AI Brief, 14 October 2026: a million-token output budget that arrives capped at 65,536

Gemini 4 Argon can now be called. Google announced it on 30 September with a price, state-of-the-art claims and no general availability, rolling out only to "a set of trusted cyber defenders" through its Fairwind Program. The question this brief put down on 1 October was when it would become callable, and at what effort tiers. The answer is 13 October: Argon appeared on OpenRouter at 16:04:37 UTC under the canonical slug google/gemini-4-argon-20261013, and the Hacker News submission for Google's model page took 1,137 points the same day.

Three things the announcement left open are now settled. The context window, which Google's post never stated at all, is 1,048,576 tokens. The price is $2 per million input tokens and $10 per million output, which is still the introductory rate rather than the $4 and $20 it was said to rise to. And reasoning effort is a request parameter, so the single "high" setting the model was measured at is not the only one exposed.

The number that does not survive the transition is the one the announcement led with. Google raised Argon's output budget to 1 million tokens from 64K, and described it as letting the model "generate hundreds of thousands of tokens in a single trajectory". On the first third-party surface to carry the model, the standard route reports a maximum completion of 65,536 tokens, the old figure, while only the batch tier, at half the token price, goes anywhere near the claim, at 943,718. Artificial Analysis reported at announcement that reaching a million output tokens depended on Long Decode Continuation, an undocumented feature that pauses a long response and resumes it across calls. Whatever Google's own API permits, the headline budget is not what a developer gets by calling the model the ordinary way.

Google also shipped Gemini 4 Argon for Robotics the same afternoon, at 206 points on Hacker News, plus a companion post on the model's safety evaluation.

• Argon became callable on 13 October at 16:04:37 UTC: 1,048,576-token context, $2/$10 per million tokens, text, image, audio, video and file input, text out.
• Its million-token output budget is not exposed on the standard route, which reports a 65,536-token completion cap. The batch tier reports 943,718.
• Argon scores 52.56 on Artificial Analysis's Intelligence Index, measured at announcement, which puts it within a point of GPT-6 Astra and Muse Wonder. The cost-per-task figures published on 13 October put it at $1.99 against Claude Opus 5.5's $5.98.
• Datalab's Indra OCR landed at 14:47 UTC on 13 October: a roughly 1M-parameter recognizer, 18 MB of ONNX, 49 pages per second on 32 CPU cores, under an OpenRAIL-M weights licence.
• autotrust/JEV-35B, Apache-2.0 and explicitly unaffiliated with TypeSafe AI, claims 51.3% on ERQA with no robot training. There are still no Jev weights from TypeSafe itself.

The spread has moved to price #

The index scores below were established at announcement and covered here on 1 October; what 13 October adds is the other axis. Artificial Analysis's changelog records its performance results for Argon landing that day, and those are what complete the cost column, because cost per task is a function of price and of how many tokens a model spends getting to an answer.

One caveat belongs on the chart rather than under it: these are each model's charted configuration, not a matched setting. Argon is measured at "high" and most of the others at "max", so the ordering is not like for like; the 1 October edition made the same point.

Read down the left column and the top of the table is a cluster: four models from four companies between 51.83 and 52.67, with Claude Opus 5.5 the only clear outlier five points clear of the rest. Read down the right column and the same set spans $0.72 to $7.63 per task, a factor of eleven. Argon at $1.99 is the cheapest of the four at the top and costs a third of Opus 5.5. The comparison that should worry Google more is the one directly below it: GPT-6.1 Sol scores 1.4% lower and costs 36% as much.

A million-parameter recognizer that reads 49 pages a second #

The quietest release in the window is also the one most likely to change somebody's pipeline this week. Datalab published Indra OCR on 13 October: the weights landed on Hugging Face at 14:47 UTC and the write-up went up the same afternoon. It took 35 points on Hacker News and 45 likes on Hugging Face, which is to say almost nobody noticed.

