Baidu's newest foundation model, released November 11, 2026: a 280B-total / 14B-active sparse MoE with a 1M-token context window, built as the execution layer for agent workloads. The Qianfan API and OpenRouter (paid and free tiers) are both live — but the positioning needs stating first: in Baidu's own words, Preview is an intermediate checkpoint in post-training that "may have capability gaps due to insufficient training," with the final version to follow — every one of the roughly 20 benchmark rows on this page is vendor-reported, and only 1 of the 8 cross-checked tests is a win; as of November 16 it has no Artificial Analysis, LiveBench or LMArena score, the Preview weights are not public, and the license accounts conflict (several outlets say Apache 2.0; the official release page does not confirm it). This page collects the official specs, pricing, benchmark analysis and 6 in-depth reviews — with unverified claims clearly marked.
What is confirmed: this site's first Baidu ERNIE topic — and an "unfinished" one. "Preview is an intermediate build, all benchmarks are self-reported, and the weights are not out yet" is the one-line summary of ERNIE 5.5 Flash Preview.
Where it stands among its closest Chinese peers: AA Index, price, context and open weights, side by side. Yuan and dollar prices are separate list prices — do not convert directly.
| Model | AA Index | Price (input · output, per 1M tokens) | Context | Open weights |
|---|---|---|---|---|
| ERNIE 5.5 Flash Preview | Not rated | ¥1 / ¥4 (Qianfan, after 50%-off promo, peak) | 1M tokens | No (weights promised Nov 25) |
| Tencent Hunyuan Hy4 | 35 | ¥6 / ¥18 (list; 50% off through Dec 31) | 1M tokens | No (preview weights only) |
| Qwen4-Max | 43 | $1.50 / $4.50 | 1M tokens | No |
| Kimi K3.1 Pro | 46 | $2 / $10 | 1M tokens | Yes (custom license) |
| LongCat 2.5 | Not rated | $0.30 / $1.20 (promo) | 1M tokens | Yes (MIT) |
AA Index from Artificial Analysis. ERNIE 5.5 Flash Preview and LongCat 2.5 were not rated as of Nov 16; ERNIE’s headline benchmark is Terminal-Bench 4.0 at 46.5 (vendor-reported) — see section 12. Prices per the official pages.
The grouped comparison from our benchmarks article, up front: across four tests, the gap between 14B-active ERNIE and the flagships is plain to see — it wins only AutomationBench.
Short on time? Start with these three: the benchmark tally, the flagship comparison, and the pricing guide.
A four-act release: the base weights on November 5, Preview on November 11, the Baidu World stage on November 12, and an open-source promise for November 25 — the last act has not happened yet, so this page states it as a promise, not a fact.
ERNIE 5.5 Flash Base releases: the pre-training endpoint weights land on Hugging Face (baidu/ERNIE-5.5-Flash-Base, FP8 and BF16 builds) under a custom Baidu license, not Apache 2.0; Base is base-model capability only, with no instruction following.
Baidu first discloses its RSI (recursive self-improvement) progress: the automated research system ERNIE Evolve.
OpenRouter lists baidu/ernie-5.5-flash-preview early (paid tier $0.28/$1.13) plus the :free tier; Qianfan documentation is dated the same day.
Official release of ERNIE 5.5 Flash Preview: the first post-trained version of Base. The official release page (ernie.baidu.com) is dated Nov 11, the official WeChat post goes out the same day, and 每日经济新闻, 新华网, 新智元 and 量子位 all cover it the same day.
Baidu World 2026 opens in Beijing, Robin Li keynotes, and ERNIE 5.5 Flash Preview shares the stage with the "Token Exchange" (词元所) marketplace.
Hands-on reports arrive: madechango / 火星财经 ("good taste, strong coding"); OpenRouter's free-tier measured median speed is p50 ~75–79 tok/s (Nov 12 data).
