ERNIE 5.5 Flash vs MiniMax M3.1 Flash: Two Million-Token Windows, One Published Score

Both of these models claim a million-token context window, both are Flash-tier, both cost less than a dollar per million input tokens, and exactly one of them has a number attached to its name. MiniMax M3.1 Flash has been scored at 42 on the Artificial Analysis Intelligence Index since its 19 October 2026 launch, at $0.22 per completed index task. ERNIE 5.5 Flash went on sale at Baidu Qianfan on 10 November 2026 with a price, a specification and no evaluation of any kind — not a vendor table, not an independent run, not a knowledge cutoff. So the useful question is not which model is better. It is how much the missing score is worth, and whether the two windows are the same window.

Where the two specs actually diverge

The headline figures are close enough to look interchangeable, and they are not. Four of the rows below decide real workloads.

• Input window — ERNIE 1,024,000 tokens vs MiniMax 1,000,000. A 24,000-token difference is noise; treat these as equivalent and stop comparing them.

• Output ceiling — ERNIE 921,600 tokens vs MiniMax 524,288. ERNIE's ceiling is nearly double, and the reason is architectural rather than generous: its reasoning cannot be switched off, so long agent loops need somewhere for the thinking to go.

• Price per million — ERNIE $0.28 in / $1.13 out / $0.028 cached vs MiniMax $0.15 / $0.60 / $0.015. ERNIE is 1.87× the input rate and 1.88× the output rate. Cached reads scale the same way.

• Modality — ERNIE is text only. MiniMax accepts text, image and video and returns text. For any workload containing a screenshot, a chart, a UI capture or a recorded call, this row ends the comparison before price is discussed.

• Reasoning control — ERNIE's reasoning is mandatory, enabled by default, offered at one effort level: high. MiniMax exposes tunable thinking depth, which means a classification call can skip the thinking it does not need.

• Weights — MiniMax published 428B parameters (23B active) on 23 October, 252.4 GB across 48 shards under a community licence with a $20 million revenue line. ERNIE's base checkpoint is downloadable; the post-trained ERNIE 5.5 Flash is not, and no date has been given.

• Independent score — MiniMax 42 on the Intelligence Index, 41st of 173 at time of scoring, $0.22 per task. ERNIE: none. Artificial Analysis carries ERNIE 4.5 and ERNIE 5.0 rows and no ERNIE 5.5 row at all.

Run the bill three times, not once

A rate card is the least useful way to compare two models that both think before they answer. The same task, priced three ways, is more honest.

Take a document-review workload: 400,000 input tokens, 3,000 output tokens per request, a hundred requests a day. On ERNIE 5.5 Flash that is 40 million input tokens at $0.28 and 300,000 output tokens at $1.13 — $11.20 plus $0.34, so $11.54 a day. On MiniMax M3.1 Flash it is $6.00 plus $0.18, or $6.18. ERNIE costs $5.36 more per day, or about $1,950 a year, for the same 100 documents.

Now repeat it with a system prompt worth caching. If 380,000 of those input tokens are stable instructions and only 20,000 change, ERNIE's cached reads at $0.028 turn the input line from $11.20 into roughly $1.62; MiniMax's $0.015 cached rate takes its $6.00 to about $0.87. The gap narrows to well under a dollar a day, and both models become cheap enough that the choice stops being financial. Caching is not a micro-optimisation on a million-token window — it is most of the bill.

The third run is the one nobody publishes. Add reasoning tokens. ERNIE cannot be asked to skip them; on a task where it thinks for 6,000 tokens before writing 3,000, output volume is 9,000 tokens and the output line triples. MiniMax can be told to think less, and its documented behaviour is already unusually verbose — 180 million output tokens across the Intelligence Index against a median of 99 million — so the two may converge in practice. Nobody knows, because only one of these models has been measured at all. Until you count output tokens on your own prompts, the second decimal place of any comparison here is decoration.

Mandatory reasoning is a design position, not a bug

It is tempting to read ERNIE's mandatory high-effort reasoning as a limitation, and for some workloads it is exactly that — there is no fast path for a one-word classification, and every request pays a latency and token tax that MiniMax lets you avoid. But the same decision explains the output ceiling, and it says something about who Baidu built this for.

A model that always reasons, with a 921,600-token output allowance, is built for long agentic runs: repository-scale refactors, multi-step research, planning loops that need somewhere to think. Baidu's base checkpoint release leans the same way — the pre-training corpus emphasises "repository-scale code and long-context," and the benchmark rows that flatter it most are HumanEval+ at 82.3 and RepoQA at 91.7, both base-model figures that say nothing directly about the chat model but do say what the training data prioritised.

MiniMax's design points elsewhere. Tunable thinking depth, native image and video input, 125.5 output tokens per second and a 1.04-second time to first token describe a model built for throughput — autonomous loops, screenshot-driven interfaces, document pipelines where the input is a scan rather than a diff. Its own positioning language is about cost per token at volume, and the $0.22 per index task is the number that backs it up.

Two Flash-tier models, two different bets: one on depth per request, one on requests per second. Neither is wrong, and the specification sheets are unusually clear about which is which.

Both are cheap; only one is a decision you can make today

The asymmetry that matters is evidentiary. MiniMax M3.1 Flash has a published index score, a published per-task cost, a published model card, downloadable weights and an independent measurement of how many tokens it burns to finish a benchmark. A procurement decision can be made on that. ERNIE 5.5 Flash Preview has a rate card, a context window, and a Baidu press page that still leads with the base checkpoint and the phrase "coming soon" about the post-trained model — which is, awkwardly, the model now being sold.

That does not make the preview a bad model. It makes it an unmeasured one, and the correct posture toward an unmeasured model at twice the price of a measured competitor is not to assume it is worse. Baidu's base checkpoint beat DeepSeek V4 Flash Base on 26 of 30 rows; a strong post-training run on that base could plausibly land above 42. It could also land below. The point is that nobody outside Baidu can currently tell, and a rate card is not evidence.

Testing both without signing two contracts

One of these models is on a routed key and one is not. OrcaRouter serves MiniMax M3.1 Flash at provider list price with 0% markup, alongside GLM-5.3-Flash, DeepSeek V4.1 Flash, Qwen 3.8 Flash and Gemini 3.8 Flash — more than 200 models behind a single OpenAI-compatible endpoint, with automatic failover across providers and a routing DSL for composing several models into one call. ERNIE 5.5 Flash Preview is not in that catalogue; there is no ERNIE entry and no Baidu endpoint for us to route to, so it has to be called on Baidu's own API with a second key and a second bill.

The practical route through that asymmetry is to build the evaluation first and the integration second. Put your twenty prompts behind the routed key, run MiniMax and the rest of the Flash tier, and record output tokens per task — not just price per million. Then point the same twenty prompts at the Baidu endpoint and compare on the same axis. If ERNIE wins on the tasks you care about, its 1.88× rate is a premium worth paying and the missing model card becomes a nuisance rather than a blocker. If it does not, you have spent an afternoon and one small invoice finding out.

What would settle it properly is a single row in a tracker that runs its own harness. MiniMax has had one since launch day. Baidu could publish one tomorrow and this article's central uncertainty would evaporate — which is the tell that the uncertainty is not really about the models at all. It is about which vendor chose to show its work.