A generated title card for 'Tencent Hy4 vs Gemini 4 Argon'. Left card for Tencent Hy4: 'Released 4 Nov 2026, weights open 20 Nov', '770B total / 49B active', '1M context, text image video in', '$0.834 / $2.501'. Right card for Gemini 4 Argon: 'Released 30 Sep 2026', 'Intelligence Index 50 at high effort', '1M context, text image audio video PDF in', '$2.00 / $10.00'. Footer: 'Index figures per Artificial Analysis v5.0.1.'
Guides & Insights

Tencent Hy4 vs Gemini 4 Argon: Open Weights vs a Wider Modality Set

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Take a thousand pages of scanned invoices and ask two questions of them: what does it cost to read, and what does it cost to reason over. The answers point in opposite directions, and that is the whole of the Tencent Hy4 versus Gemini 4 Argon decision. Tencent Hy4, generally available since 4 November 2026 and open-weighted since 20 November, is a text-and-image model at $0.834 per million input tokens. Gemini 4 Argon, which Google released on 30 September 2026, accepts text, images, audio, video and PDFs, and charges $2.00 per million input tokens. One is 2.4 times the price of the other on input and four times on output. The cheaper one can be downloaded and run on your own hardware. The more expensive one reads formats the cheaper one cannot see at all.

On the independent index the two are close enough that the price gap is hard to justify on capability grounds alone: Artificial Analysis puts Gemini 4 Argon at 50 on its Intelligence Index at high effort and 48 at medium, against 35 for Tencent Hy4. A 15-point gap at matched-ish effort is real, but it is far smaller than the 2.4x input and 4x output multiples suggest it should be. The cost-per-completed-task figure is where the price gap actually lands: $4.97 for Argon against $1.03 for Hy4 on the same index.

The invoice, worked through

Concrete arithmetic beats a rate-card screenshot, so here it is for one plausible job: 1,000 scanned invoice pages, roughly 800 tokens of image tokens each, a 600-token prompt, a 300-token structured output per page.

• Input volume — 1.4 million image tokens plus 0.6 million prompt tokens, so 2.0 million input tokens in total.

• Output volume — 300,000 tokens of extracted fields.

• Tencent Hy4 — 2.0 million at $0.834 is $1.67, plus 0.3 million at $2.501 is $0.75. Total $2.42.

• Gemini 4 Argon — 2.0 million at $2.00 is $4.00, plus 0.3 million at $10.00 is $3.00. Total $7.00.

• With cache — if the prompt prefix is stable across pages, which in an extraction pipeline it usually is, Hy4's cached input at $0.042 per million cuts the prompt share to almost nothing; Argon's cached rate of $0.10 per million does the same in proportion. The ratio does not move much, because the image tokens dominate and neither vendor discounts those heavily.

A single run is $4.58 apart. At a million pages a month it is $4,580 against $7,000, and the interesting question becomes whether the 15-point index gap buys enough accuracy to justify $2,420 a month — which depends entirely on whether Hy4's extra points of error are in a field you can post-validate.

That last clause is the useful part. Invoice extraction with a schema and a validation pass is a workload where a weaker model plus deterministic checks beats a stronger model alone, because a failed parse is cheap to retry and a wrong number that passes validation is not. Tencent Hy4 at 35 is not too weak for that job. It is probably too weak for the job two paragraphs down.

Where the modality gap actually bites

Hy4 reads text, images and video frames through one vision encoder. Gemini 4 Argon reads text, images, audio, video and PDF documents natively, and Google's model documentation lists all five on the model's capability line. The difference is not a marketing line; it changes what pipelines are possible without a preprocessing stage.

• Audio. A call recording, a meeting, a voicemail — Argon takes the file. Hy4 needs a transcription step first, which means a second model, a second contract and a second failure mode.

• PDF as a document rather than as pages. If a PDF has selectable text plus embedded images plus a scanned signature block, Argon handles the mixture in one pass. Hy4 processes rendered pages as images, so the text layer is lost and the token cost rises.

