Hy3 API Pricing: ¥1/¥4 Rates, 4 Providers, and GPT-5.6 Cost Comparison

Decision first: Hy3 is worth testing when a team processes large volumes of Chinese text, builds workplace agents, or needs a low-cost domestic API route. Tencent's current published rate is ¥1 per million input tokens and ¥4 per million output tokens. ComputeUnion currently tracks four Hy3 providers; the official Tencent Hunyuan route is also the lowest recorded input price at about $0.1379 per million tokens. For 10 million input tokens and 2 million output tokens, the official bill is about ¥18 ($2.48).

That price does not make Hy3 the automatic answer for every application. GPT-5.6 Sol has a higher dated external Intelligence Index, while GLM-5.2 has broader tracked provider coverage. The useful question is narrower: can Hy3 complete your actual Chinese and agent tasks reliably enough that its lower hosted cost survives retries and human review?

What the current data actually says

Signal	Hy3 snapshot	Why it matters
Official hosted API price	¥1 input / ¥4 output per 1M tokens	Baseline for a real API bill
ComputeUnion coverage	4 tracked providers	Current buying routes, not theoretical availability
External Intelligence Index	41, #7 of 97 in its comparison class	Dated independent evidence, not a universal verdict
Measured output speed	45.5 tokens/second	Latency may matter for interactive workloads
Official context	Up to 256K tokens	Long documents still require cost and latency testing
Model architecture	295B total / 21B active parameters	Official architecture fact, not a quality score

Snapshot date: 2026-07-17. Provider count and lowest price can change; the linked Hy3 price page is the current source of record.

Open weights do not mean a hosted API costs zero

Artificial Analysis displays a $0 token price for the open-weights Hy3 entry. That is useful when comparing openness and self-hosting, but it is not the bill a developer receives from a managed API. ComputeUnion records the current Tencent hosted API at ¥1/¥4 per million input/output tokens. These figures describe different products: downloadable weights versus a managed endpoint with infrastructure, availability, and service terms.

Practical reading: Hy3 can be free to download under its license while still costing money to call through Tencent Cloud or another hosted provider.

Monthly Hy3 cost at three usage levels

The formula is input tokens divided by one million, multiplied by ¥1, plus output tokens divided by one million, multiplied by ¥4.

Workload	Monthly input	Monthly output	Official Hy3 cost	What to check
Small production test	10M	2M	¥18 / about $2.48	Integration quality matters more than price
Medium agent workload	1B	200M	¥1,800 / about $248	Retries, concurrency and output length
Large-scale processing	10B	2B	¥18,000 / about $2,483	Enterprise limits, invoicing and SLA

Output is four times the input rate. Hy3 is therefore most attractive when prompts are large but answers remain short and structured. Repeated long explanations can erase much of the apparent price advantage.

Who should test Hy3 now?

More likely to benefit	Should compare carefully first
High-volume Chinese classification, extraction and summarization	Applications tied to OpenAI-specific tools or response behavior
Workplace agents, document workflows and knowledge-base Q&A	Global multi-region services with strict international SLA requirements
Domestic teams that need RMB billing, local access or invoices	Small workloads where migration saves only a few dollars
Teams building a second model route for resilience	Core systems that have not run task-level regression tests

Tencent documents progress in software development, office productivity, financial modeling, front-end design, game production, knowledge-base Q&A, task planning and tool coordination. Those claims justify what to test; they do not prove that every workload will outperform GPT or GLM.

Hy3 vs GPT-5.6 Sol vs GLM-5.2

Model	External Intelligence Index	Current tracked input price	Tracked providers	Decision role
Hy3	41	$0.1379 official hosted API	4	High-volume Chinese and agent candidate
GPT-5.6 Sol	59	$5 official; tracked channels from $0.0966	19	Frontier capability candidate
GLM-5.2	51	Tracked channels from $0.0579	18	Coding-focused Chinese competitor

The price scopes are deliberately labelled. Hy3 has a verified official hosted rate. GPT-5.6 shows official and channel prices. GLM-5.2 currently has a tracked channel floor on ComputeUnion, but this article does not invent a monthly bill where a complete comparable output-rate record is unavailable.

How to use Hy3 without wasting the price advantage

1. Route by task difficulty. Start common Chinese extraction and summarization on Hy3; escalate failed or high-value tasks to a more expensive model.
2. Constrain output. Request schemas, bullet limits or maximum sections instead of open-ended explanations.
3. Measure successful-task cost. Include retries, tool failures, latency and manual review—not just token price.
4. Test the exact provider route. A low quote is not proof of the same model ID, quota, privacy policy or support responsibility.
5. Use a 20-task regression set. Ten common tasks, five difficult tasks and five known failures expose many migration problems before production.

What gives Hy3 future potential—and what could limit it

The strongest positive signal is usage, not marketing language. Tencent reports that average daily Hy3 token consumption increased twenty-fold after the preview launch, while active WorkBuddy users selecting Hy3 grew sixfold. Tencent has integrated it into multiple products and made the weights available under Apache 2.0. Those facts support a reasonable case for continued ecosystem growth.

The limits are equally concrete. Artificial Analysis reports about 45.5 output tokens per second and describes the model as verbose in its benchmark workload. ComputeUnion currently tracks only four providers, far fewer than GPT-5.6 Sol or GLM-5.2. Hy3 may become a strong large-scale Chinese workhorse, but it still needs wider hosted availability and task-level evidence before being treated as a universal replacement.

Clear conclusion

Hy3 is currently a low-cost, high-volume candidate—not an automatic frontier-model replacement. It is most compelling for Chinese workloads, workplace agents and domestic enterprise procurement. Begin with the official ¥1/¥4 rate, run the same task set against GPT-5.6 Sol and GLM-5.2, and keep Hy3 only when its successful-task cost remains lower after retries and review.

Sources and verification method

Official positioning, architecture, context and adoption figures come from Tencent's Hy3 release announcement. Independent Intelligence Index and speed figures come from Artificial Analysis. Hosted prices and provider coverage come from ComputeUnion records, with official, channel, live_api and unverified states kept separate. Prices do not create capability scores, and open-weight cost is not presented as hosted API cost.