Correction: the CodeArena 1668 is an AutoEval placeholder: Stack Futures checks the Nov 11 board snapshot — blank rank, N/A votes, AutoEval label, Proprietary license tag; Arena said on Nov 12 that the model was "trending towards becoming the #2 open model", so "No. 2 open, No. 4 among Chinese models" is a forecast, not an established ranking. This site has corrected the wording accordingly.
BrowseComp (web research) listing: ERNIE 5.5 Flash Preview scores 87.4%, #15 of 62 (Inference360 board; vendor-reported). See section 12.
Promised, not delivered: open-sourcing the Preview weights — the English release page says "by Nov 25," the Chinese WeChat post says "on Nov 25"; as of Nov 16 the weights have not landed and the license is unconfirmed.
As published on the official release page and cross-checked against secondary sources that read it directly (this research round could not read the official page directly — see section 17). Benchmarks are in section 12; this section lists the specs only.
Four routes: a free web chat, the official Qianfan API, OpenRouter (paid and free), and self-hosting once the weights land. The cheapest hands-on route is the web chat or OpenRouter's free tier — but the free tier allows only 50 requests a day.
ernie.baidu.com/web/new-chat offers direct access with no API key — the quickest way to see the model's style and Chinese-language behavior before wiring up an API.
Baidu AI Cloud's Qianfan platform, model ID ernie-5.5-flash-preview, listed for both text generation and deep thinking; default rate limits are 60 RPM and 1,000,000 TPM. Pricing has peak/off-peak tiers (see section 08); the exact peak/off-peak windows are per the Qianfan console.
Paid tier baidu/ernie-5.5-flash-preview ($0.28/$1.13, cache reads $0.028), listed Nov 10; free tier baidu/ernie-5.5-flash-preview:free ($0, 20 requests/min, 50 requests/day; buying 10 credits raises it to 1,000/day). The free tier is the shortest zero-cost path to hands-on testing.
Preview weights are not public as of Nov 16, with open-sourcing promised for around Nov 25; Base weights are downloadable (baidu/ERNIE-5.5-Flash-Base, FP8 + BF16) but under a custom license with restricted commercial terms. The vendor says Preview supports SGLang and vLLM — verifiable only after the weights land. See our deployment and open-source outlook.
The rate itself is cheap — Qianfan peak ¥1/¥4 and off-peak ¥0.5/¥2 after the limited-time 50% discount (original ¥2/¥8), OpenRouter $0.28/$1.13, and a $0 free tier. Prices checked Nov 17, 2026 (51AllAI, citing Qianfan's model-service billing page); the checkout page price is authoritative.
| Item (per 1M tokens) | Qianfan peak | Qianfan off-peak | OpenRouter paid |
|---|---|---|---|
| Input | ¥1 | ¥0.5 | $0.28 |
| Output | ¥4 | ¥2 | $1.13 |
| Cache-hit input | ¥0.1 | ¥0.05 | $0.028 |
| Free tier | — | — | :free $0 (20 req/min, 50 req/day) |
51AllAI, citing Qianfan's model-service billing page, settles the wording: ¥2/¥8 is the original price, and the limited-time 50% discount brings peak hours to ¥1 in / ¥4 out and off-peak hours to ¥0.5 in / ¥2 out; peak hours are 8:00–22:00 and off-peak hours 22:00–8:00 the next day; cache-hit input after the discount is ¥0.1 at peak and ¥0.05 off-peak. Earlier, aiho.net and todayforai only reported "the vendor says limited-time 50% off" without stating whether ¥2/¥8 was the discounted price, which left two readings in play — that ambiguity is now resolved. The Qianfan console is authoritative, including whether the discount has an end date.
OpenRouter's free tier baidu/ernie-5.5-flash-preview:free costs nothing but is rate-limited to 20 requests per minute and 50 requests per day; buying 10 credits raises it to 1,000/day (verified by apidog on Nov 12). Fifty requests a day is enough to try it, not enough to run it as a workhorse — use the paid tier or Qianfan for heavier use.
