Gemini Decision 1 vs Laya: 262,144 Tokens Rented Against 512 Tokens Owned

Two decision models, two numbers that describe the same design choice from opposite ends. Gemini Decision 1 takes a 262,144-token input and up to 1,000 typed questions in one call, at $0.10 per million input tokens, on Google's own API. Laya takes 512 tokens of text on its English checkpoint, runs on a T4 you already pay for, and is Apache-2.0. That is a factor of 512 on the window and infinity on the price, in opposite directions, and no benchmark rescues the comparison because the two models are not built for the same job.

Convai Innovations published Laya on 18 September 2026 as three checkpoints in one repository: a 421M ModernBERT-large English model, a 322M mmBERT-base multilingual model, and a typed-decisions variant. Google released Gemini Decision 1 on 11 November 2026 and did not publish weights, a parameter count or a base model. The only number both can be placed against is a scoreboard that carries just one of them — Gemini Decision 1 sits at 66 on the Artificial Analysis Decision Index v2.1, ninth of eighteen, and Laya is not on that board.

What a 512-token window rules out before breakfast

Start with the constraint, because it eliminates entire product categories faster than any accuracy figure.

• Shared input — Gemini Decision 1 reads up to 262,144 tokens covering the input and every question. Laya's English checkpoint reads 512 tokens, full stop. The multilingual checkpoint defaults to 1,024 and can be pushed to 8,192 with max_len=8192, at which point Convai's own card reports 16 to 18 of 20 long-document requests correct at up to roughly 4,000 tokens of context and "beyond that results vary".

• Modality — Gemini Decision 1 accepts text, image, video, audio and PDF in the same request. Laya accepts text. A scanned claim form, a call recording or a product photo is not an input until something upstream has converted it into words, and that conversion is a separate model with a separate cost and its own errors.

• Questions per call — up to 1,000 for Gemini Decision 1, isolated from one another, in one pass. Laya answers the questions you send with one state, and its own documentation sets the practical ceiling around fifty options per choice question before the model's per-option token budget gets thin.

• Price — $0.10 per million input tokens for Gemini Decision 1, measured at $345 per million tasks by Artificial Analysis, with no context caching and no Batch API. Laya is a download: roughly 808 MB for the English checkpoint, no per-call charge, and a T4 running at about 32.8 milliseconds per decision.

• Licence and distribution — Apache-2.0 weights, no hosted dependency, and a pip install laya entry point. Gemini Decision 1 is available only on the Gemini Developer API at v1beta, with a free tier at 200 requests per minute and a documented list of supported and unsupported capabilities.

The scoreboard has one of these two on it, and that is not a tie

Gemini Decision 1's 66 is a composite of six weighted evaluations, 45% of which is built on private data, scored as accuracy minus the Brier penalty attached to the answer. Its strengths and weaknesses are unusually clean for a launch-day model: first of eighteen on AA-Omniscience-Decision at 66.3 and first on HLE-MultiChoice at 48.7, but ninth on AA-Paperclip at 56.8 with a 33% all-or-nothing document pass rate, and fourteenth on When2Call at 80.5. Ninth overall, first where knowledge and abstention are tested, mid-table where extraction and tool choice are.

Laya has no comparable figure, and the project says why in its own card. The base checkpoints score 0.362 zero-shot on the typed-decisions benchmark — below the 0.461 majority-class baseline, and near random. The headline 0.766 requires fine-tuning on that benchmark's own training split. Convai's summary sentence is the one to carry into any evaluation: Laya is a fast base to specialise, not a zero-shot decision engine.

That is not a weakness to be scored against a frontier lab's model; it is a different product description. Gemini Decision 1 is sold as something that is already good at decisions. Laya is sold as something that becomes good at your decisions after you train it, at which point its published numbers — 0.766 against a sibling's 0.727 on typed decisions, 0.950 against 0.910 on AG News — are competitive on the tasks it was fitted to.

The two pipelines people actually build

Because the capabilities barely intersect, the honest framing is not "which one" but "which of these two jobs do you have".