• Long video. Tencent's card describes frame-feature adaptation for video, which is a real capability and better than most open models have. But it is frame sampling, not native temporal modelling, and for anything where the question is about motion over a long clip, the difference shows.

• The reverse gap. Hy4 takes 64K output tokens against Argon's 64K default with a 1M ceiling available on the output side per Google's documentation. Neither is the binding constraint for extraction work.

So the honest framing is that Gemini 4 Argon is the broader instrument and Tencent Hy4 is the cheaper one. On a workload that stays inside text and images — which is most code, most documents, most agent work — the breadth is unused and the price is what you pay for.

Capability, measured and mis-measured

Artificial Analysis puts Gemini 4 Argon at 50 (high effort) and 48 (medium) on the v5.0.1 Intelligence Index, with a blended price of $1.47 per million tokens, 68 tokens per second, a 47-second wait to first token at high effort, and $4.97 per completed index task. Tencent Hy4 sits at 35, with 93 tokens per second, a blended $0.45 and $1.03 per completed task. Neither figure carries an effort split on the Hy4 side, which is the same limitation that dogs every comparison involving it: Tencent does not publish its own model across effort levels and the evaluator has only one number.

Google's own announcement carried a nineteen-row benchmark table comparing Argon against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. It leads on Vals Index, AutomationBench, Harvey's Legal Agent Benchmark and several vision suites; it loses on FrontierSWE v2 and Terminal-bench 4.0. Those are Google's numbers on Google's methodology, with the non-Gemini figures taken from the other providers' own scoring — a caveat the methodology PDF states explicitly. They are useful for knowing what Google optimised for, not for ranking.

Tencent's evidence base is thinner and different in kind. The Hy4 announcement dropped the preview's Terminal-Bench and SWE-bench rows entirely and substituted internal task suites: a 213-expert finance panel where Hy4 averaged 3.04 of 4, a hallucination rate cut from 15.76% to 5.45%, a +19.1% gain on an internal game-development suite. The model card repeats those. None of them are reproducible outside Tencent. The one independent security-specific result, a Cyber Index score of 58 and a first-place 76% on CWE-Bench-AA, does not have a Gemini 4 Argon counterpart because Google's model is not on that board.

What that means for a buyer: Argon has a measured aggregate and a vendor table; Hy4 has a measured aggregate, a security-specific result, and a vendor table about internal work. Both aggregates come from the same evaluator, which is what makes the 15-point comparison worth something.

Why this matchup is only weeks old

Worth stating plainly, because the record matters. When Google announced Gemini 4 Argon on 30 September, it described a three-stage rollout — cyber defenders through the Fairwind Program first, then paid API customers and AI Ultra subscribers, then everyone else — with no dates past the first stage. At that point the name had a model-family page, a benchmark table and no model identifier you could put in a request.

That has changed. Google's model documentation now lists gemini-4-argon as a current model with no preview label, a 1M-token input window, and the five-input modality set described above. Artificial Analysis lists it as released and available via API through Google. So the comparison in this article is between two models you can call, which it would not have been six weeks ago. The correction is worth making because a lot of pages still describe Argon as a gated teaser.

Tencent Hy4's own status changed in the other direction: it went from closed API on 4 November to open weights on 20 November, under the Tencent Hy License 1.0 rather than the Apache 2.0 that governed the August preview. Both models moved, in opposite directions on openness, inside the same six weeks.

The openness question, priced honestly

Hy4's 770-billion-parameter checkpoint is downloadable. Running it takes eight H100-class GPUs for the FP8 build at tensor-parallel size 8, per Tencent's own vLLM recipe. That is a real machine, and for most companies the honest answer is that open weights buy optionality rather than savings: the model is self-hostable in principle, cheap to rent in practice, and worth downloading only if data residency, fine-tuning or an air-gapped deployment is a hard requirement.

The licence adds a term worth reading. Section 10 of the Tencent Hy License requires a separate written licence if you operate a model-as-a-service business with worldwide gross revenue above US$50 million in any twelve-month period, or more than 100 million monthly active users. Internal use is carved out. So an enterprise running Hy4 behind an internal tool is unaffected; a platform reselling inference on it is negotiating.

