What shipped

Twelve days after the full Hy4, Tencent has added the cheaper tier under it. Hy4 Flash is built for routine office work: documents, slide decks, spreadsheets, deep search and clicking through apps. It keeps the flagship's 1M-token window and multimodal input, activates 12 billion parameters per token instead of the flagship's 49 billion, and is priced accordingly.

On Tencent Cloud's Singapore card, as transcribed by the9.pro, Hy4 Flash costs $0.070 per million input tokens and $0.281 per million output tokens, with cached input at $0.003. Against the full Hy4 on the same card ($0.834 / $2.501 / $0.042), that is about one-twelfth the input price, about one-ninth the output price and about one-fourteenth the cache price (ALFW arithmetic). Until 12:00 Beijing time on November 30, 2026 the API is half price again, and the model is free inside WorkBuddy through November 30. Those are published prices with dates attached; the listed rates provide a basis for a trial, subject to confirming the applicable billing terms.

Where 18x comes from

The figure traveling with this launch is not a price ratio. It is a per-task ratio from one test: a reported ¥0.26 per task for Hy4 Flash against ¥4.73 for Claude Haiku 5.5, in a blind test Tencent ran with 85 users of its own WorkBuddy app across 259 everyday tasks.

Three things about that test deserve to stay attached to the number. First, the comparison models were run at their high tier. Second, per-task cost depends on settings, token consumption and prices together, and Tencent reports that its post-training halved Hy4 Flash's average output length; the published results do not isolate how much of the cost gap comes from each factor. Third, the same test did not put Hy4 Flash first on satisfaction. Its point estimate was 2.79 out of 4, close to Haiku 5.5's 2.81 and behind DeepSeek-V4.1-Flash-1017's 3.09, with no uncertainty information supplied. The accurate sentence is narrower than the headline: in Tencent's own test, Hy4 Flash's satisfaction estimate sat next to Haiku 5.5's and its reported cost per task was far lower, while a DeepSeek flash model was rated higher at about four and a half times Hy4 Flash's per-task cost (ALFW arithmetic).

None of this makes the result useless. A vendor-run blind test with named task counts and published scores for all four models, including the one that scored highest, tells a reader more than a chart with one bar. It is still a home fixture: Tencent's app, Tencent's users, Tencent's task set, and an internal WorkIndex suite.

The trade Tencent states itself

Flash tiers are a trade, and Tencent's own numbers show its size. On Terminal-Bench 4.0, Hy4 Flash scores 19.2 against the full Hy4's 44.4. On ProgramBench it scores 16.0 against 30.0. Tencent recommends the flagship for complex, long-horizon software engineering and positions Flash for everyday productivity. A buyer who routes coding-agent work to Flash because the launch headline said "matches Claude" would be reading the wrong test: the Haiku comparison was on office tasks rated by office-app users, not on engineering benchmarks.

The other open item is openness. Tencent says it plans to release the weights. At launch there are no weights and no license, and the model is not yet on OpenRouter. API access is through TokenHub; Tencent also lists access in its own applications. Until the weights land, Hy4 Flash is a hosted price offer, and its quality claims stay exactly as checkable as Tencent chooses to make them.

Sources

  1. [1] the9.pro, "Hy4 flash: 224B/12B MoE, 1M context, $0.07/1M input," November 16, 2026 — a research page compiling Tencent's official launch page, TokenHub pricing and the blind-test chart, with self-reported flagsRead source
  2. [2] TechNode, "Tencent launches Hy4 flash, cutting AI task costs by 95% in internal tests," November 16, 2026Read source
  3. [3] Tencent Hy (Hunyuan), "Hy4 flash: Ultimate cost-efficiency for everyday productivity," official launch page, November 16, 2026 (read via the the9.pro compilation above)
  4. [4] Tencent Cloud, TokenHub model pricing and Hy4 Flash model detail pages (prices, discount terms, context and limits; read via the the9.pro compilation above)