V4.1-Flash-1017 and Gemini 3.1 Flash-Lite. One footnote concedes that ERNIE 5.5 Flash Preview and GLM-5.3 lack native multimodal capability, and that image understanding during evaluation was provided through a tool powered by Claude Opus 5.5. So at least some vision-dependent rows measure a system in which the model under test calls a different vendor's model to look at things. That is a defensible engineering choice. It is also a reason those rows are not a like-for-like score.

The omniscience pair has the same problem in a subtler form. Baidu reports AA-Omniscience-Public at 25.2; Artificial Analysis records −17.1 for GPT-5.6 Luna on its Omniscience index. The labels differ, one number is a vendor run and the other is not, and the metric penalises confident wrong answers, which is exactly the behaviour a self-reported evaluation is least likely to surface. Treat these as two unrelated data points, not as a 42-point spread.

The one comparison that is genuinely fair is the internal one: GPT-5.6 Luna at 23.8 versus GPT-6 Luna at 23.5, both measured the same way by the same people. That pair is how you can tell the successor is a price move rather than a capability move — and it is the pair that tells you what ERNIE 5.5 Flash Preview is really being asked to beat.

What each side actually wins

ERNIE 5.5 Flash Preview wins on the ceiling and on freshness. A 921,600-token output allowance against a 1,024,000-token window means a single call can emit nearly as much as it can hold, which is the mechanical precondition for the eleven-hour Unreal Engine build Baidu put at the centre of its launch post. Nothing on GPT-5.6 Luna's card competes with that, and nothing on GPT-6 Luna's does either. ERNIE also wins the only contest that is not about numbers: it is the model with something to prove this month, and Baidu has promised Apache 2.0 weights, a date it has not given.

GPT-5.6 Luna wins on verification and on input price. There is a third-party score you can check, a published cost per completed task, a measured time per task, and a 40% cheaper input rate. Those are the things that make a model safe to put behind a product, and ERNIE 5.5 Flash Preview currently has none of them — no model card, no knowledge cutoff, no latency figure and no independent row.

GPT-6 Luna wins on price outright, and that is the finding that reframes the page. If the question is "which cheap fast model should I call", the answer inside the Luna family is not GPT-5.6 Luna; it is the successor at $0.10 and $0.50 with the same measured index and a cost per task under half its predecessor's. A comparison that stops at GPT-5.6 Luna is comparing ERNIE against the wrong half of the family.

The decision, stated without hedging

If you need an independently scored model at a known cost per task today, you do not need ERNIE 5.5 Flash Preview — you need a Luna, and specifically the current one. There is nothing to evaluate, because somebody else already did.

If you have a task whose output genuinely runs to hundreds of thousands of tokens in one pass — a long build, a full-repository refactor, a document nobody wants to split — ERNIE 5.5 Flash Preview is the only model on this page that publishes an output ceiling allowing it, and its output rate is slightly below GPT-5.6 Luna's. That is a narrow, specific case, and it is the only case the spec sheet supports.

If you are somewhere in between, the honest answer is that neither card settles it, because neither vendor published tokens per completed task for the workload you have in mind. GPT-5.6 Luna's $0.33 is Artificial Analysis' suite, not yours.

One key, two rate cards, and the migration that is not a project

ERNIE 5.5 Flash Preview is not something OrcaRouter serves. There is no ERNIE entry in the catalogue and no Baidu endpoint behind it, so Baidu's own Qianfan API and third-party platforms that resell it are where you would go. Nothing here is an offer to route it.

The Luna situation is the opposite, and it is instructive. Both GPT-5.6 Luna and GPT-6 Luna are in the OrcaRouter catalogue, at provider list price with no markup — $0.20 and $1.20 for the deprecated SKU, $0.10 and $0.50 for the one that replaced it. That is what a superseded rate card looks like from the inside: the model string changes, the key does not, and nobody has to sign anything. When a vendor retires a model mid-quarter — as happened here on 22 September — the swap costs one line in a config rather than a procurement cycle, and you can price the old and new cards against each other on the same bill.

What would settle this comparison is not another benchmark table. It is an ERNIE 5.5 row produced by Artificial Analysis, a published tokens-per-task figure for the eleven-hour build, and a rate card for ERNIE on the tier GPT-6 Luna now occupies. Until one of those exists, ERNIE 5.5 Flash Preview versus GPT-5.6 Luna is a two-day-old preview measured against a retired card by two sets of people who never shared a harness — and the useful version of the question is whether the preview can hold its output-ceiling advantage once somebody with no stake in the answer runs the test.