ERNIE Light DT 1.0 vs Ling 3.1 Flash: A 25x Price Gap in the Same Weight Class

Two Chinese fast-tier models, both text-only, both advertising a million-token window, both built around sparse attention tricks to keep long context affordable — and one of them costs twenty-five times more per input token than the other. ERNIE Light DT 1.0, published by Baidu on 10 November 2026, lists at $1.13 per million input tokens and $3.94 per million output. Ling 3.1 Flash, Ant Group's 560-billion-parameter mixture of experts announced on 30 September 2026, lists at $0.045 and $0.135 on the same units and carries an independent score of 41 on the Artificial Analysis Intelligence Index.

That is not a modest gap to be explained away by context quality. It is the largest per-token spread between two models in this tier that either lab currently sells, and it demands an explanation before anyone signs a purchase order.

The cheapest way to see the problem: one request, two bills

Take a single call of the kind these models exist for — a 120,000-token retrieved document set with a 2,000-token answer, run once.

On ERNIE Light DT 1.0 that request costs 120,000 × $1.13/1M plus 2,000 × $3.94/1M, which is $0.1356 plus $0.0079, or about 14.4 cents.

On Ling 3.1 Flash the same request costs 120,000 × $0.045/1M plus 2,000 × $0.135/1M, which is $0.0054 plus $0.00027, or about 0.57 cents.

Both models claim to fit that request. One of them is twenty-five times cheaper to run it on. At ten thousand such requests a day the difference is $1,356 a day against $57 — roughly $39,000 a month against $1,700.

There are only three possible explanations. Either Baidu knows something about quality that it has not published, or the compression scheme is buying something Ling's architecture does not, or the price is a placeholder that will move. The rest of this piece is about which.

What each one actually is

• Ling 3.1 Flash — Ant Group's hybrid-reasoning MoE, roughly 560 billion total parameters with about 25 billion active per token, text in and text out, from the inclusionAI lab. Announced 30 September 2026 and live on Ant's own API and in its consumer products. The marketing claim is a 1M-token ceiling; what the endpoint actually served at launch was 256,000, and current international listings show a 1,000,000-token context with 262,144 tokens of maximum output.

• ERNIE Light DT 1.0 — Baidu's small-tier endpoint, listed 10 November 2026, built on open-source G​LM architecture with research from Baidu Search's "Lightning" line, using Baidu Sparse Attention and a feedback-driven history compression scheme. 1,024,000-token context, 131,072-token maximum output, text only. Four reasoning effort levels, tool calling, structured outputs.

The parameter counts are not comparable — Ling has not published Light DT's, so the honest statement is that one side is known to be 560B total and the other is undisclosed. What is comparable is the shape: both are sparse, both are text-only, both are aimed at long-context, high-volume work.

Sparse attention, twice, for two different reasons

The architecture stories rhyme, which makes the price gap harder rather than easier to explain.

Ling 3.1 Flash pays for 25 billion of its 560 billion parameters on each token. That is a capacity play: keep a very large knowledge store resident and route narrowly through it, so you get big-model world knowledge at small-model compute. The long-context claim comes from the same design — if each step is cheap, holding a long sequence open is less of a burden.

ERNIE Light DT's story is about memory rather than capacity. Baidu says the model "improves the storage and indexing of historical information" through feedback-driven history compression, "significantly reduces KV cache storage and computational costs while maintaining model performance," and lets you hold longer contexts "at a lower cost." Note the phrasing: lower cost to the operator. That is a claim about Baidu's serving bill, not necessarily about yours.

This is the crux. Every compression scheme in this class — Baidu's Heavily Compressed Attention in ERNIE 5.5, Q​wen's micro-block sparse attention, Ling's own routing — reduces the vendor's cost per token of context. Whether the vendor passes that saving on is a commercial decision, and Baidu has chosen not to. It is charging a premium on precisely the capability it says it made cheaper.

That is not automatically bad faith. A compression scheme that preserves quality at a million tokens is expensive to build and expensive to validate, and the vendor may be recovering that research cost. But it does mean the buyer is paying for a capability whose economics were supposedly improved, and the improvement is not visible in the rate card.

