A generated hero card comparing Hy4 flash and Gemini 3.8 Flash, with a left tile reading 'Hy4 flash — $0.070 / $0.281, open weights, no independent score' and a right tile reading 'Gemini 3.8 Flash — $0.75 / $3.75, Intelligence Index 31, deprecated', plus a footer reading 'Gemini figures per Artificial Analysis on index v5.0.1; Hy4 flash prices are Tencent list rates.'
Guides & Insights

Hy4 flash vs Gemini 3.8 Flash: A Deprecated Model Is Still the One With a Price List

Author

jinhao.song

Date Published

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

Artificial Analysis labels Gemini 3.8 Flash "deprecated" and points readers toward Gemini 4 Hydrogen. Google still sells it at $0.75 per million input tokens and $3.75 per million output. Hy4 flash is two days old, costs a tenth of that, and has no independent score at all. So the honest question in this matchup is not which model is better — nobody can answer that yet — but which of two uncomfortable positions you would rather occupy: paying flagship-adjacent prices for a model its own maker has superseded, or paying almost nothing for a model nobody outside Tencent has measured.

Hy4 flash went onto Tencent's API on 16 November 2026 and shipped as open weights on 17 November: 220B total parameters, 12B active, a 1,048,576-token context, image input with text output, under the Tencent Hy License. Gemini 3.8 Flash was released on 2 September 2026 with a 1M-token context, text, image, video, audio and file input, and an Artificial Analysis Intelligence Index score of 31 at its high reasoning setting. It is the only model in this comparison with an independent measurement, and it is also the only one marked deprecated.

What the deprecation label actually means here

Artificial Analysis's "deprecated" flag on Gemini 3.8 Flash means the board considers it superseded, and the page points at Gemini 4 Hydrogen as the replacement. It does not mean Google has switched it off. The model resolves, it is priced, and it is on the OrcaRouter catalogue at $0.75 and $3.75 with a 90% cache discount, which means it is still being served to paying customers by at least one platform.

What it does mean is that the roadmap has moved past it. A model released in early September and flagged as superseded by mid-November had a shelf life of roughly two and a half months as the current generation. Any team standardising on it now is standardising on something the vendor has already replaced in its own recommendations, and the practical consequence is straightforward: the next capability improvement will not arrive as an update to this model, it will arrive as a different model you have to migrate to.

Hy4 flash has the opposite problem. It is not deprecated; it is unmeasured. Tencent's own evaluation is a WorkIndex internal suite and an 85-user blind test, both vendor-run, and Artificial Analysis's Hy4 flash URL returns 404 as of 18 November. The model card's public benchmark tables are base-versus-base and describe the pre-trained checkpoint, not the instruction model you would actually call.

Neither position is comfortable. One is a known quantity on a closing curve; the other is an unknown quantity on an opening one.

The price gap is real, and it is not the interesting part

Tencent's mainland card for Hy4 flash reads ¥0.5 per million input, ¥2 per million output, ¥0.02 per cached input — about $0.070, $0.281 and $0.003. Google's list price for Gemini 3.8 Flash is $0.75 and $3.75, with cache reads at a 90% discount.

• Input — $0.070 against $0.75. Roughly 10.7x cheaper for Hy4 flash.

• Output — $0.281 against $3.75. Roughly 13.3x cheaper.

• Cached input — $0.003 against about $0.075. Roughly 25x cheaper.

• Blended at a 7:2:1 input/output/cache mix — roughly $0.10 per million for Hy4 flash; Artificial Analysis puts Gemini 3.8 Flash at $0.58.

That is not a discount, it is a different category of product, and it explains the second number that matters: cost per completed Intelligence Index task. Artificial Analysis measures Gemini 3.8 Flash at $7.77 per task, over 14.2 minutes. There is no Hy4 flash figure because there is no Hy4 flash score — but Tencent's WorkIndex arithmetic puts its average cost per task at one-thirteenth of Hy4 preview, one-twentieth of Hy4 and one-eighth of DeepSeek-V4.1-Flash-1017 at peak-hour pricing.

