OpenRouter's ERNIE-5.5-Flash-Preview listing says over 200 tokens a second; its own median reads 74

The text says more than 200 tokens a second; OpenRouter's median reads 74 (at least 2.7 times apart, derived)

Baidu's ERNIE-5.5-Flash-Preview appeared on OpenRouter on 10 November 2026 (the page's release date; its data gives 17:58:53 UTC). The listing page carries a model description whose last sentence ends with "delivering generation speeds of more than 200 tokens per second". Further down, in the performance panel, OpenRouter's own measurement for the single provider that serves it, Baidu Qianfan, is a median of 74 tokens per second over a 30 minute window (the page data as read at 12:38 UTC on 13 November 2026: 75,741 requests with a throughput reading). The distance between the two is at least 2.7 times (derived: 200 divided by 74 = 2.70; the claim says "more than" 200, so the true ratio is higher), and the measurement is 37 percent of the claimed floor (derived: 74 divided by 200).

There is no Baidu primary behind this briefing. Baidu's own announcement and model card were not available to this site: the Hugging Face card answers 401 to scripts and was not requested, and no Baidu page was saved. Everything below comes from OpenRouter's two listing pages, OpenRouter's public models API, and three independent leaderboards that were checked for the two model ids. The description is Baidu's wording as OpenRouter copies it; the speed measurement is OpenRouter's.

What this does not establish. It does not show that either number is wrong. They are not the same kind of number: one is a sentence in a vendor description with no stated method, hardware, load or token definition, and the other is a median over live traffic, with whatever prompt lengths and other customers the endpoint had in that window (the page does not say how it counts reasoning tokens). Neither the listing nor the API prints a benchmark for the model, so nothing here says how good its answers are.

The size figures are in the same description: "280 billion total parameters", "activates 14 billion parameters per token" and "a context window of 1 million tokens". 14 of 280 is 5 percent (derived). The API prints the context as 1,024,000 tokens, which is 2.4 percent above one million (derived: 1,024,000 divided by 1,000,000 = 1.024). The price is free.

Whose interest. Baidu sells access to its ERNIE models, is the provider on both listings, and wrote the description; OpenRouter is a marketplace that lists this free preview next to priced models, and the performance panels are its own measurements. The description is therefore a vendor claim, and the panels are a marketplace's record, and this briefing keeps the two apart.

Read times. The lead's saved copies of the two listing pages (rendered in a browser and converted to text) were read between 12:30 and 13:00 UTC on 13 November 2026. This site then read the server rendered HTML of both pages and the endpoints API at 12:38 UTC (the page data is stamped 12:38:09 UTC), and the three leaderboards at 12:39 UTC. Throughput, uptime and usage figures move minute to minute; anything after those times is not included.

Zoom

Two numbers that are not the same measurement

OpenRouter's page defines throughput as "how fast the model writes", in tokens per second. The panel's headline reads "P50, best across providers", and with one provider that is the median for Baidu Qianfan. The saved page prints 72 tokens per second in the provider row and the headline; this site's read of the page data at 12:38 UTC gives 74 for the 30 minute window. The page also prints, for a three day window, a P50 average of 89 tokens per second. So the page itself shows the figure moving between 72, 74 and 89 depending on the window and the minute. Against the claim, those give at least 2.78 times, 2.70 times and 2.25 times (derived: 200 divided by 72, 74 and 89).

The page does not say, in the text saved, which direction the percentiles run, so this briefing reports the values and does not interpret them. What can be said is that the page's own figures fall from 74 to 8 across the columns, so a reader's experience depends on where in that spread their requests land, and that the claim of more than 200 sits above all of them.

Three things the claim does not say. It does not say what a token is counted as: the model has mandatory reasoning (the API prints reasoning as mandatory, with high as the only supported effort), so a speed that counts only visible output differs from one that counts reasoning tokens. It does not say what prompt length or concurrency applied; the page's cache panel shows an average cache hit rate of 93.01 percent for this provider, which says real traffic reuses long prompts, not that Baidu's figure did. And it does not say what hardware ran it. The description's own wording ties the speed to a "new architecture" that "combines strong model capabilities with efficient inference", which is a sentence about design, with no method behind the number.

The second Baidu model, listed the same day

The same pages list a second Baidu model released on 10 November 2026, ERNIE-Light-DT-1.0 (17:21:00 UTC in the API's created field, about 38 minutes before the preview, derived). It is not free: the page prints $1.13 per million input tokens and $3.94 per million output tokens, and $0.113 per million for cache reads. The cache read price is one tenth of the input price (derived: 0.113 divided by 1.13 = 0.10). The page's weighted average input price actually paid is $0.2373 per million, which is 21 percent of the listed $1.13 (derived), consistent with the page's 87.8 percent cache hit rate (derived check: 0.122 x 1.13 + 0.878 x 0.113 = 0.2371). A request with 100,000 input tokens and 1,000 output tokens lists at $0.117 on the Light model (derived: 0.1 x 1.13 + 0.001 x 3.94 = 0.11694) and at $0 on the preview.

