
Tencent Hy4 vs Claude Opus 5.5: A Tenth of the Cost per Task
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 182 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 749 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 1902 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 149 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 109 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Tencent Hy4 stopped being a rented model on 20 November 2026. That is the date Tencent published the weights — tencent/Hy4, Hy4-FP8 and the pre-trained Hy4-base — on Hugging Face, ModelScope and GitCode, sixteen days after the model itself went generally available on 4 November. Claude Opus 5.5, by contrast, has been downloadable never: Anthropic released it on 22 September 2026 as an API-only flagship and has published nothing about opening it. So the interesting question is no longer which model is better. It is what you are buying when you pay roughly ten times more per completed task for a model that scores 24 points higher on the independent index — and whether that trade still holds once the cheap side can be run on your own hardware.
The numbers that make the comparison concrete both come from Artificial Analysis, on the v5.0.1 revision of its Intelligence Index. Claude Opus 5.5 sits at 59 at its maximum effort setting, 57 at xhigh, 52 at high, 48 at medium and 35 at low. Tencent Hy4 sits at 35, with no effort split published. On cost per completed Intelligence Index task the ordering inverts hard: $10.04 for Claude Opus 5.5, $1.03 for Tencent Hy4.
What actually changed on 20 November
Tencent's own framing of the release is a single line in the model card's news section: the company open-sourced Hy4, Hy4-FP8 and Hy4-base. What that means in practice is more specific, and worth separating from the marketing.
• Weights — full BF16 checkpoints. The Hugging Face repository reports 774.4 billion parameters in the safetensors index, which is the 770B backbone plus the vision encoder and the draft model. A 4-bit GGUF build is already up in a separate repository for people who want to test the model without a multi-node cluster.
• Licence — the Tencent Hy License, version 1.0. It is derived from Apache 2.0 but is explicitly not Apache 2.0, and Section 10 adds commercial terms that matter. More on that below.
• Serving — Tencent ships prebuilt images rather than leaving you to fight the architecture: vllm/vllm-openai:hy4 and lmsysorg/sglang:hy4, both with the DFly2 speculative-decoding drafter wired in. The vLLM recipe wants eight GPUs at tensor-parallel size 8 for the FP8 checkpoint.
• What did not change — the price. Tencent Hy4 has been $0.834 per million input tokens and $2.501 per million output since it went GA, with cache reads at $0.042. Opening the weights did not come with a rate cut, and there is no indication one is coming.
The GA build and the open build are the same model. That was not true a month earlier: the August release, Tencent Hy4 Preview, shipped under Apache 2.0 with weights on day one, and the November GA build arrived as a closed API with a promise. The 20 November upload is that promise being kept, with a stricter licence attached.
The gap on the independent index
Twenty-four points is a large gap, and it is larger than the headline suggests once you look at what each number measures. Artificial Analysis publishes Claude Opus 5.5 at five effort levels and Tencent Hy4 at one. The 35 for Hy4 is not an effort-matched figure; it is the only figure the lab has published, and it presumably reflects whatever reasoning setting the model defaults to, which on Tencent's own API is high.
Two things follow. First, the honest comparison is 35 against Opus 5.5's 52 at high effort, not against its 59 — a 17-point gap rather than 24. Second, and more useful, the low-effort row for Claude Opus 5.5 is also 35. At its fastest, cheapest setting Anthropic's flagship lands on the same index score as Tencent's open flagship at its default. That is a genuinely useful fact for anyone building a router table, because it means the choice is not "good model versus weak model" but "which end of the effort dial you can afford".

The sub-scores point the same way. On Artificial Analysis's Cyber Index, published in November, Tencent Hy4 scores 58 and ranks third of 26 models — behind Kimi K3.1 Pro at 64 and a trusted-access GPT-6 Sol variant at 69, and just ahead of DeepSeek V4.1 Flash 1017 at 57. It is first outright on CWE-Bench-AA at 76%. Those are security-specific evaluations, not general reasoning, but they are the only independent scores Tencent's model has that are not a single aggregate, and they are respectable.