Those are different tasks and different harnesses and must not be divided into each other. What they establish is the shape of the trade: Gemini 3.8 Flash is expensive per task because it is expensive per token and spends a lot of tokens getting there. Hy4 flash is engineered — hybrid sparse attention, a 1/6-of-Hy4 KV cache, 50% shorter outputs from RL — specifically to make the per-task number small.

A generated reference card headed 'Price favours one, modality favours the other', with a Hy4 flash panel of four rows and a Gemini 3.8 Flash panel of four rows, and a split band beneath reading '10.7x cheaper on input, 13.3x on output' beside 'four more input modalities'.

Capability, on the one axis we can measure

Gemini 3.8 Flash's independent scores, all Artificial Analysis on v5.0.1, are effort-dependent and the effort setting matters more than the headline:

• Intelligence Index at high — 31.

• Intelligence Index at medium — 30.

• Intelligence Index at low — 23.

• Class rank — 66th of 92 on the Intelligence Index.

• Output speed — 116 tokens per second.

• Time to first answer token — 22.4 seconds.

• Context — 1M tokens.

An index of 31 at high effort is mid-pack, and the deprecated flag exists because Google has shipped better since. Earlier coverage of this model quoted 59 at launch; that figure predates the 27 October 2026 index rebase and is not comparable to today's 31. Anyone citing 59 and 31 in the same breath is comparing two different instruments.

For Hy4 flash there is no row to put underneath. The nearest thing to an external anchor is Tencent's own blind test, where Hy4 flash scored 2.79 out of 4 for user satisfaction against 2.81 for Claude Haiku 5.5 — and Claude Haiku 5.5 scores 36 on the same index where Gemini 3.8 Flash scores 31. If you chain those two facts together you get an implied position for Hy4 flash somewhere near the low thirties, which is exactly the kind of inference this blog exists to flag as an inference rather than a measurement. Two vendor numbers and one third-party number do not make a benchmark.

Screenshot of the Artificial Analysis model page for Gemini 3.8 Flash, showing its Intelligence Index scores at high, medium and low effort, its class rank, output speed, time to first answer token, context window and cost per task on index v5.0.1.

Modality is where Gemini 3.8 Flash still wins outright

This is the row that survives the price gap, and it is worth being concrete about it.

• Hy4 flash input — text and images, through a 27-layer SigLIP2-derived encoder with 2x2 feature merging.

• Gemini 3.8 Flash input — text, image, video, audio and file, per the OrcaRouter catalogue entry and Google's model documentation.

• Gemini output — text.

If your pipeline ingests a screen recording, a call recording or a PDF with an audio track, Hy4 flash cannot accept it and Gemini 3.8 Flash can. The openness of the weights does not help; the encoder that is not there is not there. Tencent's own recommended use cases for Hy4 flash are text-and-screenshot shaped — deep search, document processing, slide creation, simple frontend demos, computer use — which is a coherent product decision and also a boundary.

On documents specifically the two are closer. Both take images, both hold a million tokens of context, and Hy4 flash's hallucination profile on document work is what its WorkIndex suite is actually measuring: spreadsheet repair, document editing, slide co-writing. That is a more honest claim than a generic benchmark score, provided you remember it is Tencent's rubric.

Two very different ways to be a "flash" model

Google's flash tier means fast and cheap relative to the pro tier, on a closed API, with the vendor absorbing the serving complexity. Tencent's flash means something more specific: an architecture whose entire purpose is to make a 1M-token context cheap to serve.

• Attention — Hy4 flash interleaves sparse IndexPool attention with 512-token sliding windows at a 1:3 ratio across 42 layers; Gemini's architecture is undisclosed.

• KV cache — Hy4 flash is 1/24 of Hy3's and 1/6 of Hy4's, vendor-reported; no comparable public figure for Gemini 3.8 Flash.

• Speculative decoding — Hy4 flash ships DFly2, a 0.85B drafter that reuses the target model's sparse KV and indexer selection so the draft cost stays constant with context; Google's equivalent is internal.