Its description, as printed in the page's structured data, says it "is based on open-source GLM models and includes research findings from Baidu Search's Lightning large model". It says it "features Baidu Sparse Attention (BSA), a unique sparse attention mechanism developed by Baidu ERNIE", and that this "significantly reduces KV cache storage and computational costs while maintaining model performance". Those are claims by Baidu. The description prints no parameter count, does not say which GLM models or which versions, and prints no licence for the result. The API gives it the same 1,024,000 context as the preview and a maximum completion of 131,072 tokens. Its median throughput on the page is 63 tokens per second, from only 123 throughput readings in the 30 minute window (the preview had 75,741), so that figure rests on very little traffic.

Zoom

The two entries differ in size claim, price, reasoning behaviour (the preview's reasoning is mandatory with only high supported; the Light model's is optional with max as the default) and weights link. The listing says nothing about how the two relate. This briefing therefore does not call the Light model a smaller version of the preview, a sibling or a distillation: none of those is stated.

What the listing states, and what it does not

Independent measurements. A model launch brief looks first for scores from sources that did not write the description. This site read the robots.txt of each host and looked once for the two model ids. Artificial Analysis's models index (robots.txt 200, all paths allowed) lists 526 model entries, the newest dated 12 November 2026; its only two Baidu entries are ERNIE 4.5 300B A47B (30 June 2025) and ERNIE 5.0 Thinking Preview (13 November 2025). Neither ERNIE-5.5-Flash-Preview nor ERNIE-Light-DT-1.0 is on it. The Vals Index page (robots.txt 200 after a redirect to the www host) and the CWE-bench home page (robots.txt 200) contain no ERNIE or Baidu entry at all. This is the plain result for a model that is three days old, not a finding against it.

OpenRouter's own models API shows the same gap. Its models feed has 472 entries; 266 carry a benchmarks object and 209 of those carry Artificial Analysis data (derived counts from the saved JSON). Neither Baidu entry carries one. The API entry for the preview also prints no expiry date, while 22 of the 472 entries do carry one.

Friendly labels doing the work of evidence

"Flash" and "Preview". The listing's description defines neither. "Flash" suggests a fast, small tier; the 14 billion active of 280 billion is consistent with a sparse design, but the description does not say the model is a mixture of experts, so that is this site's inference, not Baidu's statement. "Preview" tells a developer to expect change; nothing on the page says what changes, when, or whether the identifier continues to work. The API prints the dated identifier ending 20261110 and no expiry date.

"Free". The price is zero for prompt and completion tokens, and the FAQ text on the page adds "Free endpoints are rate limited". A zero price per token is not a zero cost per accepted answer. It also says nothing about what happens to prompts: the provider's data record on the same page says the provider retains them (next section).

"Model weights". The preview's page has a button with that label that links to a Hugging Face repository for the model. A label is not a licence. This site did not request Hugging Face, so it cannot say what the weights are, whether they can be downloaded, or under what terms. The Light model's page has no such link, and its description says it is "based on open-source GLM models", which says something about its base and nothing about the licence of what Baidu built on it.

"1 million tokens" and "100 percent uptime". The context is 1,024,000 as the API prints it. The maximum completion differs by source: the API prints 921,600 (90 percent of the context, derived: 921,600 divided by 1,024,000), and the page's FAQ text says it supports up to 1,024,000 completion tokens. The provider row shows uptime 100.00 percent; the saved page's three day panels print Uptime 99.99 percent and Availability 99.87 percent, and the page defines the first as the share of time that at least one provider was responding and the second as the share of time that inference was successfully served. With one provider, the stricter number is the second.

The free endpoint's data record

The visible text of the page does not mention a data policy. The structured data embedded in the page does. For the provider Baidu Qianfan it records training as false, "trainingOpenRouter" as false, "retainsPrompts" as true, "canPublish" as false and "safetyRetention" as false, with a provider headquarters of "CN" and an empty list of datacentres. It also records links to Baidu Cloud's terms of service and privacy policy, which this site did not read. The saved page adds, in its API panel, "Not currently available through EU in-region routing."

So the page's own record says that, on this endpoint, prompts are retained by the provider, with no retention period, no storage location and no access rule printed. It does not say prompts are used for training; the record says they are not. It says nothing about law. This site's caution, not the page's, is that a provider headquartered in China may be subject to Chinese data law, and a team that sends it customer or regulated data should take that advice from its own counsel before the first request, not after.

What the claim implies for memory bandwidth (derived, an illustration)

The listing does not state the numeric format of the served weights, and the endpoint prints quantization as "unknown", so the arithmetic below tries three common sizes and claims none of them. At one request at a time, each generated token needs the active parameters read from memory once. Bytes per token is 14 billion times the bytes per parameter, and the bandwidth needed is that multiplied by tokens per second (1 GB is 10 to the power 9 bytes).

It is a ceiling style calculation, not a benchmark. It ignores batching (a busy endpoint shares each weight read across many requests), speculative decoding or multi token prediction, how experts are placed across accelerators, interconnect, and the KV cache. This site did not read an accelerator datasheet for this briefing, so the table is not compared with any named hardware. What it shows is that the gap between 74 and 200 is a factor of 2.7 in memory traffic per stream at the same format, which a vendor can close with serving choices the listing does not describe, or not close at all.

