BaiduNovember 11, 2026

ERNIE 5.5 Flash Preview

Baidu ERNIE 5.5 Flash Preview

Baidu's post-trained ERNIE 5.5 Flash Preview released Nov 11, 2026: 280B/14B MoE, 1M context, 200+ tps; BrowseComp 87.4, DeepSWE v1.1 71.2, AutomationBench 61.9; CNY 2.0/8.0 per 1M tokens on Qianfan.

PreviewText inputText outputOpen weights announced · release pending

At a glance

Context window
1M tokens
Max output
900K tokens
Text input / output · per 1M tokens
— / —
Knowledge cutoff
—

Benchmark snapshot

Best result in 6 of 7 categories

  • AcademicHLE

    46.4

    Thinking High (No Tools) · #61 of 151

  • CodingNL2Repo-Bench

    64.8

    Thinking High (With Tools) · #5 of 21

  • Computer UseMCP-Atlas

    84.6

    Thinking High (With Tools) · #8 of 55

  • Long ContextAA-LCR

    79.0

    Thinking High (No Tools) · #16 of 22

  • ProfessionalSpreadsheetBench 2

    36.4

    Thinking High (With Tools) · #1 of 3

  • Research & SearchBrowseComp

    87.4

    Thinking High (With Tools + Internet) · #10 of 61

View benchmark details (15)

Benchmark details

15 results · Best published score per benchmark, grouped by category

Benchmark · settingScoreRank
Academic
HLE

Thinking High (No Tools) · Knowledge Exams

46.461 / 151
Terminal-Bench-Science 0.1

Thinking High (With Tools) · Scientific Computing

21.020 / 32
Coding
DeepSWE

Thinking High (With Tools) · Repository Engineering

71.222 / 73
NL2Repo-Bench

Thinking High (With Tools) · Repository Engineering

64.85 / 21
ProgramBench

Thinking High (With Tools) · Code Generation & Editing

29.510 / 23
Terminal-Bench 4.0

Thinking High (With Tools) · Agentic Development

46.521 / 90
Computer Use
Agents' Last Exam v1.1

Thinking High (With Tools) · pass-rate/claude-code · Tool Orchestration

29.014 / 17
APEX-Agents

Thinking High (With Tools) · Tool Orchestration

64.213 / 61
AutomationBench-AA

Thinking High (With Tools) · Tool Orchestration

61.913 / 18
MCP-Atlas

Thinking High (With Tools) · Tool Orchestration

84.68 / 55
Toolathlon-Verified

Thinking High (With Tools) · Tool Orchestration

75.011 / 39
Long Context
AA-LCR

Thinking High (No Tools) · Long Reasoning

79.016 / 22
Professional
SpreadsheetBench 2

Thinking High (With Tools) · Data Analysis

36.41 / 3
Research & Search
BrowseComp

Thinking High (With Tools + Internet) · Fact Finding

87.410 / 61
Safety & Alignment
AA-Omniscience

Thinking High (No Tools) · Honesty & Factuality

25.211 / 16
5 more results in Computer Use, Long Context, Professional and 2 more

Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.

Back to benchmark snapshot ↑

Specifications

Provider
Baidu
Model type
Reasoning model
Released
November 11, 2026
Total parameters
280B
Active parameters
14B
Architecture
Mixture of experts
Context window
1M tokens
Max output
900K tokens
Input
Text
Output
Text
Weights
Open weights announced · release pending
Licence
Unspecified (weights release pending)
Commercial use
Not documented

Capabilities

ReasoningMultilingualCodingFunction callingTool useStructured outputLong context

Model analysis

ERNIE 5.5 Flash Preview (ernie-5.5-flash-preview) is Baidu's first post-trained preview model in the ERNIE 5.5 Flash family, released on November 11, 2026, built on ERNIE-5.5-Flash-Base for high-throughput reasoning, agentic coding, 3D modeling, deep research, and end-to-end knowledge work.

Architecture and Specifications

ERNIE 5.5 Flash Preview uses a hybrid-attention sparse MoE architecture with 280B total parameters and 14B active parameters per token, supporting a 1M-token (1,024,000) context window and up to 921,600 output tokens (64K default on Qianfan). It supports built-in reasoning (defaulting to high effort), tool calling, and structured outputs, achieving 200+ tokens/s single-user decoding on Qianfan via FP4/FP8 expert weights, HCA + MLA-DSA hybrid attention, and ECHO speculative decoding.