The headline is that the main recognizer, rec.onnx, carries about 1 million parameters, and all the models together are about 18 MB running on ONNX Runtime with no PyTorch in the dependency tree. It turns PDFs and images into markdown, plain text, or JSON with a bounding box and a label for every line and layout block.

pip install indra-ocr

indra-ocr paper.pdf                       # markdown to stdout
indra-ocr scan.png --format json          # lines, boxes, layout blocks, confidences
indra-ocr paper.pdf --mode ocr --timings  # ignore the PDF text layer, OCR every page

The --mode flag is the one worth knowing. In the default auto, Indra reads a digital PDF's own text layer where it passes its checks and OCRs only what is missing, which is why its median page takes 228 ms on four x86 vCPUs. Forcing ocr drops the headline score from 62.1 to 57.5 and raises the median page to 303 ms: the text-layer shortcut is worth five points and a quarter of the latency.

On OmniParseBench, 16,288 pass/fail tests over 2,937 pages, Indra's headline score of 62.1 beats PaddleOCR's full PP-StructureV3 pipeline at 58.4 while its median page takes 228 ms against PP-StructureV3's 40,958 ms, roughly 180 times faster per page. The honest part is the breakdown: PP-StructureV3 reads plain text better, 55.6 to 41.3. Indra wins the headline on tables (56.4 to 47.8) and layout (88.6 to 71.7). Turning on --math lifts text to 57.5 and the headline to 66.8, at roughly 1.7 times the latency.

Two caveats carry real weight. First, every figure here is self-reported, and OmniParseBench is Datalab's own benchmark, published from its own GitHub organisation; no third party has scored this model. Second, the weights are not open in the way the code is. The code is Apache 2.0, but the weights use a modified OpenRAIL-M licence that is free for research, personal use and companies under $20M in funding or revenue, and explicitly bars use "competitively with our API".

To Datalab's credit the card leads its limitations with the failures rather than burying them: handwriting scores about 8%, degraded old scans about 20%, tables with merged cells come out with cells in the wrong places, and scripts outside Latin, Cyrillic, Greek, Devanagari, Bengali and Chinese work only through a digital text layer. A Hangul line sent to OCR, it says, "comes out as wrong Chinese characters".

The arithmetic is what makes it interesting. At 49.4 pages per second a single 32-core box with no GPU gets through about 4.3 million pages a day; PP-StructureV3 on the same 32 cores, at its reported upper bound of 0.18 pages per second, manages about 15,500. Where 41.3 on text is good enough, that is the difference between a cluster and a laptop.

An open answer to Jev, from nobody in particular #

Jev has been covered here since TypeSafe AI launched it in September as a hosted, closed product with no downloadable artifact of any kind. There is still no Jev release. What exists, uploaded on 11 October, is autotrust/JEV-35B: an Apache-2.0 fine-tune of Qwen/Qwen3.5-35B-A3B from an outfit called AutoTrust AI that reimplements Jev's typed primitives and its endpoint shape. The card is direct about the first half of its provenance, calling itself "an independent open-weights model ... not affiliated with, endorsed by, or a product of TypeSafe AI."

Organisationally that holds. Derivationally it is murkier. AutoTrust's sibling models declare a training corpus, SargeDev/jev-distill-corpus-v3, whose own card states that 498,010 of its 740,957 rows carry labels "distilled from Jev 1.13 (TypeSafe) via OpenRouter". JEV-35B itself declares no training data at all, so for this particular checkpoint the distillation route is a strong inference from its family rather than a documented fact. The Apache-2.0 tag is clean with respect to the Qwen base; whether anyone held the right to relicense labels harvested from a closed commercial API is addressed nowhere.