The job Gemini Decision 1 fits. A long input arrives — a 60-page contract, an hour of call audio, a folder of scanned invoices — and many judgements have to be attached to it. Forty policy rules over one ticket in a single billed read; a rubric scored across a hundred candidate answers; a routing decision that depends on evidence spread across a document nobody wants to chunk. Google's own documentation names exactly these: request routing, policy checks across many rules at once, candidate scoring against a rubric, tool selection, low-latency classification over live media. The 1,000-question ceiling and the multimodal input are the whole reason the product exists, and they are the two things Laya cannot do at any price.

The job Laya fits. A short state and one decision, at a volume where the per-call cost is the constraint and the label set is yours. Ticket triage to your own taxonomy, intent classification, a yes/no gate before an expensive model runs. A 421M encoder on a T4 at 32.8 milliseconds is roughly free per call, and because the answer space is assembled per request rather than baked into a vocabulary head, a schema you invent this afternoon needs no retraining to be accepted — though it does need training data to be accurate.

The combination is not exotic: Laya as the fast first stage that handles the 80% of traffic it was trained for, and a large hosted scorer for the items that need evidence from a long or non-text input. That pattern costs less than sending everything to Google and fails more gracefully than running a 421M encoder on a contract.

Where the cost comparison actually lands

At low volume, Laya's economics are unbeatable and irrelevant. A T4 is a few hundred dollars a month whether you use it or not; a million Gemini Decision 1 tasks at $345 is a line item somebody has to approve. Below a few hundred thousand decisions a month, the hosted endpoint usually wins on engineering hours alone — there is no training run, no label set and no serving stack.

At high volume the arithmetic flips, with one caveat that matters more than the crossover point: Laya's marginal cost is near zero only if the model already performs on your task. If it does not, the labels are the real bill. Convai's card is explicit that fine-tuning is what produces the published numbers, so the honest comparison is $345 per million hosted tasks against a GPU plus the cost of labelling enough examples to reach a calibration you trust. Neither of those is a per-token rate, and a comparison that pretends otherwise is comparing a rental price to a construction project.

Then there is the latency shape, which is where the two products stop being substitutes at all. Gemini Decision 1 measures 211 milliseconds per task on the index — fine for a routing hop behind a customer message, marginal for a frame-by-frame loop. Laya's 32.8 milliseconds on a T4 is the number that makes per-frame and per-keystroke decisions possible, and it is the one advantage no amount of Google infrastructure closes, because it comes from being 421M parameters rather than a frontier-scale multimodal model.

Routing the part that is neither

OrcaRouter carries neither model. Laya is Apache-2.0 weights you run yourself, and Gemini Decision 1 is a probabilities-only endpoint on Google's own API — a decision model is not a chat-completions target, and we do not list what we cannot serve. The decision model in the catalogue is TypeSafe's typesafe/jev-1.14, at 66.23 on the same index, if a scoring call on one key matters to you.

The call that does belong behind a router is the one both of these models depend on and neither performs: the conversion step. Laya's 512-token window means an upstream model has to turn the document, the recording or the screenshot into a short state before Laya can judge it, and Gemini Decision 1's cost per task is dominated by how much context you hand it. Choosing and swapping that upstream model is the recurring decision in this pipeline, and it is the one we are built for: more than 200 models behind one OpenAI-compatible key, provider list price passed through at 0.0% markup so a vendor cut is live the same day, and automatic failover when a provider degrades. Write once against the router, and the encoder you own and the scorer you rent both stay replaceable.

What to do this week

If your input is longer than a page, contains anything that is not text, or needs more than a handful of questions answered about it, Laya is not a candidate and the decision is between Gemini Decision 1 and its hosted peers on price and pass rate. Google's documentation lists what the model supports in one paragraph; read it before you design around it, because the unsupported list — no caching, no batch, no Flex or Priority, no media inside question instructions — is longer than the supported one.

If your input is a sentence or a short JSON blob and your label set is your own, the hosted model is the expensive way to do it. Download Laya, label 500 examples, fine-tune, and check the calibration before you trust the confidences — that is the workflow its authors designed for, and their own card is unusually clear about the gap between the base checkpoint and the fine-tuned one.

What you should not do is compare 66 against a number from Laya's card. They were produced on different instruments, one of them is a zero-shot score below the majority baseline, and the other is a weighted composite of six private-data evaluations. The only comparison worth running is on your own labelled items, and only one of these two models can be run that way at all.