China-based users paying in RMB who want the deep-thinking tier: Qianfan (model ID ernie-5.5-flash-preview, default limits 60 RPM / 1,000,000 TPM). Users paying in USD, wanting a free trial, or already on OpenRouter: OpenRouter — the paid $0.28/$1.13 tier converts to roughly Qianfan's peak rate, and the free tier is the way to validate the model first. It is a text-only model with a 1M context; for long documents, note that output tokens (max output 921,600) drive the bill.
Official pages: Baidu official release page · OpenRouter free-tier model page · PR Newswire release
Everything public from November 11 to 16: Chinese tier-1 coverage is thorough, English coverage is niche-media only — but independent scores are still zero. The headline first: the volume of evidence is sufficient, its independence is not here yet. Every entry carries a verdict badge; "Read on this site" opens our rewritten deep-dive article, with the original source at the end.
Source: Baidu official release page (~20 benchmark rows, all vendor-reported); 8 tests cross-checked via aiho.net, todayforai and madechango
The official release page lists about 20 benchmark rows; 8 are cross-confirmed by multiple secondary sources: Terminal-Bench 4.0 46.5 (Opus 5.5 63.6), AutomationBench 61.9 (No. 1 in that test; Opus 5.5 58.0), DeepSWE v1.1 71.2, DRACO 84.9 (third; Hy4 90.7, Opus 5.5 87.9), SpreadsheetBench V2 36.4, OfficeQA Pro V2 67.9, Agents' Last Exam (CLI) 29.0, APEX-Agents 1.1 64.2 (Hy4 75.0, Opus 5.5 72.9). Only 1 of the 8 is a win (AutomationBench); the rest trail the flagships, and the TB 4.0 gap to Opus 5.5 (46.5 vs 63.6) is wide.
Read on this site → (original ModelBrief rewrite)
Source: QbitAI reporting, via madechango / news.marsbit; behind Kimi K3.1 Pro, Qwen3.8-Max and other large Chinese models
CodeArena WebDev scores 1668 (±16), but on the Nov 11 board snapshot the row shows a blank rank, N/A votes and an AutoEval label — a placeholder used while human votes are still missing, not a ranking earned through battles; the license tag is Proprietary (weights not yet released, API-only). Arena said on Nov 12 that the model was "trending towards becoming the #2 open model", so "No. 2 open, No. 4 among Chinese models" is a forecast, not an established ranking. It is the only third-party board signal found, but it cannot be cited as a confirmed ranking.
Read on this site → (original ModelBrief rewrite)
Source: this site's check, Nov 16, 2026; aireiter: no standalone Preview row found in the AA list; europesays (Nov 5) also reports no AA page
As of November 16, 2026, Artificial Analysis has not listed ERNIE 5.5 Flash Preview (no AA Index score); no LiveBench or LMArena listing was found; no independent body has reproduced any benchmark score. The CodeArena 1668 is the only third-party signal, but only an AutoEval placeholder (zero human votes), with a Proprietary license tag.
Read on this site → (original ModelBrief rewrite)
Source: Baidu's vendor claim; OpenRouter measurement (Nov 12 data, free tier)
The vendor claims decoding at 200+ tokens/s; OpenRouter's measured median p50 is ~75–79 tok/s (Nov 12 data, free tier). The two figures differ by nearly 2× and come from different measurement bases (a vendor claim vs a third-party free-tier measurement), so this site states both and takes no side — measure it yourself.
Read on this site → (original ModelBrief rewrite)
Source: Baidu World 2026 (Beijing, Nov 12); IT之家 and others
Baidu World 2026 opened November 12 with Robin Li keynoting, and ERNIE 5.5 Flash Preview was presented alongside the Token Exchange (词元所), a marketplace for model services. The vendor says the exchange signed 100+ vendors in its first month, with 50M calls and 700B tokens per day — company-reported figures with no independent audit; this site reports them, it does not vouch for them.