What Tencent Hy4 does not have is a public benchmark table from its own lab. The GA announcement in November replaced the preview's Terminal-Bench, DeepSWE and SWE-bench Pro rows with internal task suites — a 213-expert finance panel, a hallucination-rate comparison, an internal WorkIndex score of 42.1. Those are vendor-reported claims about Tencent's own tasks, and the model card repeats them without the public-benchmark rows that would let you check them against anything. The independent index is the check.
What the gap costs, per task
Here is the part that decides most real deployments, and it does not favour the cheaper token price as cleanly as the rate card implies.
• Token price — $0.834 in and $2.501 out for Tencent Hy4 against $4.00 and $20.00 for Claude Opus 5.5. Just under five times cheaper on input, eight times on output.
• Cache — $0.042 per million cached input tokens for Hy4, a 95% discount, against $0.20 for Opus 5.5. On an agent loop that re-reads the same 200,000-token repository on every turn, that difference is the whole cost model.
• Cost per completed index task — $1.03 for Hy4 against $10.04 for Opus 5.5. Note that this is a 9.7x gap, wider than the token-price gap, because Hy4 is also more verbose in absolute terms: 160 million output tokens across the index against 218 million for Opus 5.5, at a verbosity rank of 87th of 94 models for Hy4 and 90th of 94 for Opus 5.5. Neither model is terse.
• Speed — 93 tokens per second for Hy4 against 112 for Opus 5.5, but the latency story is worse than the throughput story. Tencent Hy4 takes 115.9 seconds to produce its first answer token, where the tier median is 12.7 seconds. For interactive work that is the number that matters, and it is the one place where the cheaper model is clearly the worse product.
• Time per task — 14.0 minutes for Hy4 against 25.7 minutes for Opus 5.5. Hy4 finishes index tasks faster in wall-clock terms despite the first-token delay, because it thinks for less time overall.
So the cost claim needs a qualifier that the rate card hides. Tencent Hy4 is roughly a tenth of the cost per completed task, but it will make a user wait nearly two minutes for the first token, and it will answer a question 17 points worse at matched effort. If your workload is batch — document extraction, classification, overnight agent runs, synthetic data generation — the first-token latency is irrelevant and the economics are overwhelming. If it is interactive, the latency is the product and the price advantage does not rescue it.
Where Hy4 wins outright
Three dimensions, and none of them are about being smarter.
• Context economics. Both models advertise a 1M-token window. Only one of them charges $0.042 per million tokens to re-read the same window. Long-context work that involves repeated passes over a fixed corpus — code review against a whole repository, contract analysis across a document set, a research agent that keeps its notes in context — is dramatically cheaper on Hy4 even before the base rate.
• Modality. Tencent Hy4 accepts text, images and video frames through a single ViT path and returns text. Claude Opus 5.5 accepts text, images and files. Video is the difference: Hy4's model card describes a vision encoder with an adapter that converts video frame features into visual tokens, and no equivalent statement exists for Opus 5.5.
• Output length is the one place Opus 5.5 wins cleanly: 128K maximum output tokens against 64K for Hy4. For a task that needs a long generated artefact in one call, that ceiling decides.
Claude Opus 5.5's advantages are the ones Anthropic has always sold: a 17-point index gap at matched effort, a first token in 24 seconds at high effort against 116, twice the output ceiling, and a much larger body of third-party evaluation behind it. If the task is hard software engineering or multi-step analysis where a wrong answer is expensive, the ten-times price difference is not the relevant number. If the task is volume, it is the only number.
The licence is not Apache

This is where the 20 November release differs from the August one, and it is the detail most comparison pages will skip.
