GPT-6.1 Sol vs Tencent Hy4 Preview: Paying $10 an Output Token Against an Open Model at $2.50

Put the two rate cards side by side and the gap is not subtle. Tencent Hy4 Preview lists at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input on its international endpoint. GPT-6.1 Sol lists at $2.00, $10.00 and $0.10 respectively. That is a 4× spread on output tokens and a 2.4× spread on input, and because both models are priced per token, the question the spread raises is not which is cheaper — Hy4 Preview is cheaper on every line — but whether the expensive one finishes enough more work per call to justify it. Neither vendor has published a figure that answers that directly, and this article is largely about why the obvious numbers on both sides are not the answer.

Tencent released and open-sourced Hy4 Preview on 2026-08-28; Open​AI released GPT-6.1 Sol on the API on 2026-09-29. Both are text-capable agentic models with million-token context windows, both are reachable today, and both are in our catalogue — which makes this the rare comparison in this series where the per-token arithmetic can be run against two live rows rather than against a press release. Figures below come from Tencent's release material and model documentation, Open​AI's model and pricing pages, our own catalogue records, and published evaluations as noted, read on 2026-10-12. Every number that is a vendor's own run is labeled as one.

What each model actually is

Start with the architecture, because the two models are solving the same serving problem with the same technique.

• Parameters — Hy4 Preview is a 770B mixture-of-experts with 49B active per token across 78 layers, using 256 routed experts plus 1 shared expert. GPT-6.1 Sol is a proprietary Open​AI model positioned one tier below GPT-6 Astra in the same generation; Open​AI does not publish parameter counts.

• Context and output — both document roughly a million tokens of input context. Our catalogue record for Hy4 Preview carries 1,048,576 tokens of context with a 64,000-token maximum completion; Tencent's own material describes a context window exceeding 1M tokens with a 960,000-token maximum input. GPT-6.1 Sol documents 1,050,000 tokens of context with a 128,000-token maximum output — double Hy4's completion ceiling, which is the spec that bites when you want a long structured artifact out of a single call.

• Modality — both are text-only on the input side, with the tool use and structured output you would expect of a current agentic model. Neither is a vision model.

• Reasoning controls — Hy4 Preview exposes high, low and none, with high as the default; a temperature of 0.9 and top_p of 1.0 are the recommended sampling settings. GPT-6.1 Sol exposes low, medium (default), high, xhigh and max, with no off position at all — reasoning is always on, and sampling parameters such as temperature and top_p are rejected outright.

• Weights — Hy4 Preview is open-sourced under Apache 2.0 and downloadable today. GPT-6.1 Sol is closed, API-only, and will never be self-hosted.

That last line is the one that separates these two models more than any benchmark. A 770B/49B Apache 2.0 model is a possession; a proprietary API model is a subscription. Everything else on this page is about which of those two things your workload needs.

The price lines, with the arithmetic done

Both rate cards are published and both are passed through by us without markup, so this section is verifiable rather than reported.

• Output — GPT-6.1 Sol $10.00 per million against Hy4 Preview $2.501 per million. For every million output tokens a workload emits, the difference is $7.50, and output is the meter a reasoning model spends most heavily.

• Input — $2.00 per million against $0.834. A 2.4× gap on the line that dominates any long-context pipeline, because a repository-scale or document-set prompt re-sends a large prefix on every turn.

• Cached input — $0.10 per million against $0.042. Both vendors discount re-read prefixes steeply, and Hy4's cache rate is less than half of GPT-6.1 Sol's.

• Cache writes — GPT-6.1 Sol charges $2.50 per million to write a cached prefix. Hy4 Preview's published international rates do not carry a separate cache-write line, which means the first write is priced as ordinary input rather than as a premium.

Run a shape that is ordinary for either model: 500,000 input tokens of which 460,000 are a cached prefix, and 20,000 output tokens, in a single call.

• On GPT-6.1 Sol — 460,000 cached tokens at $0.10/M is $0.046, 40,000 fresh input tokens at $2.00/M is $0.080, and 20,000 output tokens at $10.00/M is $0.200. Total $0.326.

• On Hy4 Preview — 460,000 cached tokens at $0.042/M is $0.019, 40,000 fresh input tokens at $0.834/M is $0.033, and 20,000 output tokens at $2.501/M is $0.050. Total $0.102.

The Hy4 call is a little under a third of the GPT-6.1 Sol call on this shape. Now the caveat that makes the comparison honest: this is a comparison of two rate cards applied to an identical token count, and token counts are not identical between models. A model that reasons longer before answering will emit more output tokens for the same task, and GPT-6.1 Sol is documented as producing 10–30% more output tokens than its predecessor GPT-6 Sol on comparable work. If Hy4 Preview's default of reasoning "high" makes it equally verbose, the real-world gap narrows; if Open​AI's model is the more verbose of the two, the gap widens past the 3.2× the arithmetic shows. Rate cards tell you the price of a token. They do not tell you how many tokens a task costs, and only running your own workload settles that.

Where the benchmarks do and do not help

This is the section where the two models' evidence quality diverges, and the divergence is the point.

