ERNIE-5.5-Flash vs Claude Haiku 5.5: A Base Checkpoint Meets a Price Floor

Baidu put ERNIE-5.5-Flash-Base on Hugging Face this morning, and the first line of the model card is the one that matters: the weights are the end-of-pretraining checkpoint, with no mid-training and no post-training applied. The post-trained ERNIE 5.5 Flash, Baidu's release post says, "will follow." That single sentence makes a direct comparison with Claude Haiku 5.5 impossible today, because Claude Haiku 5.5 is a finished, instruction-tuned model that Anthropic sells at $0.10 per million input tokens and $0.50 per million output tokens. One side of this matchup is a research artifact. The other side is a product with a price list. What follows is what each one actually is, what the numbers Baidu published can and cannot be set against, and what would have to change before the comparison people are searching for becomes a real one. ERNIE-5.5-Flash is not in it today. browse the full model catalogue It lists every model OrcaRouter can route, with live provider pricing.

What Baidu actually published today

The release is two repositories — baidu/ERNIE-5.5-Flash-Base in FP8 and baidu/ERNIE-5.5-Flash-Base-BF16 — plus a technical post under Product Updates on ernie.baidu.com. It is a sparse mixture-of-experts language model with 280B total parameters and 14B activated per token, pretrained from scratch on 33.5T tokens using the Muon optimizer and FP8 mixed precision. The Hugging Face sidebar rounds the total to 286B; the card says both figures exclude 0.36B of what Baidu calls ECHO parameters.

The architecture is specific, which is unusual for a quiet drop. Forty-three layers, one dense and 42 MoE. Thirty-seven layers use the compressed attention Baidu calls HCA, caching one entry per 128 tokens; six use MLA with DeepSeek-style sparse attention. Hidden size 4,096, 64 heads, 512 routed experts with top-10 routing plus one shared expert, and a vocabulary of 201,216. Training context is 128K; the released configuration enables 1,048,576 tokens for fine-tuning and inference.

Tooling shipped with it. The card carries SGLang and vLLM deployment examples, and NVIDIA's Megatron-Bridge repository already has an ERNIE 5.5 example for fine-tuning. That is the practical footprint of a base release aimed at people who intend to train on it, not people who intend to call it.

What did not ship: an instruct model, a chat template worth the name, an API listing, a price, or a single independent benchmark. There is no ERNIE entry in OrcaRouter's 216-model catalogue as of publication, and no Artificial Analysis page. The only scores in existence are Baidu's own, run on 30 benchmarks under a unified harness against three other base models: DeepSeek-V4-Flash-Base at 284B-A13B, Hy3-preview-Base at 295B-A21B, and MiMo-V2.5-Base at 310B-A15B.

What Claude Haiku 5.5 is selling

Claude Haiku 5.5 arrived in early October 2026 as Anthropic's small, fast tier, and its commercial shape is the opposite of Baidu's release: everything is published, priced, and served. The headline was a price cut of roughly 90 percent against Haiku 4.5 at the short-context tier — $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, rising to $0.50 and $2.50 above that. Cached input reads at $0.01, batch at half the standard rate. Anthropic's own framing is that a new tokenizer counts roughly 30 percent more tokens for the same text, so the effective saving per task is smaller than the sticker suggests — around 75 percent by Anthropic's own estimate.

The spec sheet is a product sheet. One-million-token context, 128K maximum output, text and image input with text out, a knowledge cutoff of June 2026, and five effort levels from low to max. Anthropic positions it as its fastest model at standard speed; Artificial Analysis measures it at 258 output tokens per second on its current run. It is available through Anthropic's own API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.

Those are the two objects. A 280B-parameter base checkpoint under a bespoke licence, and a hosted 1M-context model at ten cents per million input tokens. Putting them in the same sentence is defensible only if you are explicit about which question you are answering.

Why 93.3 MMLU is not a Haiku number

Baidu's card reports MMLU 93.3, MMLU-Pro 78.4, GPQA 60.2, BBH 77.1 in direct mode and 89.2 with chain of thought, HumanEval+ 82.3, MBPP+ 81.5, LiveCodeBench 40.1, HELMET-RAG 71.5, RepoQA 91.7 and LongBench-v2 48.7. Those are strong numbers for a base model. They are also, structurally, not comparable to anything Anthropic publishes for Haiku 5.5.

Three reasons, and all three matter. First, base models are evaluated with different prompting — few-shot completion rather than instruction following — so a base score measures the pretrained prior, not the assistant a developer calls. Second, the harnesses differ: Baidu ran its own pipeline with its own prompts and few-shot counts, which is the right way to compare its model to other base models and the wrong way to compare it to a vendor's instruct benchmarks. Third, MMLU at 93 is deep into saturation; the differences between top models on that benchmark are now smaller than the formatting effects between harnesses.

