Last updated Nov 20, 2026

ERNIE 5.5 Flash Preview vs Claude Opus 5.5 & Hy4: The Tier Gap, in Plain Sight

ModelBrief · Published 2026-11-11 · Fact-check 2026-11-16 · Reference: Baidu official release page, aiho.net, aireiter

Every number in this comparison comes from one chart: Baidu's official release page, which puts ERNIE 5.5 Flash Preview on the same table as Claude Opus 5.5, Tencent Hy4, DeepSeek V4.1 Flash and GPT-6.1 Sol. The headline first: the tier gap is plain to see — ERNIE 5.5 Flash Preview is a 14B-active lightweight execution layer, and its rivals are flagship tiers; of the four benchmarks with head-to-head scores, it wins 1. But the value of this comparison is not who wins — it is by how much: the gaps are all within low double digits, and a 14B-active model sitting that close to the flagship tier is itself a conclusion. All figures are vendor-reported (rival scores as the vendor chart lists them), not independently measured.

The four head-to-head tests

Test (vendor-reported)ERNIE 5.5 Flash PreviewClaude Opus 5.5Tencent Hy4
Terminal-Bench 4.046.563.6—
AutomationBench61.9 (No. 1)58.0—
DRACO84.9 (3rd)87.990.7
APEX-Agents 1.164.272.975.0

Test by test: Terminal-Bench 4.0 46.5 vs Opus 5.5's 63.6, a 17.1-point gap — the widest of the four, and the clearest sign that on terminal agent work the lightweight execution layer sits below the flagship tier; AutomationBench 61.9 vs Opus 5.5's 58.0, a 3.9-point lead, the only No. 1; DRACO 84.9 vs Opus 5.5's 87.9 and Hy4's 90.7, gaps of 3.0 and 5.8 points, third place; APEX-Agents 1.1 64.2 vs Opus 5.5's 72.9 and Hy4's 75.0, gaps of 8.7 and 10.8 points. Hy4's TB 4.0 and AutomationBench figures were not verified in this research round, so those cells are left blank rather than filled in.

What the AutomationBench No. 1 is worth

61.9, No. 1 in that test, against Claude Opus 5.5 (58.0) and DeepSeek V4.1 Flash (57.2) — leads of 3.9 and 4.7 points. That No. 1 has to be read two ways: on one side, it is the only No. 1 of the 8 cross-checked tests, showing real strength on automation-style tasks and matching the "execution layer for agent workloads" positioning; on the other side, it is the only No. 1 — TB 4.0, DRACO and APEX-Agents all trail, so it should not be read as an overall lead. One No. 1 plus three trails is the complete picture.

The other four tests: ERNIE's scores only

DeepSWE v1.1 71.2, SpreadsheetBench V2 36.4, OfficeQA Pro V2 67.9, Agents' Last Exam (CLI) 29.0 — these four have ERNIE 5.5 Flash Preview's scores only, with rival scores unconfirmed. They describe the model's absolute level (e.g., coding ability at DeepSWE 71.2, office ability at OfficeQA 67.9) but cannot rank it against anyone; this site does not fill in unverified rival scores. Two more stray rival figures appear on the vendor chart: GPT-6.1 Sol's TB 4.0 56.1 (ERNIE trails by 9.6) and DeepSeek V4.1 Flash's AutomationBench 57.2 (ERNIE leads by 4.7) — both also vendor-chart figures, not independently measured.

The gap: no independent scores in the same harness

The biggest gap in this comparison table is not Baidu's but the independent boards': as of Nov 16 ERNIE 5.5 Flash Preview has no AA Index score (not listed), and no LiveBench or LMArena listing; and same-harness AA Index, LMArena and LiveBench scores for Opus 5.5 and Hy4 were not verified in this research round, so this site leaves them blank. In other words, this comparison can currently only run inside the "vendor-reported" harness — rival scores are as Baidu's own chart lists them, not independently measured head-to-head runs. When AA lists it and independent reproductions appear, this comparison gets a second edition.

Who should pick it

Pick by tier, not by wins: for flagship-tier terminal agent ability, Opus 5.5 at TB 4.0 63.6 fits better; for top DRACO and APEX-Agents scores, Hy4 (90.7 / 75.0) fits better; for a 14B-active, 1M-context lightweight execution layer at Qianfan ¥2/¥8 (peak) or OpenRouter $0.28/$1.13, with automation-style work as the main scenario, ERNIE 5.5 Flash Preview is a reasonable pick — provided you accept the Preview intermediate state (scores may change with the final version) and the weights not being out yet (no self-hosting).

FAQ

Is ERNIE 5.5 Flash Preview stronger than Opus 5.5?

No. Of the four head-to-head tests it wins 1 (AutomationBench 61.9 vs 58.0); TB 4.0 46.5 vs 63.6, DRACO 84.9 vs 87.9 and APEX-Agents 64.2 vs 72.9 all trail. The tiers differ: it is a 14B-active lightweight execution layer, Opus 5.5 is a flagship tier.

And against Tencent Hy4?

DRACO 84.9 vs 90.7 and APEX-Agents 64.2 vs 75.0, both trailing; Hy4's TB 4.0 and AutomationBench figures were not verified in this research round. The same tier gap: Hy4 is a flagship MoE, ERNIE 5.5 Flash Preview a lightweight execution layer.

Are these comparison scores independently measured?

No. All figures are vendor-reported on Baidu's official release page, with rival scores as the vendor chart lists them; as of November 16, ERNIE 5.5 Flash Preview has no AA, LiveBench or LMArena score and no independent reproduction.

ModelBrief's take: the right title for this comparison is not "who won" but "by how much" — TB 4.0 behind by 17.1 points, DRACO by 3.0–5.8, APEX-Agents by 8.7–10.8, AutomationBench ahead by 3.9. A 14B-active model sitting that close to the flagship tier is the selling point; consistently trailing is the ceiling; both sentences hold at once, and that is ERNIE 5.5 Flash Preview's true position against its rivals.

Reference: Baidu official release page — ERNIE 5.5 Flash Preview · aiho.net — 280B deep dive · aireiter.com — API deep review (rewritten and organized by ModelBrief; facts per the official release page and reporting)