Read on this site → (original ModelBrief rewrite)
Source: 新智元, aiho.net; madechango / 火星财经 hands-on, Nov 13
The official demos: the third-person wuxia game 《云岫行》 (UE 5.8 + C++, completed in a single 11-hour agent session), ~100 phone mini-games generated from design documents, and airport/airliner Blender modeling (dual-agent). All are end-to-end results inside Baidu's own agent framework — a different harness or prompting strategy could change results by an order of magnitude; and the model is text-only, so it cannot see the screen. madechango / 火星财经's Nov 13 hands-on verdict: "good taste, strong coding."
Read on this site → (original ModelBrief rewrite)
Source: this site's search, Nov 16, 2026
Chinese tier-1 tech media are all present: 新智元 (Nov 11 launch-day deep dive), 量子位, 每日经济新闻, 新华网, IT之家, 雷锋网, QbitAI; English coverage is niche tech media (Unite.AI, aiaffairs, artificialwatch, singularity.kiwi, freshsynapse, aireiter, apidog, digintoai), with no TechCrunch / The Verge / Reuters / Bloomberg / Wired piece. The strength of the Chinese coverage is the key difference from Nex-N2.6 (only 2 second-tier Chinese pieces).
The six articles below are original ModelBrief rewrites and reorganization — restructured with added conclusions and buying advice, with facts and numbers checked against the original reviews. Read them on this site; each ends with its original source.
The third-party boards are nearly empty: the only third-party score is the AutoEval-placeholder CodeArena 1668, and AA, LiveBench and LMArena have no listing. The vendor release page's roughly 20 benchmark rows are all self-reported — read the gap in independent numbers first, then the vendor table.
| Third-party board | ERNIE 5.5 Flash Preview | Note |
|---|---|---|
| CodeArena WebDev | 1668 | AutoEval placeholder (Nov 11 board snapshot: blank rank, N/A votes, ±16; license Proprietary; QbitAI via madechango); "No. 2 open, No. 4 among Chinese" is Arena's Nov 12 "trending towards" forecast, not an established ranking |
| Artificial Analysis Index | Not listed | No AA Index score as of Nov 16; no AA page |
| LiveBench | Not listed | No listing found |
| LMArena | Not listed | No listing found |
| Independent reproduction | None | No independently reproduced benchmark score |
| BrowseComp (web research) | 87.4% | Inference360 board, #15 of 62 (vendor-reported; read by this site Nov 19, 2026) |
| Test (vendor-reported) | ERNIE 5.5 Flash Preview | Main rivals (vendor chart, self-measured) |
|---|---|---|
| Terminal-Bench 4.0 | 46.5 | Claude Opus 5.5 63.6, GPT-6.1 Sol 56.1 |
| AutomationBench | 61.9 (No. 1) | Claude Opus 5.5 58.0, DeepSeek V4.1 Flash 57.2 |
| DeepSWE v1.1 | 71.2 | Rival scores not cross-confirmed |
| DRACO | 84.9 (3rd) | Hy4 90.7, Claude Opus 5.5 87.9 |
| SpreadsheetBench V2 | 36.4 | Rival scores not cross-confirmed |
| OfficeQA Pro V2 | 67.9 | Rival scores not cross-confirmed |
| Agents' Last Exam (CLI) | 29.0 | Rival scores not cross-confirmed |
| APEX-Agents 1.1 | 64.2 | Hy4 75.0, Claude Opus 5.5 72.9 |
Full tally: our benchmarks article →
| Dimension | ERNIE 5.5 Flash Preview | Claude Opus 5.5 | Tencent Hy4 |
|---|---|---|---|
| Terminal-Bench 4.0 (vendor-reported) | 46.5 | 63.6 | — |
| AutomationBench (vendor-reported) | 61.9 (No. 1) | 58.0 | — |
| DeepSWE v1.1 (vendor-reported) | 71.2 | — | — |
| DRACO (vendor-reported) | 84.9 (3rd) | 87.9 | 90.7 |
| APEX-Agents 1.1 (vendor-reported) | 64.2 | 72.9 | 75.0 |
| SpreadsheetBench V2 (vendor-reported) | 36.4 | — | — |
| OfficeQA Pro V2 (vendor-reported) | 67.9 | — | — |
| Agents' Last Exam (CLI, vendor-reported) | 29.0 | — | — |
| CodeArena WebDev (third-party, AutoEval placeholder) | 1668 | — | — |