So the honest reading of "ERNIE-5.5-Flash vs Claude Haiku 5.5 on benchmarks" is: there is no such comparison yet, and anyone publishing one this week is either comparing a base checkpoint to an instruct model without saying so, or inventing the instruct numbers.

One detail from Baidu's own table is worth keeping, because it cuts against the launch narrative. The post concedes that competitors win several rows — ChineseSimpleQA, INCLUDE-lite, HELMET-Rerank, InfiniteBench and FullStackBench go to one of the other base models. A vendor that publishes its losses in the same table as its wins is giving you a usable signal about how much weight the rest of the table deserves.

The price that does not exist yet

"Flash" in Baidu's naming is a tier label, not a specification. The ERNIE line has an established price ladder — ERNIE 5.1 sits around $0.59 input and $2.65 output per million tokens on the international Qianfan listing, with ERNIE 4.5 Turbo far below that — but ERNIE 5.5 Flash has no published rate, no Qianfan endpoint, and no third-party listing. Searching for its pricing returns the base release and nothing else.

That gap is the entire commercial question. Haiku 5.5's $0.10/$0.50 reset what a small model is allowed to cost, and it is the number every Flash-tier launch is now measured against. If Baidu prices ERNIE 5.5 Flash above that, the model has to justify it on Chinese-language quality or on data residency. If it prices below, it becomes a serious option for mainland workloads. Until a rate appears, neither branch can be checked.

Artificial Analysis, whose index was rebuilt on 27 October 2026, places Claude Haiku 5.5 at 36 at max effort — a figure that belongs to the current index and cannot be compared with pre-rebase numbers still circulating for other models. The point is not the exact value. It is that Haiku has a place on that scale and ERNIE-5.5-Flash does not, because there is no instruct model to test.

This is also where routing changes the decision. OrcaRouter passes provider list prices through at 0% markup, so when a vendor cuts a rate the new price is live on the platform the same day rather than after a reseller margin is recalculated. ERNIE-5.5-Flash is not in that catalogue today — 216 model IDs, none of them ERNIE — so the practical advice for anyone planning a cost model is to build it on the models that are actually priced, and revisit when Baidu publishes a rate card.

Who should clone the repo this week

Not application developers. The base checkpoint has no instruction-following behaviour to speak of, and getting it to behave like an assistant means a post-training pipeline you would have to design yourself. Three groups do have a reason to pull the weights now.

• Domain-adaptation teams working in Chinese — 33.5T pretraining tokens and a 200K vocabulary, continued-pretraining on a base is a legitimate starting point, and the Megatron-Bridge example lowers the cost of trying.

• RL and post-training researchers who want a 14B-active model with published architecture details — the 512-expert, top-10 routing and the 37/6 HCA-to-MLA split are documented well enough to reproduce.

• Evaluation teams who want a clean baseline before the instruct model lands — running the base yourself now gives you a reference point that vendor marketing cannot move later.

Everyone else should wait. A model you cannot call, price, or benchmark against an instruct model is not a procurement option, and the 280B total parameter count means self-hosting it is a multi-GPU proposition even at FP8.

What would change the verdict

Four events, in the order they are likely to matter.

• The post-trained ERNIE 5.5 Flash ships. Baidu says it will follow. Until it does, every capability comparison is speculative, and the gap between a base checkpoint and a finished Flash model is exactly where most of the user-visible quality lives.

• A price appears on Qianfan. That is the moment the Haiku comparison becomes a real decision rather than a category error.

• An independent evaluation lands. Artificial Analysis or LMArena coverage would convert Baidu's harness into something cross-checkable. Note that the AA index itself was rebased on 27 October 2026, so pre-rebase figures circulating for Haiku 5.5 and others should not be compared to post-rebase ones.

• A third-party host picks it up. That is usually the signal that serving costs are understood, and it is the point at which a model becomes routable rather than merely downloadable.

The comparison in the title will be worth writing when two of those four have happened. What exists today is a well-documented base model, a licence that is not Apache, and a product-tier name with no product behind it. Claude Haiku 5.5, meanwhile, is a finished model with a price that other vendors now have to answer. On the only question a buyer asks — what does a token cost and what does it do — one side has an answer and the other has a promise.

The interesting part is not that Baidu shipped quietly. It is that the company chose to ship the base weights first, before the model anyone can use, and priced nothing. That is a research-lab release cadence, not a product launch, and it should be read as one.