ERNIE 5.5 Flash Preview is a good cheap model with one real strength: speed. In our 8-task tournament it finished second of four, behind DeepSeek V4.1 Flash, at almost the same cost per answer ($0.0035 against $0.0032) and about three times faster (9.5 seconds against 30.5). Head to head it split 2-2 with DeepSeek, beat Qwen3.7 Plus 2-1, and left the pricier ERNIE-Light-DT-1.0 behind at a sixth of its cost.
| ERNIE 5.5 Flash Preview | ERNIE-Light-DT-1.0 | DeepSeek V4.1 Flash | Qwen3.7 Plus | |
|---|---|---|---|---|
| Input / output per 1M tokens | $0.28 / $1.13 | $1.13 / $3.94 | $0.30 / $1.20 | $0.32 / $1.28 |
| Pairwise wins (of 24 pairings each) | 5 | 2 | 7 | 2 |
| Answers completed | 8/8 | 8/8 | 8/8 | 8/8 |
| Avg cost per answer | $0.00350 | $0.0227 | $0.00315 | $0.00479 |
| Avg latency | 9.5 s | 61.4 s | 30.5 s | 60.7 s |
| Avg output tokens | 2,178 | 4,389 | 3,627 | 2,824 |
Prices come from OpenRouter's model list. Everything else is from our runs.
What Baidu shipped
Both models appeared on OpenRouter on 2026-11-10. ERNIE 5.5 Flash Preview is described there as the newest model in the ERNIE family, improved for coding, agentic work, knowledge work and deep research, with 280 billion total parameters. The context window is 1,024,000 tokens, and the listing shows tool calling, structured outputs and adjustable reasoning effort. Cached input is billed at a tenth of the normal input price, $0.028 per 1M tokens.
The second model, ERNIE-Light-DT-1.0, launched the same day. OpenRouter says it builds on open-source GLM models and adds Baidu Sparse Attention, a sparse attention mechanism from Baidu Search's Lightning model. It costs four times as much per token. We treat the pair as one release because they arrived together. Baidu earlier published open weights for a separate base model, which we have not tested because it has no hosted endpoint. For other models at this price tier, see our cheap LLM shootout.
How we tested
We wrote 8 prompts from everyday engineering and ops work: a failing Jest test, German invoice extraction to strict JSON, a Postgres job-queue deadlock, a window-function SQL query, a half-open interval merge, a refund email under policy limits, an agent deciding whether to ask or act before deleting files, and a monthly cost calculation with caching. We ran baidu/ernie-5.5-flash-preview, baidu/ernie-light-dt-1.0-intl, deepseek/deepseek-v4.1-flash and qwen/qwen3.7-plus through OpenRouter with a 3,500-token output cap. The judge was anthropic/claude-sonnet-5.5. It compared every pair on every prompt twice with positions swapped, and a win counts only when both orders agree; otherwise it is a tie. All 32 answers finished normally, with none cut off. Total spend for this post was $1.91 of a $5 cap, including an interrupted first attempt, and we ran it on 2026-11-10. Eight prompts is a small sample, so read it as a direction. The saved runs are in the repo under server/blog/runs/ernie-5-5-flash-preview-review-2026/.
Most pairs were ties. ERNIE Preview appeared in 24 pairings (8 prompts, 3 rivals), and the judge separated only 9 of them: 5 wins and 4 losses. That is the main caveat on the rankings above.
Where ERNIE Preview won and lost
It won the Postgres deadlock prompt against DeepSeek, Qwen and ERNIE-Light, and the refund email against DeepSeek and Qwen. On the deadlock, the judge's comparison with DeepSeek's answer began:
Both responses identify the same ABBA lock-order inversion and give the same core fix: move
UPDATE workersto the top of the worker transaction so both paths lockworkersbeforejobs.
The judge credited ERNIE's answer with a step-by-step cycle and a complete, runnable code block, while noting it left out one condition DeepSeek spelled out. On the refund email, the judge penalised Qwen for what ERNIE avoided:
However, it offers a replacement outside the 30-day window, which the policy does not authorize, so it invents an entitlement.
It lost the invoice extraction to DeepSeek and Qwen on a formatting detail. Both of those returned raw JSON, while ERNIE wrapped its answer in a markdown code fence, which bends "Return only JSON" and would break a strict parser. DeepSeek also won the cost-arithmetic prompt, where all models got the same $1,052 monthly total but DeepSeek kept more precision in its intermediate steps. Fence-wrapping is easy to strip, but if your pipeline parses output directly, check for it.
The odd result is the Jest prompt, where ERNIE Preview used 9,389 output tokens and 38 seconds before answering, far more than on any other prompt. The answer was correct and the judge scored it a tie. Its reasoning is adjustable on OpenRouter, so a lower effort setting is worth trying for short fix-it tasks.
Who should use it
Pick ERNIE 5.5 Flash Preview when latency matters at a budget price: support drafting, quick code explanations, and agent steps where waiting half a minute hurts. Pick DeepSeek V4.1 Flash when accuracy on strict-format tasks matters more than waiting, since it won 7 pairings to ERNIE's 5. Skip ERNIE-Light-DT-1.0 for general work; at $0.0227 per answer and 61 seconds it cost about six times as much as the Preview model for fewer wins. The label "Preview" also means Baidu can change the model under that ID.
If you would rather pay for stronger results, Gemini 4 Carbon beat Argon 5-0 in our hard-prompt test, and our DeepSeek V4 review covers the earlier model in DeepSeek's line.