Where the sources differ

What to do, in order

Technical details

The listing is a marketplace page, not a model card, so the artifacts it prints are descriptions, identifiers, prices and measurements. Each block below is copied from the saved pages or the models API as read on 13 November 2026, and says what a developer does with it. This listing prints no exploit, indicator or command beyond the example request in the last block.

The description in full. This is the text in the page's structured data. The visible page and the API cut it after 280 billion. It is Baidu's wording as OpenRouter prints it. Use it as the vendor's claim, to be tested, not as a result.

ERNIE 5.5 Flash Preview is the latest foundation model in the ERNIE family, with comprehensive improvements in coding, agentic capabilities, knowledge work, and deep research. It has 280 billion total parameters, activates 14 billion parameters per token, and supports a context window of 1 million tokens. Its new architecture combines strong model capabilities with efficient inference, delivering generation speeds of more than 200 tokens per second.

The identifiers and limits. These are fields from the models API entry for the preview. A developer pins the id, sets max_tokens below the printed completion limit, and records that the pricing strings are zero.

{
 "id": "baidu/ernie-5.5-flash-preview:free",
 "canonical_slug": "baidu/ernie-5.5-flash-preview-20261110",
 "hugging_face_id": "baidu/ERNIE-5.5-Flash-Preview",
 "context_length": 1024000,
 "pricing": {
  "prompt": "0",
  "completion": "0"
 },
 "top_provider": {
  "context_length": 1024000,
  "max_completion_tokens": 921600,
  "is_moderated": false
 },
 "expiration_date": null
}

The reasoning and parameter fields. The reasoning object says reasoning is mandatory and high is the only supported effort. A developer budgets for reasoning tokens in every request and does not expect to switch reasoning off.

{
 "reasoning": {
  "mandatory": true,
  "default_enabled": true,
  "supported_efforts": [
   "high"
  ],
  "default_effort": "high"
 },
 "supported_parameters": [
  "include_reasoning",
  "max_tokens",
  "reasoning",
  "reasoning_effort",
  "response_format",
  "structured_outputs",
  "temperature",
  "tool_choice",
  "tools",
  "top_p"
 ]
}

OpenRouter's measurement object. These are the endpoint statistics embedded in the page, with latency in milliseconds and throughput in tokens per second over the stated 30 minute window. A developer logs the same two quantities from their own traffic and compares.

{
 "endpoint_id": "d8f1ef6c-9e39-443a-adb0-e5c9d716c9b1",
 "latency_metric": "latency",
 "p50_latency": 2886.5,
 "p75_latency": 4058,
 "p90_latency": 6162.900000000001,
 "p95_latency": 7897,
 "p99_latency": 11683.710000000003,
 "p50_throughput": 74,
 "p75_throughput": 37,
 "p90_throughput": 21,
 "p95_throughput": 15,
 "p99_throughput": 8,
 "mean_latency": 3453.2299992640023,
 "latency_request_count": 81522,
 "throughput_request_count": 75741,
 "request_count": 81522,
 "window_minutes": 30,
 "throughput_window_minutes": 30
}

The provider's data record. This is the object the page embeds for the provider, as printed. A reviewer reads retainsPrompts first and takes the linked policy to counsel.

"dataPolicy":{"training":false,"trainingOpenRouter":false,"retainsPrompts":true,"canPublish":false,"safetyRetention":false}
"headquarters":"CN","datacenters":[]

The Light model's pricing. The models API prints prices per token as strings. A developer multiplies by 1,000,000 to compare with the page's per million figures: 0.00000113 is $1.13.

{
 "id": "baidu/ernie-light-dt-1.0-intl",
 "pricing": {
  "prompt": "0.00000113",
  "completion": "0.00000394",
  "input_cache_read": "0.000000113"
 },
 "context_length": 1024000,
 "top_provider": {
  "context_length": 1024000,
  "max_completion_tokens": 131072,
  "is_moderated": false
 }
}

The example request. The listing prints this streaming request for the free preview, with the key left as a placeholder. A developer substitutes their own key from the environment and never pastes it into code. Sending data to it is the step the data record above bears on.

curl -N https://openrouter.ai/api/v1/chat/completions \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer $OPENROUTER_API_KEY" \
   -d '{
  "model": "baidu/ernie-5.5-flash-preview:free",
  "stream": true,
  "messages": [
     {"role": "user", "content": "Hello"}
  ]
}'

Where the weights link points. The page links the model weights to this repository. This site did not request it (the host answers 401 to scripts), so nothing about its contents or licence is known here.

huggingface.co/baidu/ERNIE-5.5-Flash-Preview

The question the listing leaves open

Under what prompt length, concurrency, hardware and token definition did Baidu measure more than 200 tokens per second, and which party outside Baidu has reproduced it, given that OpenRouter's own median for the only provider is 74 and the page prints no benchmark, no licence and a provider record that says prompts are retained?

Key facts