| Artificial Analysis Index | Not listed | — | — |
| Total / active parameters | 280B / 14B | — | — |
| Context | 1M tokens | — | — |
| API price (input / output) | Qianfan ¥1/¥4 after 50% off (peak); OpenRouter $0.28/$1.13 | — | — |
| Weights / license | Not public (promised Nov 25); license accounts conflict | — | — |
The tier gap is the theme of this table: ERNIE 5.5 Flash Preview is a 14B-active lightweight execution layer against flagship tiers — TB 4.0 46.5 vs Opus 5.5's 63.6, DRACO 84.9 vs Hy4's 90.7, APEX-Agents 64.2 vs Hy4's 75.0, all trailing in the same direction, though within a low-double-digit gap; the single No. 1 is AutomationBench 61.9. Blank cells are comparable figures this research round could not verify; this site won't invent numbers. See our comparison article.
| Dimension | ERNIE 5.5 Flash Preview | ERNIE 5.5 Flash Base |
|---|---|---|
| Nature | Base's first post-trained version; an intermediate post-training checkpoint | Pre-training endpoint weights (the base model) |
| Released | Nov 11, 2026 | Nov 5, 2026 |
| Instruction following | Yes (a post-trained product) | No (base capability only) |
| Reasoning tier | Fixed high-effort tier, no toggle | — |
| Weights | Not public as of Nov 16; open-sourcing promised for around Nov 25 | Public (Hugging Face, FP8 + BF16) |
| License | Not stated on the official release page; several outlets say Apache 2.0 — unverified | Custom Baidu "ERNIE 5.5 Flash Base Model License Agreement v1.0," not Apache 2.0 |
| Pre-training data | Same base (33.5 trillion tokens) | 33.5 trillion tokens |
| Context | 1M tokens | 128K training, fine-tuning/inference supports 1M |
| API | Qianfan and OpenRouter live | No API (weights download) |
| Benchmarks | ~20-row vendor table on the official release page (all self-reported) | No official benchmark table |
The Base–Preview relationship is one sentence: the same 280B base, with Preview as its first post-trained version — different in capability, availability and license status, all three at once. Base can be downloaded but not chatted with; Preview can be chatted with and called via API, but its weights are not out and its license is unsettled. See our comparison article.
Read the deep comparison: ERNIE 5.5 Flash Preview vs Claude Opus 5.5 & Hy4 → · Read the deep comparison: ERNIE 5.5 Flash Preview vs ERNIE 5.5 Flash Base →
Most confusion about ERNIE 5.5 Flash Preview comes from six places: the open-source date, the license, how to read the scores, speed, the 50%-off price, and the CodeArena score — clarified one by one below.
November 25 is only a promise: as of November 16 the Preview weights have not landed (Hugging Face hosts Base only). And the vendor's own wording is inconsistent: the English release page says "by Nov 25," the Chinese WeChat post says "on Nov 25" — whether it is "that day" or "by that day" can only be settled once the weights actually land.
Several secondary sources say Apache 2.0, but aiho.net — which read the official release page directly — states plainly that "the license type has not yet been stated by the vendor." The Preview license is not confirmed by the official release page and must not be stated as fact; and Base ships under a custom Baidu license (not Apache 2.0), so a license change still needs the license file itself to prove it.
Of the 8 cross-confirmed benchmarks, ERNIE 5.5 Flash Preview wins only 1 (AutomationBench 61.9, No. 1); TB 4.0 46.5 trails Opus 5.5's 63.6, DRACO 84.9 trails Hy4's 90.7, and APEX-Agents 64.2 trails Hy4's 75.0. A 14B-active model reaching this range is itself the selling point — reading it as "beats the flagships" is a misreading.
