Сгенерированная титульная карточка для «Tencent Hy4 vs GPT-6.1 Sol — what the token price hides». Левая карточка: Tencent Hy4, «выпущена 4 Nov 2026, веса открыты 20 Nov», «Intelligence Index 35», «$0.834 / $2.501 за 1M», «$1.03 за задачу индекса». Правая карточка: GPT-6.1 Sol, «выпущена 29 Sep 2026», «Intelligence Index 55 при максимальном усилии / 35 при низком», «$2.00 / $10.00 за 1M», «$1.27 за задачу индекса, устарела в пользу GPT-6.2 Sol». Нижний колонтитул: «Показатели индекса и стоимости за задачу по данным Artificial Analysis v5.0.1, проверено 20 November 2026».
Guides & Insights

Tencent Hy4 против GPT-6.1 Sol: что скрывает цена токена

Автор

jinhao.song

Дата публикации

Новые модели · 20Все модели →
Бенчмарки: Artificial Analysis · обновляется ежедневно
Назад ко всем статьям

O​penAI has already retired the model this comparison is about. GPT-6.1 Sol, released at DevDay on 29 September 2026 at $2.00 per million input tokens and $10.00 per million output, is marked deprecated on Artificial Analysis, which recommends GPT-6.2 Sol instead — a model released on 17 November 2026 at exactly the same $2.00 and $10.00 with an Intelligence Index of 57 rather than 55. Tencent Hy4, generally available since 4 November 2026 and open-weighted since 20 November, is not deprecated by anything and does not have a successor waiting. So the honest version of this matchup has two halves: Hy4 against the Sol tier as it stands today, and a note that the Sol tier's own vendor replaced it nine days ago at no change in price.

The arithmetic here has a trap in it, and it is the reason this comparison is worth writing. On the rate card GPT-6.1 Sol charges $2.00 and $10.00 per million tokens against Tencent Hy4's $0.834 and $2.501 — a blended $1.47 against $0.45 at a 7:2:1 cache ratio, so Hy4 looks three times cheaper. On the completed task it is not. Artificial Analysis measures GPT-6.1 Sol at $1.27 per Intelligence Index task against $1.03 for Hy4, because Sol finishes each task in 63.7 million output tokens where Hy4 burns 160 million. Twenty index points cost 23% more per answer, not three times more, and the model that looks expensive per token is the one that uses fewer of them.

The Sol tier is a price point, not a ceiling

O​penAI's GPT-6 family has three tiers and they are not distinguished by architecture so much as by where they sit on a rate card. GPT-6 Astra is the flagship at $10 and $50 per million. GPT-6 Sol — both 6.1 and now 6.2 — sits at $2 and $10. GPT-6 Luna is at $0.10 and $0.50. Sol is the middle: five times cheaper than Astra, twenty times more expensive than Luna.

That matters because the Sol tier's value proposition has never been "the best model". It has been "the model that is nearly as good as the flagship at a fifth of the price", which is the framing O​penAI used when it introduced GPT-6.1 Sol a week after GPT-6 Sol shipped. The 6.1 release changed almost nothing on the card — same $2.00 and $10.00, same 1M-token context, same 128K output — except cached input, halved from $0.20 to $0.10, and a modest bump from 50 to 55 on the index at maximum effort.

Tencent Hy4 is in a different structural position. It is not a tier of anything. It is a single open-weight flagship, 770 billion total parameters with 49 billion active, sold at one price with one reasoning default. There is a cheaper sibling, Hy4 Flash, at $0.0175 and $0.070 per million — roughly 48 times cheaper than Hy4 itself — but Hy4 Flash is a 220-billion-parameter model with two-thirds the index-relevant capability, not a tuned-down copy of the flagship.

So the rate card is comparing a vendor's deliberate mid-market product against another vendor's deliberate low-margin open flagship — and the open flagship's low margin buys it nothing on the metric that matters, because its verbosity spends the saving. Both facts are worth holding at once, and the second one only becomes visible when you price the answer rather than the token.

What $2 buys on each side

Convert to outputs rather than tokens, because that is where the money actually goes.

• Per million output tokens — $2.501 for Tencent Hy4 against $10.00 for GPT-6.1 Sol. Four times. Blend input, output and cache at a 7:2:1 ratio and the multiple compresses to 3.3x, $0.45 against $1.47, because Hy4's cached rate of $0.042 is doing heavy lifting on the input side.

• Per million input tokens — $0.834 against $2.00. Two point four times.

• Per million cached input tokens — $0.042 against $0.10. Hy4 is 2.4 times cheaper on cache, and both vendors treat caching as a first-class discount rather than an afterthought.

• Per completed Intelligence Index task — $1.03 for Hy4 against $1.27 for GPT-6.1 Sol. This is the number to quote, and it is the one that runs against the headline: Hy4 spends 160 million output tokens across the index where Sol spends 63.7 million, so Sol's four-times output rate buys a quarter of the tokens and lands at 1.23 times the cost per answer. Artificial Analysis rates Hy4 at the top of its verbosity scale, four of four units.