• Output length — Tencent says RL post-training cut average output tokens by 50%, and the blind test reports a 6.7-minute median task completion; Gemini 3.8 Flash's index run averages 14.2 minutes per task.

The last row is the one to sit with. A large fraction of what makes Hy4 flash cheap per task is that it writes less, and a model that writes less is not automatically a model that is worse — but it is a model whose outputs you should read before you route a customer-facing workflow through it. Verbosity and quality are different axes and only one of them shows up in a cost column.

Deployment is the axis where they are not comparable at all

Gemini 3.8 Flash is an API. There is no checkpoint, no FP8 build, no GGUF, no fine-tuning path. The OrcaRouter catalogue entry shows the model itself and nothing else.

Hy4 flash is an API, three Hugging Face repositories — the instruct model, an FP8 build and a base checkpoint — a GGUF quantisation from AngelSlim, a fine-tuning guide, pinned vLLM and SGLang images, and a Transformers integration currently living on pull request 50903. That list is not uniformly good news: pinned images and a PR-ref install mean you are running builds Tencent tested rather than builds that have been released.

But it changes what "cheap" means. On Gemini 3.8 Flash, cheap means Google's price, and Google can change it. On Hy4 flash, cheap means Tencent's price or your own GPUs, and the second option does not have a price at all — it has a capacity plan, and the architecture is an argument that the plan is small. The Tencent Hy License permits that, with one caveat: Section 10.2 requires a separate written licence above $50 million in worldwide gross revenue as a model-as-a-service business, or above 100 million monthly active users. Internal use is exempt. A European or Korean team has no territorial restriction, which is more than can be said for some Chinese open-weight licences. The licence is Apache 2.0 with a revenue trigger bolted on, which means the two deployment paths are priced under the same document rather than two — a detail that matters when your finance team asks why the cheap option is cheap.

Where to actually run either one

Gemini 3.8 Flash is on the OrcaRouter catalogue at google/gemini-3.8-flash, $0.75 per million input tokens and $3.75 per million output with a $0.075 cache read, routed at Google's list price with 0% markup on our side — so if Google cuts the price of a model it has already deprecated, you see it the day it happens rather than at the next catalogue sync. Hy4 flash is not on the catalogue; what is there is its teacher, tencent/hy4-preview, at $0.834 and $2.501 per million with a $0.042 cache read.

The useful pattern for this specific pair is a routing rule rather than a choice. Put the audio and video ingestion tier on the Gemini route, because it is the only one of the two that accepts those inputs. Put the high-volume text-and-document tier on the cheap route once you have measured it yourself. The routing DSL exists to express exactly that split in one endpoint, and automatic failover means the cheap route degrading does not take the pipeline down with it — which matters more than usual when the cheap route is two days old and running on a pinned container image.

Screenshot of the google/gemini-3.8-flash model page, showing the per-million-token input, output and cache-read prices the model is listed at and the modalities it accepts.

The decision, stated as three cases

If you need audio or video input, the comparison ends: Gemini 3.8 Flash is the only option here, and the fact that it is deprecated is a migration cost you should price in rather than a reason to rule it out.

If you need a model with an independent score before you put it in front of customers, the same answer applies for a different reason. Thirty-one on the Intelligence Index is a modest number, but it is a measured one, and Hy4 flash does not have a measured anything.

If your workload is high-volume text and documents, latency-tolerant, and you can run your own evaluation, Hy4 flash is worth the afternoon it takes to test. A 13x output-price gap is not a rounding error, and the fact that Tencent published a blind test it lost by 0.02 points is a better signal about the model's honesty than any number on Google's card. Just do not skip the evaluation: the model has no third-party score, its public benchmark tables describe a different checkpoint than the one you will call, and the only outside measurement of it is a satisfaction survey run by its maker.

And watch the deprecation. If Google retires Gemini 3.8 Flash the way it retired the models before it, this page's real answer changes from "the expensive one is measured" to "the measured one is gone," and a two-day-old open-weight model with no score starts looking considerably more attractive than it does today.

Compared in this article3

Detected from this article · Benchmarks: Artificial Analysis · updated daily