ERNIE-5.5-Flash vs GPT-6 Luna: What a Flash Tier Actually Owes You

Flash is not a parameter count. It is a commercial promise with four clauses: a price per million tokens, a latency you can plan a request around, a context window you can size an application to, and a score someone independent has checked. GPT-6 Luna, released by OpenAI on 22 September 2026, publishes all four. ERNIE-5.5-Flash, whose base weights Baidu released on 5 November 2026, publishes none of them — no price, no served latency, no independent evaluation, and a context figure that describes a checkpoint configuration rather than an endpoint. Setting ERNIE-5.5-Flash against GPT-6 Luna is therefore not a head-to-head. It is a useful test of what the word "Flash" is supposed to guarantee, and a demonstration of how much of that guarantee Baidu has not yet made. Luna is served at the provider's list price with routing policies for effort and context held in one place.

Clause one: price

GPT-6 Luna lists at $0.10 per million input tokens and $0.50 per million output. Cached input reads at roughly $0.01, cache writes at about $0.125, and prompts above 272K input tokens move to a higher tier at $0.20/$0.75. Against GPT-6 Sol at $2/$10 and GPT-6 Astra at $10/$50, Luna is roughly one-twentieth and one-hundredth the cost respectively — OpenAI frames the launch as a permanent 50 percent cut against GPT-5.6 promotional pricing, not a temporary discount.

ERNIE-5.5-Flash has no price. Not an unpublished one, not an enterprise-quote one: the post-trained model does not exist, so there is nothing to price. What exists is ERNIE-5.5-Flash-Base, a 280B-parameter sparse MoE with 14B active per token under the ERNIE 5.5 Flash Base Model License Agreement v1.0, free to download in FP8 and BF16.

That is the whole of clause one. A model at ten cents per million input tokens on one side, and a repository on the other.

Clause two: latency

Luna's latency profile is unusually honest about its own trade-offs. Output runs around 154 tokens per second. At low reasoning effort, the median time to first token is about 1.5 seconds. At maximum reasoning effort it stretches to roughly 124 seconds — a figure that makes the model unusable for interactive work at that setting and perfectly reasonable for batch analysis. A Flash tier that publishes both numbers is telling you which deployments it is for.

ERNIE-5.5-Flash publishes no throughput and no time-to-first-token, because there is no served endpoint to measure. What can be inferred is architectural: 14B active parameters per token out of 280B total is a design aimed at cheap, fast decoding, and 14B active is a smaller working set than GLM-5.3 Flash's 18B or DeepSeek V4.1 Flash's 8B-input/16B-output split. On paper it is built to be quick. On paper is where that claim has to stay until someone serves it.

Clause three: context

GPT-6 Luna's window is around one million tokens, with output capped near 128K. The detail worth noticing is that this is larger than GPT-6 Sol's 872K — the cheaper model accepts more input than the expensive one. That inverts the usual escalation logic. The standard routing pattern is "send the big payload to the big model"; Luna breaks it, and any router configured on the old assumption will overspend.

ERNIE-5.5-Flash-Base reports 128K training context and 1,048,576 tokens enabled for fine-tuning and inference in the released configuration. Those are not the same kind of number as Luna's. One is a configuration flag on a checkpoint that has never been served at that length; the other is a product limit with a price tier attached at 272K. Baidu's long-context evidence is a benchmark sweep — HELMET-RAG 71.5, RepoQA 91.7, LongBench-v2 48.7, InfiniteBench 46.5 — run under its own harness, and the post concedes that at least two of those long-context rows go to a competitor base model.

Clause four: a score someone else checked

Luna's published position is mixed and that is what makes it credible. BenchGecko reports an average score of 55.9 with a rank around 116th; vendor-adjacent figures include DeepSWE v1.1 at 66.6, OSWorld 2.0 at 52.7, LiveBench coding 79 and math 89.1. Terminal-Bench 4.0 comes in around 13.6 percent on Vals AI's run, and AutomationBench at 20.7. Artificial Analysis puts its intelligence index at roughly 37 to 38, below Sol at about 48 and Astra at about 53, and its coding agent index at about 41 — a regression from GPT-5.6 Luna's 43. Hallucination rate improved from 93 percent to 77 percent, which is better and still not good.