• Long-context surcharge — GPT-6.1 Sol has a tier that doubles input to $4.00 and lifts output to $15.00 above 272,000 prompt tokens. Tencent Hy4 has no such tier; a 1M-token request is billed at the flat rate. For a workload that routinely exceeds 272K, the real multiple on the O​penAI side is not four but closer to five or six.

That last row is the one most comparisons miss, and it is the one that matters for the long-context work both models advertise. A 500,000-token request costs $2.00 per million input on GPT-6.1 Sol by the headline rate and $4.00 per million in practice. T​encent charges $0.834 either way.

Effort levels are the real comparison

GPT-6.1 Sol has five reasoning settings and they span twenty index points. At maximum effort it scores 55; at xhigh, 52; at high, 50; at medium, 46; at low, 35. Tencent Hy4 publishes one number, 35, and its API exposes a binary choice — high reasoning on by default, or a no-think mode for direct answers.

Put those two facts together and something useful falls out. GPT-6.1 Sol at its lowest effort setting scores exactly what Tencent Hy4 scores at its only setting. The cheaper model is not competing with the expensive model's best configuration; it is competing with the expensive model's cheapest one, and on the index it draws.

That reframes the decision. If your workload would run GPT-6.1 Sol at low or medium effort — classification, extraction, short agent steps, anything where latency and token spend dominate — then the 20-point gap to its maximum setting is not a gap you were ever going to pay for, and Hy4 at a quarter of the output rate is the obvious choice. If your workload needs maximum effort, the gap is real and the price is what it costs to close it.

The absence of an effort ladder on Hy4 is itself a limitation. There is no way to ask T​encent's model for a cheaper, faster answer short of switching to Hy4 Flash, which is a different model at a different size. O​penAI's five-rung dial lets a single deployment trade quality for money per request; T​encent's does not.

First token, and why it decides interactive work

Throughput numbers flatter Tencent Hy4 and latency numbers bury it, and both are true.

• Output speed — 93 tokens per second for Hy4 against 45 for GPT-6.1 Sol. Hy4 decodes twice as fast, which is the opposite of what the first-token number suggests and is why the two metrics have to be read together.

• Time to first answer token — 115.9 seconds for Hy4 where the tier median is 12.7. That is not a small deviation from the norm; it is nine times the median, and it reflects a model that reasons for a long time before emitting anything. GPT-6.1 Sol at maximum effort waits 147.6 seconds for its first token, so on this comparison Hy4 is the faster of the two — the tier's interactive settings are further down the dial, at 28.3 seconds on high, 5.4 on medium and 3.2 on low.

• Time per completed task — 14.0 minutes for Hy4 against 24.9 minutes for GPT-6.1 Sol at maximum effort. The cheaper model is also the faster one end to end, by a factor of nearly two, because it thinks for less wall-clock time per answer even though it writes more.

For batch work none of this matters — a job that runs overnight does not care whether a response took two minutes or twelve seconds to start. For anything a person is waiting on, both maximum-effort settings are unusable: 116 seconds for Hy4, 148 for Sol. The difference is that Sol has rungs below maximum where the wait drops to seconds and Hy4 does not, so the argument for paying O​penAI's rate is not its peak speed — it is the existence of a cheaper, faster setting on the same model string.

Where the open model is simply better

Три вещи, и они не маленькие.

• You can run it. The 20 November upload includes the BF16 checkpoint, an FP8 build and a pre-trained base model, with prebuilt vllm/vllm-openai:hy4 and lmsysorg/sglang:hy4 images. GPT-6.1 Sol has no weights and never will. For a regulated buyer, an air-gapped deployment or a fine-tuning plan, that is the whole decision and nothing else on this page matters.

• It reads video. Hy4's vision path adapts video frame features into visual tokens alongside image tokens. GPT-6.1 Sol accepts text, images and files. For a pipeline that ingests screen recordings or camera footage, one of these models can do the job and the other needs a preprocessing stage.

• It has no long-context price cliff. The 272,000-token threshold on GPT-6.1 Sol doubles input pricing exactly where long-context workloads start to be interesting. Hy4's flat rate is a genuine structural advantage for a 400K-token agent loop, above and beyond the headline multiple.

Against that, GPT-6.1 Sol wins on the index at every effort level above low, has a 128K output ceiling against Hy4's 64K, has a twenty-point effort ladder that Hy4 lacks, and has the first-token latency that makes interactive work possible. It also has a successor nine days old at the same price, which is either reassuring — the tier is actively maintained — or a warning, depending on whether you have already integrated 6.1.

The deprecation is the story

Screenshot of the Artificial Analysis model page for GPT-6.1 Sol, showing an Intelligence Index of 55 at max effort with a five-level effort curve, $2.00 per million input and $10.00 output, a 1M-token context window, and a deprecation note pointing to GPT-6.2 Sol.