200+ tok/s is Baidu's vendor claim; OpenRouter's measured median p50 is ~75–79 tok/s (Nov 12 data, free tier). The two figures use different measurement bases and differ by nearly 2×; this site states both and adopts no single speed number.
Verified: 51AllAI, citing Qianfan's model-service billing page, confirms that ¥2/¥8 is the original price and that the limited-time 50% discount brings peak hours to ¥1 in / ¥4 out and off-peak hours to ¥0.5 in / ¥2 out; peak hours are 8:00–22:00 and off-peak hours 22:00–8:00 the next day, and cache-hit input after the discount is ¥0.1 at peak and ¥0.05 off-peak. Earlier, aiho.net and todayforai only reported "the vendor says limited-time 50% off" without stating whether ¥2/¥8 was the discounted price, leaving two readings in play — that ambiguity is now resolved; the Qianfan console remains authoritative.
1668 comes from QbitAI reporting via madechango / news.marsbit — but the Nov 11 board snapshot shows a blank rank, N/A votes, an AutoEval label and a Proprietary license tag: 1668 is a placeholder, not a battle-earned ranking, and "No. 2 open, No. 4 among Chinese models" is Arena's Nov 12 "trending towards" forecast. The score has a source; "confirmed" it is not.
November 11, 2026 (official release page and WeChat post the same day); featured at Baidu World 2026 on November 12. Note the base model, ERNIE 5.5 Flash Base, came earlier — November 5 on Hugging Face. See our launch overview.
An intermediate checkpoint in post-training, self-described as unfinished: Baidu's own words are that Preview "may have capability gaps due to insufficient training," with the final version to follow; Sun Tianxiang, head of Baidu's foundation model R&D, said on stage that the team is working to "drop the Preview suffix as soon as possible." All benchmark scores are intermediate and may change with the final version. See our launch overview.
Qianfan: after the limited-time 50% discount, peak ¥1 in / ¥4 out and off-peak ¥0.5 in / ¥2 out per 1M tokens (original ¥2 in / ¥8 out; peak 8:00–22:00, off-peak 22:00–8:00 next day; cache-hit input ¥0.1 peak / ¥0.05 off-peak) — the console is authoritative. OpenRouter: paid tier $0.28/$1.13 (cache reads $0.028), free tier $0 (20 requests/min, 50 requests/day; 1,000/day after buying 10 credits). See our pricing guide.
No. The Preview weights are not public as of November 16, with open-sourcing promised for around November 25 (the English page says "by Nov 25," the WeChat post says "on Nov 25"); license accounts conflict — several outlets say Apache 2.0, the official release page does not confirm it, so verify against the license file when the weights land. Base weights are public, but under a custom Baidu license, not Apache 2.0. See our deployment and open-source outlook.
Base is the November 5 pre-training endpoint weight (the base model: base capability only, no instruction following); Preview is Base's first post-trained version (November 11), able to chat and be called via API. See our comparison article.
Trust them cautiously. All ~20 rows on the official release page are vendor-reported, with no independent reproduction; as of November 16 there is no AA, LiveBench or LMArena listing; the only third-party signal, CodeArena WebDev 1668, is an AutoEval placeholder (zero human votes). Of the 8 cross-checked tests it wins only 1 (AutomationBench), and TB 4.0 46.5 trails the flagships clearly. See our benchmarks article.
There is no single credible figure. The vendor claims 200+ tokens/s; OpenRouter's measured median p50 is ~75–79 tok/s (Nov 12 data, free tier). The two figures use different measurement bases, so this site adopts no single speed number — measure it yourself. See our benchmarks article.
Every fact on this page comes from the public sources below. This research round could not read the official release page directly (a browser policy block), so facts were cross-checked against secondary sources that read it directly. Unlinked items are cited by name and date from reporting this site read during research; no link is fabricated.
More indexes: All Baidu models · All Language Models · All models