ERNIE-5.5-Flash vs Gemini 3.8 Flash: One Has a Price, the Other Has a Promise

Fifty-seven days from today, on 1 January 2027, the introductory pricing on Gemini 3.8 Flash ends and its rates double — $0.75 to $1.50 per million input tokens, $3.75 to $7.50 per million output. That is the most concrete fact in this comparison, and it belongs to the model that has been on sale since 2 September 2026. ERNIE-5.5-Flash, by contrast, published its base weights on 5 November 2026 with no price at all, because the post-trained ERNIE 5.5 Flash has not been released. So the honest framing of ERNIE-5.5-Flash vs Gemini 3.8 Flash is not a contest between two Flash-tier models. It is a contest between two kinds of uncertainty: a known model with a dated cost cliff, and a promised model with no cost at all. Gemini's January increase reaches the platform at provider list prices so the vendor's own rate is what you pay.

What Gemini 3.8 Flash ships, and what it will cost after New Year

Gemini 3.8 Flash is the third Flash model Google shipped in six weeks — 3.6 on 21 July, 3.7 on 14 August, 3.8 on 2 September — and it is a full product, not a checkpoint. The input context is 1,048,576 tokens, output is capped at 65,536, and the model accepts text, images, video, audio and PDFs with text out. Knowledge cutoff is March 2026.

The price structure is where the comparison actually bites. Today the rates are $0.75 input and $3.75 output per million tokens, identical to Gemini 3.7 Flash. Cached input reads at $0.075, batch at $0.375/$1.875. From 1 January 2027, list pricing becomes $1.50 and $7.50. Anyone building a 2027 budget on today's number is budgeting on a promotional rate.

There is a second cost trap underneath the first. Gemini 3.8 Flash produces roughly 30 to 34 percent more output tokens per task than 3.7 Flash did, which turns a flat per-token price into a higher cost per completed task — by one measurement, about 45 percent more at high effort. Price per million tokens is not price per job, and this model is the clearest recent example of the difference.

Its benchmark position is real but version-dependent. Vendor figures include DeepSWE v1.1 at 73.7 percent, Terminal-Bench 2.1 at 89.4, OSWorld-2.0 at 59.0 and HLE-Verified at 54.9. Artificial Analysis coverage puts output speed at roughly 302 to 351 tokens per second, at the top of its leaderboard, with a time-to-first-token around 30 seconds at high effort. The intelligence index is where the reporting fragments: launch coverage quoted 59 at high reasoning, while Artificial Analysis's own pages under newer index versions show figures in the 41 to 47 range. That is not a scandal, it is a rebase — the index was rebuilt on 27 October 2026 — but it means any single AA number quoted for this model without its version is close to meaningless.

What ERNIE-5.5-Flash-Base ships, and what it cannot do

Baidu's release is the opposite shape. ERNIE-5.5-Flash-Base is a 280B-parameter sparse MoE with 14B active per token, pretrained on 33.5T tokens, released in FP8 and BF16 under the ERNIE 5.5 Flash Base Model License Agreement v1.0. Training context 128K; the released configuration enables 1,048,576 tokens for fine-tuning and inference, which is the same headline figure as Gemini's input window and a very different thing — it is a configuration setting on a checkpoint, not a served limit with a rate card behind it.

The capability gap is blunt. ERNIE-5.5-Flash-Base is text-only. There is no image input, no video, no audio, no PDF parsing. If your workload involves a screenshot, a scanned invoice, a call recording or a product video, the comparison ends there — Gemini 3.8 Flash does those natively and ERNIE's released checkpoint does not do them at all. The post-trained model could change that, but no multimodal ERNIE 5.5 Flash has been announced, and Baidu's spec sheet describes a language model.

What ERNIE does publish is a base-model evaluation: MMLU 93.3, MMLU-Pro 78.4, GPQA 60.2, HumanEval+ 82.3, LiveCodeBench 40.1, HELMET-RAG 71.5, RepoQA 91.7, LongBench-v2 48.7. Every one of those numbers was produced on a base checkpoint, in Baidu's own harness, against other base models. Setting them beside Gemini's instruct-model scores would be the single most common error available in this comparison, and it is worth naming plainly: the two sets measure different things and cannot be subtracted.

The two uncertainties are not equivalent

A price cliff is a known unknown. You can model it: take today's rate, double it on 1 January, and decide whether the workload still clears your margin. A missing price is an unknown unknown, and it cannot be modelled at all. That asymmetry is the whole decision.

There is a second asymmetry in the benchmark record. Gemini 3.8 Flash's numbers are disputed — vendor benchmarks against independent index figures that moved when the index was rebuilt, with a documented gap between the 59 quoted at launch and the 41-to-47 range now visible on Artificial Analysis. ERNIE 5.5 Flash's numbers are not disputed. They are absent, because there is nothing to dispute yet beyond a base-model table. Disputed data is more useful than no data: at least you know the range.

The one place ERNIE has a defensible edge on paper is the parameter-efficiency story. A 280B model activating 14B per token is engineered for cheap inference, and 33.5T pretraining tokens is a serious data budget. If Baidu prices the post-trained model aggressively, the active-parameter count is the mechanism that would let it undercut a $1.50/$7.50 Gemini. Whether it will is unknowable today, and any article that tells you otherwise is guessing.

When the price doubles, routing is the lever

The useful thing about a dated price change is that it can be planned for. Workloads that are marginal at $0.75/$3.75 and unviable at $1.50/$7.50 need a migration path that exists before January, not after. That is a routing problem more than a model problem: keep the expensive model for the requests that need its multimodal input or its long-context recall, and send the rest to something cheaper with a compatible interface.

OrcaRouter carries google/gemini-3.8-flash at the provider's list price with 0% markup, so the January increase will appear at the same rate the vendor charges rather than being absorbed into a reseller margin. The routing DSL lets you compose a fallback chain — Gemini first for the multimodal path, a cheaper text model behind it — and automatic failover covers the case where a provider degrades. ERNIE-5.5-Flash is not in that catalogue; the 216 model IDs include no ERNIE entry. So the routing advice here is about the model you can actually switch away from, which is the one with the price clock attached.

A decision rule rather than a verdict

If your workload touches images, audio, video or PDFs, use Gemini 3.8 Flash and budget for the January doubling. ERNIE-5.5-Flash-Base is not a candidate and will not be one until a multimodal post-trained model ships.

If your workload is text-only, Chinese-heavy, and cost-sensitive, the correct action today is still not ERNIE — it is to wait for a price. The base weights are available and free to download, which makes them useful for a fine-tuning experiment, not for a production decision. The moment Baidu publishes a Qianfan rate for the post-trained model, this comparison becomes a genuine cost question, and the 14B active parameter count suggests the answer could be interesting.

If you are simply trying to understand the ERNIE release, the thing to internalise is that "Flash" here is a marketing tier inherited from a model that has not shipped. Gemini's Flash tier has a price, a context window, a modality list and a dated increase. ERNIE's has a Hugging Face repository and a sentence promising more. Between a known cost that is about to rise and an unknown cost that does not yet exist, the rational move is to plan against the known one and keep an eye on the other.