Tencent prices its new AI model below Anthropic's cheapest and promises a free download. It hasn't said when
Tencent has launched a cheaper AI model that lists below the base price of Anthropic's cheapest Claude model. It also says it will let companies download the model and run it themselves. It has not said when, or under what license.
The model, Hy4 flash, went on sale on November 16. Tencent charges about $0.07 for every million tokens, the small chunks of text AI models read, and about $0.281 per million tokens it writes. Claude Haiku 5.5 lists at $0.10 and $0.50 for prompts under 100,000 tokens. That makes Tencent's list price about 30% lower for reading and 44% lower for writing.
For buyers, the price is only half the pitch. The other half is the promise of a free, downloadable version. A company that runs a model on its own servers pays for hardware, not per use, and keeps its data in-house. That promise remains unfulfilled for now.
What Tencent claims, and who has checked
Tencent pitches Hy4 flash for everyday office work: research, document handling, slide decks and operating software on a user's behalf. It steers harder jobs, such as long software projects, to its larger flagship, Hy4.
The cost claims are steep. On WorkIndex, a set of office tasks Tencent built itself, the company says Hy4 flash costs one-twentieth as much per task as Hy4. It puts the cost at one-eighth of a comparable DeepSeek model at that model's peak-hour pricing. Tencent also says the new model slightly outscores an earlier, much larger Hy4 test version on the same tasks.
Those per-task figures go beyond list prices. They also count shorter answers and fewer retries, so they depend on how Tencent built its test. Tencent says its training cut the average length of task output by half.
On quality, Tencent ran a blind test with 85 internal users of WorkBuddy, its own office assistant. They rated answers on 259 tasks across 27 categories. Tencent says satisfaction came close to Claude Haiku 5.5. Dataconomy reports the scores as 2.79 for Tencent against 2.81 for Anthropic's model.
None of this has been checked by outsiders. When the model launched, Artificial Analysis, an independent testing firm, had not scored it, and no third party had reproduced any figure, according to model marketplace OrcaRouter. OrcaRouter also found no model card, technical paper or safety report.
The giveaway has no date
Tencent's announcement says only that it "plans to open-source Hy4 flash with frequent iterations." It names no license and no date. As of November 17, Tencent's page on Hugging Face, the main site for downloading free models, had no copy of the model, according to ModelFit, a site that tracks what hardware models need.
The pattern matters for anyone planning around it. Tencent made the same promise for its flagship Hy4 on November 4. Outside trackers say the final Hy4 is still offered only as a paid service, with downloadable files to follow at an unset date alongside a safety report. Only an August test version can be downloaded. Tencent's own launch post nonetheless describes Hy4 as a top open-source model.
The license, when it comes, will decide who can actually use it. Tencent's August test version came under Apache 2.0, a standard license that permits commercial use. Tencent has not said whether Hy4 flash will get the same terms.
What buyers should weigh
The near-term discounts are generous but short. Hy4 flash is free inside WorkBuddy until November 30, and Tencent's cloud service charges half price until then. OrcaRouter advises against budgeting on those rates. Tencent says listing on OpenRouter, a popular marketplace that resells many companies' AI models, is coming soon.
If the files do ship, running the model in-house would not need a large server cluster. Hy4 flash has 224 billion parameters, the internal settings that make up a model, but uses only 12 billion at a time. ModelFit estimates a compressed copy needs about 112 to 121 GB of memory, enough to fit on two of Nvidia's desktop AI computers or one high-end Mac Studio. That is an estimate, not a test.
The broader signal is price pressure. Cheap workhorse models from Anthropic and OpenAI already list at the same $0.10 entry rate. Tencent is now pricing below both and promising a free version on top. It is the same sequence it used with Hy4: sell access first, promise the files later.
For executives, the sensible stance is to treat Hy4 flash as two separate offers. The paid service is real, cheap and unverified by outsiders. The free version is a promise. Until Tencent posts the files and the license, it cannot be part of a cost plan.