Tencent's headline evaluation for Hy4 Preview is a blind side-by-side preference test, and the result is a 2.99 for Hy4 Preview against 2.92 for GLM 5.3 and 2.94 for Kimi K3. Read that carefully: it is Tencent's own evaluation, it is a preference score rather than an accuracy score, the winning margin is seven hundredths of a point on a roughly three-point scale, and the comparison set is two other open models. It is evidence that Hy4 Preview is competitive with the current open-weight tier. It is not evidence that it is competitive with a proprietary frontier model, and no suite in the release material makes that claim.

Tencent also reports a 31.8% throughput improvement from the model's own self-optimisation work, and the architecture includes a native multi-token-prediction layer for speculative decoding. Both are serving-efficiency claims from the vendor, unreproduced by anyone else, and both are plausible for a 770B/49B MoE built with that layer — but they describe how fast the model answers, not how well.

GPT-6.1 Sol has the stronger evidence position and it is worth being precise about why. Its independent Intelligence Index row at maximum reasoning effort is 52, with a measured cost to run of $0.72 per task and output speed of roughly 65 to 67 tokens per second — a third party ran it, scored it, and reported what the run cost. Its Artificial Analysis Terminal-Bench 4.0 figure is 56.06, against 43.94 for its predecessor GPT-6 Sol, which puts the generational move on shell-based agentic tasks at about twelve points. Against that, Open​AI's own headline evaluations for the model — including a Terminal-Bench Science result quoting $5.47 per task at maximum effort against $23.21 for Claude Opus 5.5 — are the vendor's runs with the vendor's baselines, and nobody has reproduced them.

So the honest summary of the evidence is lopsided and it runs one way: GPT-6.1 Sol has a third-party measurement of both quality and per-task cost, and Hy4 Preview has a vendor preference score against other open models. That does not mean Hy4 Preview is worse than the preference score suggests. It means nobody outside Tencent has measured it, which is a different statement, and a reason to test it on your own tasks rather than trusting either side's numbers.

The self-hosting question, with real numbers

Hy4 Preview's open weights are the reason to take it seriously beyond the API price, and the reason it is not a straightforward win either.

The practical figures, as published: a Q4_K_M quantisation of the model runs to roughly 435 GiB, a 1-bit quantisation to roughly 213 GiB, and the full unquantised model needs on the order of 1.5 TB of VRAM. The 213 GiB 1-bit figure is the only one of those three that a single high-memory node can hold, and it is also the one that will cost you the most quality. The 435 GiB Q4_K_M figure needs multi-GPU infrastructure that most teams do not have sitting idle.

What that means in practice: self-hosting Hy4 Preview is a real option for organisations that already run that class of hardware, and a capital project for everyone else. If you are in the first group, the API price above is irrelevant to you — you are trading a metered bill for a fixed one, and Apache 2.0 imposes no licence cost on top. If you are in the second group, the API rate is the only number that matters and the $0.834 / $2.501 / $0.042 rates are the reason to look at the model at all.

One caveat from users, and it is a cost caveat rather than a capability one: Hy4 Preview's documented rough edges are over-reasoning and over-verifying — spending more tokens than a task needs, checking work that is already correct. On a metered API that behaviour shows up on the output line at $2.501 per million; on self-hosted hardware it shows up as latency. Either way, the recommended temperature of 0.9 and the default reasoning setting of "high" are configuration choices worth tuning before you judge the model on cost.

Which card you should be holding

The two models are not substitutes for the same buyer, and the split is clean.

Take GPT-6.1 Sol if the workload is expensive to get wrong and cheap relative to the cost of failure — long agentic coding loops, computer use, document-heavy professional work where a task that fails costs more than the tokens that produced it. It has the independent measurements, the wider context, double the output ceiling and a knowledge cutoff of 2026-04-30. Budget for the fact that its output meter is where the money goes, and design around the 272,000-input-token threshold above which Open​AI reprices the entire request. It is in our catalogue at Open​AI's list rates with 0% markup: https://www.orcarouter.ai/models/openai/gpt-6.1-sol.

Take Hy4 Preview if the workload is high-volume, cost-sensitive and tolerant of a model nobody outside its maker has benchmarked — bulk extraction, classification, summarisation, tool-calling agents where the task is well-specified and the failure is cheap to retry. At $0.834 / $2.501 / $0.042 it is a third of the cost on the same token counts, it is open-sourced under Apache 2.0 if you ever want to stop renting it, and it is in our catalogue at Tencent's list rates with no markup added: https://www.orcarouter.ai/models/tencent/hy4-preview. The same 0% markup rule applies to both rows, which means a vendor price cut lands on the catalogue the same day it lands on the vendor's page rather than after a reseller notices.

The configuration most teams will actually want is both, with the cheap model carrying volume and the expensive one carrying the hard tail — and that is exactly the setup a single endpoint with automatic failover across 200-plus models exists to make cheap, since an unproven model like Hy4 Preview is far easier to try when it sits behind the same key as a model you already trust, with a route that falls back automatically rather than a second integration to maintain.