So Luna is a cheap model with a documented ceiling and at least one documented regression. That is more useful to a buyer than a strong number with no provenance, because it tells you where not to use it: agentic coding loops, high-stakes factual recall, and anything that depends on Terminal-Bench-style environment manipulation.

ERNIE-5.5-Flash's side of this clause is a base-model table. MMLU 93.3, MMLU-Pro 78.4, GPQA 60.2, HumanEval+ 82.3, LiveCodeBench 40.1 — all measured before post-training, in Baidu's harness, against other base checkpoints. It is a legitimate table. It is not an answer to "how good is the model I would call", because the model you would call has not been built.

Why "Flash" at 280B is a different animal

There is a structural point hiding under the comparison. Luna is a small model made cheap by being small. ERNIE-5.5-Flash is a large model made cheap by activating a small slice of itself — 14B of 280B per token. The economics are genuinely different: memory cost scales with the 280B total, compute cost scales with the 14B active. For a provider serving at volume, that is an attractive shape, because the expensive part is amortised across many concurrent requests and the cheap part is what each token costs.

For a self-hoster, the same shape is a problem. You need the memory for 280B parameters in FP8 — on the order of 280GB before KV cache — to get the compute profile of a 14B model. That is a data-centre deployment, not a workstation one, and it is why base checkpoints of this size tend to become APIs rather than local models.

Luna is the opposite: small enough to be served broadly, priced to reflect it, and available through OpenAI's API, ChatGPT Work, Codex and the desktop app on Free and Go tiers. The distribution is part of the product.

Where OrcaRouter fits, and where it does not

GPT-6 Luna is routable on OrcaRouter as openai/gpt-6-luna, at the provider's list price with 0% markup. That matters more than usual for this model, because the interesting routing decisions around Luna are not about price at all — they are about effort. A request that needs 1.5-second first-token latency should not be sent to the same effort setting as a batch job that can wait two minutes, and a workload that occasionally needs 800K of context should not default to Sol on the assumption that the bigger model has the bigger window. A routing layer that can hold those policies in one place, with automatic failover when a provider degrades, is doing real work here.

ERNIE-5.5-Flash is not in the OrcaRouter catalogue — 216 model IDs, no ERNIE entry — and nothing in this article should be read as a claim that it is. There is no endpoint to route to. The honest position is that the day Baidu serves the post-trained model, the routing question becomes interesting; today it is hypothetical.

What would make ERNIE a Luna competitor

Four things, and they are the same four clauses.

• A price at or below $0.10/$0.50. Anything above Luna's rate has to justify itself on Chinese-language quality or data residency, and Baidu has not made either argument for this model yet.

• A measured throughput number. The 14B active parameter design suggests it could be fast; until a provider publishes tokens per second and time to first token, that is an architectural inference, not a spec.

• A served context limit with a price attached. The 1,048,576-token configuration is a capability of the checkpoint. Luna's window is a product term, including the 272K pricing tier. Those are different commitments.

• A post-trained model with independent evaluation. This is the big one. The distance between a base checkpoint and a finished Flash model is where instruction following, tool use and formatting discipline live, and it is exactly the distance Luna has already travelled.

Until those four exist, the comparison has a clear answer for anyone buying capacity: GPT-6 Luna, with its documented weaknesses, is a real option at a real price, and ERNIE-5.5-Flash is a research release wearing a product tier's name. The more interesting observation is that Baidu chose to ship the base weights first and loudly documented the architecture — 43 layers, 37 HCA plus 6 MLA-DSA, 512 routed experts at top-10, a 201,216-token vocabulary — which reads like a lab inviting post-training work rather than a vendor opening a sales channel. Judge it by that standard and it is a good release. Judge it against Luna and it is not yet in the same market.