Public Benchmark Results

In Baidu's official release appendix (ernie.baidu.com/release/ernie-5-5-flash-preview/), evaluated at the highest available reasoning setting across 20 benchmarks:

  • Agentic Coding (7): Terminal-Bench 4.0 46.5, DeepSWE v1.1 71.2, ProgramBench (Almost Solved) 29.5, NL2Repo-Bench 64.8, Terminal-Bench-Science 0.1 21.0, CookieBench 88.2, and BlenderWorld (text-only) 79.0.
  • Knowledge Work (5): AutomationBench 61.9, APEX-Agents 1.1 64.2, DRACO 84.9, SpreadsheetBench V2 36.4, and OfficeQA Pro V2 67.9.
  • General Agent (5): Agents' Last Exam (CLI) 29.0, SkillsBench 1.1 (text-only) 73.5, Toolathlon-Verified 75.0, MCP-Atlas 84.6, and BrowseComp 87.4.
  • Reasoning & Long Context (3): HLE (w/o tools, text-only subset) 46.4, AA-Omniscience-Public 25.2, and AA-LCR v1.1 79.0.

API Pricing and Open-Source Plan

ERNIE 5.5 Flash Preview is live on Baidu Qianfan (Endpoint ID: ernie-5.5-flash-preview) and OpenRouter (baidu/ernie-5.5-flash-preview:free). Qianfan standard pay-as-you-go list pricing is CNY 2.0 input / CNY 8.0 output / CNY 0.2 cache read per million tokens during peak hours (08:00–22:00 Asia/Shanghai), and CNY 1.0 input / CNY 4.0 output / CNY 0.1 cache read during off-peak hours (22:00–08:00), with a 50% launch promotion and Token Plan discounts. Baidu also announced that post-trained weights will be open-sourced under Apache 2.0 soon.

Sources

Last reviewed 2026-11-16

Questions

What are the parameter count and inference speed of ERNIE 5.5 Flash Preview?
ERNIE 5.5 Flash Preview is a 280B total / 14B active sparse MoE model with a 1M-token (1,024,000) context window, up to 921,600 output tokens, and 200+ tokens/s single-user decoding on Baidu Qianfan.
How does ERNIE 5.5 Flash Preview score on benchmarks?
In Baidu's official 20-benchmark table, it scores 87.4 on BrowseComp, 84.6 on MCP-Atlas, 84.9 on DRACO, 71.2 on DeepSWE v1.1, 61.9 on AutomationBench, 46.5 on Terminal-Bench 4.0, and 46.4 on HLE (w/o tools, text-only subset).
What is the API pricing for ERNIE 5.5 Flash Preview?
Baidu Qianfan lists CNY 2.0 input / CNY 8.0 output / CNY 0.2 cache read per million tokens during peak hours (08:00–22:00) and CNY 1.0 / 4.0 / 0.1 during off-peak hours (22:00–08:00), with a 50% launch discount; OpenRouter also offers a free preview endpoint.
Will ERNIE 5.5 Flash Preview be open-sourced?
Yes. Baidu announced on its release blog that the post-trained weights will be open-sourced under the Apache 2.0 license soon, following the November 5 release of ERNIE-5.5-Flash-Base.

ERNIE 5.5 Flash Preview vs competitors

Comparable results only · record is ERNIE 5.5 Flash Preview's wins–losses

ERNIE 5.5 Flash Preview benchmark scores against competitors
BenchmarkERNIE 5.5 Flash PreviewThis modelvs Hy4 FlashRecord 3–1
HLE46.455.2
Terminal-Bench 4.046.519.2
DeepSWE71.264.6
APEX-Agents64.263.1

Lineage

Predecessors and successors · oldest to newest

  1. ERNIE 5.0Nov 13, 2025
  2. ERNIE 5.0 (0110)Jan 10, 2026
  3. ERNIE-5.1-PreviewApr 30, 2026
  4. ERNIE 5.5 Flash PreviewThis model · Nov 11, 2026vs ERNIE 5.5 Flash Base

Price & benchmark trends

Across curated predecessors and successors, ordered by release date. Points use listed rates and published results, not a history of price changes within one model. Related branches may appear; adjacent points do not imply direct succession. Undated models appear last.

API price across generations

Text inputText outputUSD / 1M tokens
$0$1.83$3.65ERNIE-4.5-300B-A47B-BaseERNIE 5.0ERNIE 5.0 (0110)ERNIE-5.1-PreviewERNIE 5.5 Flash BaseERNIE 5.5 Flash Preview

Input and output share one linear USD axis · uncached text rates · lower is cheaper. Retired models may retain their last recorded price. Rate tiers and context limits can differ; inspect each model's pricing conditions.

Orange marks this model. Hover, tap or focus the chart to see exact values. Missing values leave gaps. Horizontal spacing shows release order, not elapsed time.

Listed prices and selected benchmark scores across model generations
Model · releasedInput / 1MOutput / 1MPrice effective
1ERNIE-4.5-300B-A47B-BaseJun 30, 2025——Not recorded
2ERNIE 5.0Nov 13, 2025$0.840$3.382026-09-15
3ERNIE 5.0 (0110)Jan 10, 2026——Not recorded
4ERNIE-5.1-PreviewApr 30, 2026$0.560$2.532026-11-13
5ERNIE 5.5 Flash BaseNov 5, 2026——Not recorded
6ERNIE 5.5 Flash PreviewNov 11, 2026 · This model——Not recorded

Community discussion

Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.

Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.

No community comments yet.