The claim that is new in this window was added to the card at 00:24 UTC on 14 October: 51.3% on ERQA, DeepMind's 400-question embodied-reasoning set, with no robot or embodied data in training. Baselines re-measured under one protocol in the PhysBrain 1.5 report put PhysBrain 1.5, an 8B model trained on robot trajectories, at 52.8, the rest of the open embodied field between 42.3 and 47.5, and GPT-6 Astra at 75.8. Chance is about 26% and the 95% interval on the overall figure is roughly ±5 points, so "level with PhysBrain" is defensible and "beats it" is not. The card states plainly that the 51.3% is its own run, which also means it was not produced under the protocol that generated the figures it is tabled against.

The mechanism is the part worth taking away, and it explains some of the gap with no appeal to better reasoning:

# Conventional VLM: generate a letter, then hope it parses.
out = model.generate(prompt + "\nA) ... B) ... C) ... D) ...")   # "The answer is B." -> parse

# Typed decision: score the options, read the argmax. No formatting to lose.
r = client.choice(images=imgs, question=q, options=["A", "B", "C", "D"])
r.probabilities    # one score per option, summing to 1

A model that reads option probabilities directly cannot fail a multiple-choice question by formatting its answer wrongly, which over 400 questions is worth real points. That is a measurement artefact as much as a capability, and it cuts both ways: more comparable on this task, less transferable to open-ended ones.

The NVFP4 build is the practical news. It brings the download from 72 GB to 25.6 GB and weights in GPU memory from 66.5 to 23.3 GiB, for 59.18 against 59.73 on the card's decision index, with 96.6% of answers identical to the bf16 model at the same latency. That puts a 35B mixture-of-experts decision model on a single 32 GB card.

Separate the traction from the validation, though. In three days the repo has taken 587 likes and 17,426 downloads, and alongside that: no Hacker News submission, no press, no discussions on the repository itself, and one upvote on the announcement post. The model is also absent from the community Decision Index board whose numbers it quotes as its comparison. Every figure above is the uploader's own, on its own hardware, and nobody has checked any of them.

Also notable #

• Character.AI filed for Chapter 11 bankruptcy protection on 13 October, which the company presents in its own post as a restructuring rather than a shutdown, at 313 points on Hacker News. The Verge connects it to a dispute involving Disney, which is its reporting rather than the company's.
• An open letter signed by more than 400 researchers, Geoffrey Hinton and Yoshua Bengio among them, called on 13 October for an international body to set AI containment standards and incident reporting, with signatories pressing for frontier models to be cut off from open internet access. Reuters carried it first and The Decoder summarised it; the signature count reported is 406.
• Somebody turned EmbeddingGemma 2 into a speech-to-text model. The write-up took 408 points on Hacker News on 13 October, the loudest community result of the day. Routing a text embedding model to transcription is an odd enough path to be worth reading; its performance claims are the author's own and unaudited.
• Apple Machine Learning Research published a design for a virtual GPU cloud built from idle Apple Silicon devices, posted on 13 October, at 240 points on Hacker News.
• The capability-hiding paper covered here on 12 October reached a much wider audience on the 13th, at 233 points on Hacker News. The watch item attached to it then was whether its promised released models would become locatable. They have not: the paper still prints no address for them.

What to watch #

• Whether Argon's million-token output budget exists anywhere a developer can reach it, and what effort tiers sit above "high". Asked here on 1 October and now half-answered: the model is callable and effort is a parameter, but the first surface to carry it exposes a 65,536-token completion cap on the standard route. Google's own API documentation would settle it.
• Whether the $2 and $10 introductory price holds. The announcement named $4 and $20 as the real price, and two weeks of availability have not changed the listed rate yet.
• Whether anyone independently scores JEV-35B, and whether TypeSafe ships Jev weights at all. Both were flagged on 12 October. The community Decision Index board still does not list JEV-35B, so every number attached to it remains the uploader's own.
• Whether any decision-model vendor publishes a fail-open rate. Asked here on 10, 11, 12 and 13 October. JEV-35B's card, the most detailed in the class, carries no such figure either.