Ling's independent score, and what it says about this tier

Ling 3.1 Flash scores 41 on the Artificial Analysis Intelligence Index. That is a real third-party measurement, not a vendor chart, and it puts the model in the upper half of the fast tier — comfortably above models with larger marketing budgets, below the frontier small models. The score comes with a caveat that matters for a cost comparison: Ling is a talkative reasoner. Prior evaluation of the model found it burning roughly 220 million output tokens working through the index evaluations, well above the median for its price tier, and averaging several minutes per task. Cheap tokens that the model spends freely can still add up — but at $0.135 per million output, even generous reasoning stays inexpensive.

ERNIE Light DT 1.0 has no score at all. Not a vendor number, not an independent one. Baidu published a rate card and a paragraph of mechanism copy, and stopped. For a model priced at a premium, arriving with zero evidence is the single most damaging fact in this comparison — not because it is likely to be bad, but because a buyer cannot distinguish "unmeasured" from "uncompetitive," and the price gives no benefit of the doubt.

The context ceiling is the one place Baidu has room to argue

If there is a case for the premium, it is here. Ling 3.1 Flash's million-token window has been, at least through its launch period, a ceiling rather than a served figure — the endpoint gave 256K while the announcement advertised 1M. ERNIE Light DT advertises 1,024,000 and its listings show the same figure for the context and 131,072 for output, with no documented reduction.

Whether that distinction survives contact with real traffic is unknown, because Baidu has published no long-context evaluation. Neither has anyone else. The only long-context evidence in either family comes from ERNIE 5.5 Flash Base's vendor table, which reports HELMET-RAG 71.5, HELMET-ICL 85.9, LongBench-v2 48.7 and InfiniteBench 46.5 — Baidu's own numbers, on a different and much larger model, and not transferable to this one.

So the argument for Light DT reduces to: an unmeasured model with an unverified long-context advantage, at twenty-five times the input price of a measured model that at least has a real index score and a working endpoint. That is not a strong position, and it is fair to say so plainly.

What each is genuinely better at

• Output ceiling — Light DT at 131,072 tokens against Ling's 262,144. Ling gives you twice the room to generate, which matters for long-form synthesis and agent transcripts.

• Reasoning control — Light DT exposes max, high, low and none with max as the default. Ling is a hybrid reasoner with an on/off switch rather than a graded ladder; there is no documented "none" that gets you a fast non-thinking path.

• Cost at any volume — Ling at every size of request. There is no workload in which a 25x input-price premium is recovered by a context advantage of two percent (1,024,000 against 1,000,000 tokens).

• Evidence — Ling, in both directions: an independent index score and a published architecture with parameter counts. Light DT has neither.

• Serving stability — Ling is on Ant's own infrastructure plus third-party platforms; Light DT is on Baidu's international endpoint alone. A single-provider model has no failover, which is a procurement risk rather than a performance one.

• Openness — Ling's 1019 checkpoint was released as open weights in late October 2026, so the architecture is inspectable and the model can be self-hosted. Light DT has no weights and no model card.

That last point deserves emphasis. The Ling line has an open artefact and a closed one, and the open one lets you check the claims. ERNIE Light DT offers only the closed option, on a model whose central claim — that compression preserves quality at long context — is exactly the kind of thing you would want to verify independently.

If you still want to try it

Neither model is on OrcaRouter's catalogue, so this is not a recommendation to sign up anywhere. But if Light DT's long-context story is relevant to your work, the way to test it is narrow and cheap.

Send the same 100,000-plus-token prompt to both models, twenty times, with questions whose answers are buried at different depths — near the start, in the middle, and at the end. Score the answers yourself. That test costs under a dollar on Ling and about three dollars on Light DT, and it settles in an afternoon what neither vendor has published. If Light DT holds quality at the deep end where Ling degrades, the premium has a case. If both degrade, you have learned that a million-token window is a marketing number in this tier regardless of who sells it.

Whatever the outcome, keep the swap cheap. A routed endpoint that exposes both models' prices per request — with the option to send a fraction of your traffic to the new one and pull it back the same day — is the only sane way to evaluate a model with no published evidence. Buying capacity on a rate card alone is how teams end up paying 25 times over for a window they never fill.