Tencent Hy4 Preview shipped under Apache 2.0 — a genuine permissive licence with no strings. Tencent Hy4 ships under the Tencent Hy License 1.0, which is Apache 2.0 with an added Section 10. That section says that if you operate a model-as-a-service business and either your worldwide gross revenue exceeds US$50 million in any twelve-month period, or the product you are embedding the model in has more than 100 million monthly active users, you need a separate written licence from Tencent before you use it. Internal use is carved out: if neither the model nor its outputs reach third parties, Section 10 does not apply.
For a startup, a research group or an enterprise running the model behind an internal tool, that clause is inert and Hy4 is effectively open. For a platform planning to resell inference on top of it, it is a negotiation. Compare that with Claude Opus 5.5, where the question does not arise because there is nothing to license — you rent the endpoint or you do not use the model.
The practical consequence is a split that did not exist a month ago. Tencent Hy4 is now the only model in this comparison you can fine-tune, distil, quantise or run inside a network with no outbound access. Claude Opus 5.5 is the only one of the two with a 128K output ceiling and a first token inside half a minute at its high effort setting.
Choosing between them
The decision tree is short enough to state without hedging.
• You need the highest score per token and the task is interactive — Claude Opus 5.5. Nothing about the price argument survives a user staring at a 116-second first-token delay.
• You need the lowest cost per completed task and the work is batch — Tencent Hy4, and the margin is close to an order of magnitude.
• You need to self-host, fine-tune, or keep data inside your own network — Tencent Hy4, subject to reading Section 10 of the licence. There is no Claude Opus 5.5 option here at all.
• You need 128K of generated output in one call, or video input — the output ceiling points at Opus 5.5, the video input at Hy4. Few workloads need both.
• You want to hedge between the two — put a weighted route in front of them and let the cheap model take the first attempt. Hy4 at $1.03 per task is cheap enough that a failed attempt costs about a tenth of letting Opus 5.5 try first, and the escalation path only fires on the tasks that need it.
Running both on one key

Both models are callable through OrcaRouter under their own model strings — tencent/hy4 at $0.834 per million input tokens and $2.501 per million output, and anthropic/claude-opus-5.5 at $4.00 and $20.00. That means the comparison above can be run as an experiment rather than an argument: one key, one SDK, two model strings, and a routing DSL that sends cheap-first and escalates on failure.
Because OrcaRouter passes provider list prices through at zero markup, Tencent's rate card is the rate you pay on our side, and so is Anthropic's. If either vendor moves its price, the change is live the same day rather than waiting on a reseller to re-key a catalogue.
For the self-hosted case the calculation changes shape entirely. Once you are running Hy4 on your own GPUs, the marginal cost per token stops being a vendor number and becomes your electricity and utilisation figure, which is why the weights matter more than the licence debate for anyone with spare capacity.
What the next month decides
Two things will settle this comparison, and neither is a benchmark row.
The first is whether Tencent's open release pulls independent evaluation of the weights. The 20 November upload includes a system card and a finetuning guide, but the benchmark material in the model card is still mostly the pre-training table — MMLU at 91.6, MMLU-Pro at 73.9, SuperGPQA at 62.7, MMMU-Pro at 59.6 for the base model — plus the internal task suites. Nobody outside Tencent has published a score for the instruction-tuned build, and the 35 on the independent index is a single number from a single evaluator.
The second is effort. Claude Opus 5.5 already publishes five rungs from 35 to 59; the spread between its cheapest and most expensive setting is wider than the entire gap to Tencent Hy4's single score. If Tencent publishes the same kind of split — and the API already exposes a high/no-think toggle — the comparison becomes a table of matched settings rather than one number against another. Until then, the honest summary is that Tencent Hy4 is a mid-pack model on independent evaluation, a strong one on security-specific evaluation, an order of magnitude cheaper per completed task, and now genuinely open, which it was not on 4 November.
Compared in this article2
Detected from this article · Benchmarks: Artificial Analysis · updated daily