Gemini 4 Argon has no equivalent question because there is nothing to license — you call Google's endpoint or you do not use the model. That is a simpler arrangement and a strictly less flexible one. For a regulated buyer, "we can run it in our own rack" is often worth more than fifteen index points; for everyone else it is a footnote.

Which one to put in the request path

• Volume extraction inside text and images — Tencent Hy4. The arithmetic above is not close, and a schema plus validation pass absorbs most of the accuracy gap.

• Anything with audio, native PDFs, or long video — Gemini 4 Argon. Hy4 cannot see the input, and a preprocessing model costs more than the token difference.

• Interactive surfaces — Argon at medium effort, if either. Hy4's 115.9-second wait for a first token makes it unsuitable, and Argon's high-effort path is 47 seconds, which is better but still not a chat interface.

• A regulated network or a fine-tuning plan — Tencent Hy4, subject to Section 10.

• The hardest reasoning you have — Gemini 4 Argon, on the strength of a 15-point independent gap that both models' vendors agree exists in roughly that direction.

Both on one key, and what that buys

A generated two-column scoreboard titled 'Tencent Hy4 vs Gemini 4 Argon'. Tencent Hy4 column: Intelligence Index 35, $0.834 / $2.501 per 1M tokens, input modalities text image video, weights open 20 Nov 2026, 64K max output, licence Tencent Hy 1.0. Gemini 4 Argon column: Intelligence Index 50 at high effort, $2.00 / $10.00 per 1M tokens, input modalities text image audio video PDF, weights closed, 64K default output, rollout phased at announcement. Footer: 'Index per Artificial Analysis v5.0.1; modality and licence details per each vendor's documentation.'

Both models are callable through OrcaRouter under their own strings — tencent/hy4 at $0.834 and $2.501, google/gemini-4-argon at $2.00 and $10.00 — which makes the modality split a routing decision rather than an architecture decision. Send image and text work to the cheap model, and route on input type so that anything carrying audio or a native PDF goes to the model that can read it. That is the routing DSL doing the job it exists for, and it costs one key rather than two integrations.

The pass-through pricing matters here more than usual, because the two vendors move their rates on different schedules. OrcaRouter charges provider list price with zero markup, so a Google price change and a Tencent price change both land on our side the same day rather than after a reseller re-keys a catalogue.

Automatic failover is the other half. A batch job that fails a schema validation on Hy4 can retry on Argon without the client knowing, which is a cheaper way to buy accuracy than paying four times the output rate on every call.

What would change the answer

Screenshot of the Artificial Analysis model page for Gemini 4 Argon, showing the model name, Google as vendor, an Intelligence Index of 50 at high effort and 48 at medium, a price of $2.00 per million input and $10.00 output tokens, and a 1M-token context window.Screenshot of the Hugging Face model card for tencent/Hy4 showing the repository title, the Tencent Hy License badge, the open-weight tags and the download and like counts.

Two developments would move this comparison, and both are plausible on a twelve-month horizon.

The first is a Tencent model with audio. Hy4's vision path is one encoder over images and sampled video frames; adding an audio encoder is the obvious next step, and it would collapse most of Argon's breadth advantage. Tencent has already shipped Hy4 Flash, a cheaper 220-billion-parameter sibling with the same modality set, so the family is expanding rather than consolidating.

The second is independent evaluation of the open Hy4 weights. The 35 on the index is one evaluator's number for one build. With the weights public, a second lab can run the instruction-tuned checkpoint and either confirm the gap to Argon or narrow it. That has not happened yet — the model card's benchmark material is still largely the pre-training table, MMLU at 91.6 and MMMU-Pro at 59.6 for the base model, plus Tencent's internal suites.

Until then the summary is narrow and defensible. Gemini 4 Argon is the broader model and the higher-scoring one, at 2.4 times the input price, four times the output price and 4.8 times the cost per completed task. Tencent Hy4 is the cheaper one by an order of magnitude per completed task, open since 20 November, and blind to audio. Choose on the input type first, because that decides the question before price gets a vote.

Compared in this article1

Detected from this article