GPT-6.1 Sol lasted seven weeks. It shipped on 29 September 2026 and was superseded on 17 November 2026 by GPT-6.2 Sol at an identical $2.00 and $10.00, with a two-point index gain at maximum effort — 57 against 55 — and the same five-rung effort ladder. Nothing in the pricing changed. The model simply moved on.

That cadence has a cost that never appears on a rate card: integration churn. A team that tuned prompts against GPT-6.1 Sol at high effort now has a model with a different effort-to-score curve, and the only way to know whether their prompts still land is to re-run the evaluation. Twice a year is manageable. Every seven weeks is a maintenance burden.

Tencent Hy4's cadence looks different and should be read with the same suspicion. The Hy4 line shipped a preview in August, a GA build in November and an open-weight release sixteen days later, plus a Flash sibling in between. That is four events in four months. The difference is that three of them were the same underlying model in different states of release, and the fourth was a smaller sibling — so an integration against tencent/hy4 in November is still an integration against tencent/hy4 today. The model string has not moved.

For a buyer, the practical rule is to price the churn. A model that changes every seven weeks at a fixed price is cheap per token and expensive per migration; a model that changes its weights once and then opens them is the reverse.

A routing plan that uses both

The two models are complementary in a way that makes a weighted route more than a hedge.

• First attempt — Tencent Hy4. At $1.03 per completed index task against $1.27, a failed attempt costs 19% less than a GPT-6.1 Sol attempt, and it returns in 14 minutes rather than 25. On workloads where 70% of requests are easy, the saving is real but modest; the latency is the larger gain.

• Escalation — GPT-6.1 Sol, or its 6.2 successor, for anything that fails a validation check or exceeds a token budget. The escalation only fires on the hard tail, where the 20-point index gap actually earns its keep.

• Interactive surfaces — Sol at medium or low effort, always. Hy4's 116-second first-token latency disqualifies it from anything a person is watching, and Sol's own maximum setting is worse; it is the lower rungs, at five and three seconds, that make the tier usable.

• Regulated or offline — Hy4 only, weighted to 100% by policy rather than by cost.

That structure turns the argument into a measurement: log which requests escalate, and the escalation rate tells you whether the expensive model is earning its premium on your traffic. It is a more useful number than any benchmark row, because it is measured on your distribution.

One key, two tiers, no second contract

A generated two-column scoreboard titled 'Tencent Hy4 vs GPT-6.1 Sol'. Tencent Hy4 column: Intelligence Index 35, one reasoning setting, $0.834 / $2.501 per 1M tokens, $1.03 per index task, 160M output tokens on the index, 64K max output, weights open 20 Nov 2026. GPT-6.1 Sol column: Intelligence Index 55 at max effort and 35 at low, five effort levels, $2.00 / $10.00 per 1M tokens, $1.27 per index task, 63.7M output tokens on the index, 128K max output, weights closed and the model deprecated. Footer: 'Index, cost-per-task and token figures per Artificial Analysis v5.0.1; deprecation note per OpenAI's successor release on 17 November 2026.'

Both models are callable through OrcaRouter under their own strings — tencent/hy4 at $0.834 and $2.501, openai/gpt-6.1-sol at $2.00 and $10.00 with the long-context tier applied above 272,000 tokens. The routing DSL composes them into a single call path, which is the only practical way to run the escalation pattern above without writing two client integrations and reconciling two billing streams.

Because OrcaRouter passes provider list prices through with zero markup, both rate cards are the ones their vendors publish. T​encent's November weight release did not change its price, and O​penAI's 17 November successor did not change Sol's — so if either moves, it moves on our side the same day rather than waiting for a reseller to catch up.

Automatic failover is worth naming separately here. A deprecated model is a model whose endpoint can change behaviour without warning, and a route that fails over to the successor keeps a production path alive through that transition rather than discovering it in an incident review.

Короткая версия

Screenshot of the Hugging Face model card for tencent/Hy4, showing the repository title, open-weight tags, the Tencent Hy License badge and links to ModelScope, GitCode and GitHub.

Twenty points of independent capability separate these models at maximum effort, and four times the output price separates them on the invoice. Both numbers are true and neither is the decision: per completed answer the gap is 23%, and at low effort Sol's score falls to exactly Hy4's 35, which is why the effort dial rather than either headline figure should drive the choice.

Tencent Hy4 is the right model for batch work at volume, for anything with video, for any request that runs past 272,000 tokens, and for any deployment that has to be self-hosted. GPT-6.1 Sol is the right model for interactive work and for the hard tail of a task distribution, with the caveat that its own vendor has already replaced it at the same price and the successor is the one to integrate against.

The claim worth discarding is that a cheap token price makes a cheap answer. It does not here: Hy4's 160 million tokens to Sol's 63.7 million eat most of a four-times rate advantage, and the open model's headline saving shrinks to a fifth on the metric a buyer actually pays. The claim worth keeping is that most production traffic never needs maximum effort, and at the effort levels production actually runs, GPT-6.1 Sol's five-rung dial is the thing you are buying — not its peak score.

Сравнение в этой статье3

Определено по этой статье · Бенчмарки: Artificial Analysis · обновляется ежедневно