Hy4 flash: Tencent's $0.07 Model Undercuts Haiku 5.5 and GPT-6 Luna
Tencent released Hy4 flash on November 16, 2026 at $0.07/$0.281 per 1M, under Haiku 5.5 and GPT-6 Luna, with 1M context and 12B active params. Self-reported scores.

Key numbers
Hy4 flash · Nov 16, 2026
list; $0.07 input, $0.003 cached
of 224B total, including 64B Engram
960K input, 64K output
Tencent's runs, self-reported
Tencent's claim
Released Nov 16, 2026 · API on Tencent Cloud TokenHub · weights and OpenRouter planned
Tencent released Hy4 flash on November 16, 2026. It is Tencent's cheapest Hy model: $0.07 input / $0.281 output / $0.003 cached per million tokens on Tencent Cloud TokenHub, under the $0.10 / $0.50 that Claude Haiku 5.5, GPT-6 Luna, and Muse Ember all charge. It has 224B total parameters including 64B Engram, 12B active per token, a 1M context (960K input, 64K output), and image input.
Price is the headline. Capability is not. On Tencent's own self-reported appendix, where Tencent ran every comparator at each model's highest available reasoning setting, Hy4 flash trails Hy4 on all nine public rows and trails DeepSeek-V4.1-Flash-1017 on six, while beating GPT-6 Luna on seven. Use it for high-volume search, document, slide, and computer-use agents. Route long-horizon coding to Hy4 or DeepSeek. See Tencent's research page and TokenHub API guide.
What Hy4 flash is
Tencent calls Hy4 flash its most cost-effective Hy model to date: a leaner everyday workhorse for high-volume office work. The design story is cost engineering. Architecture mixes hybrid attention, LatentMoE, and Engram. That stack cuts the KV cache to 1/24 of Hy3 and 1/6 of Hy4. Post-training is self-distillation from Hy4, then reinforcement learning that cut average end-to-end output tokens by 50%. Inference uses megakernels, 4-bit quantization, and DFly to push decoding past 150 tokens per second, per Tencent.
Parameter split
224B total · 12B active per token
KV cache 1/6 of Hy4
KV cache 1/24 of Hy3
end-to-end output tokens cut 50% by RL
Tencent describes it as "natively multimodal." The TokenHub API takes text and images and returns text. Image input supports PNG, JPEG, WEBP, and static GIF, up to 10 MB per image and up to 600 images per request. Recommended size is 128 to 2048 px; images above 5500 x 5500 are rejected.
On price, Hy4 flash undercuts Hy3 as well ($0.132 / $0.528 list), not only Hy4 ($0.834 / $2.501). Tencent's routing advice is plain: use Hy4 flash for deep search, document processing, slide (PPT) generation, simple front-end display, and computer use; use Hy4, "with a higher ceiling," for long-horizon software engineering and other complex tasks. TokenHub now marks Hy4 preview as "To Be Discontinued."
Benchmarks
Every score below is self-reported by Tencent. Tencent ran Hy4 flash and every comparator in the appendix at each model's highest available reasoning setting. These are not the other vendors' own published numbers. Do not mix the two.
Public appendix · 9 rows
Public appendix table
Compact table of the nine public rows. Same caveat: Tencent-run, self-reported, November 2026.
| Benchmark | Hy4 flash | Hy4 | DeepSeek-V4.1-Flash-1017 | GPT-6 Luna |
|---|---|---|---|---|
| DeepSWE v1.1 | 64.6 | 74.9 | 73.8 | 67.1 |
| ProgramBench (almost solved) | 16.0 | 30.0 | 29.0 | 6.0 |
| Terminal-Bench 4.0 | 19.2 | 44.4 | 40.0 | 13.3 |
| APEX-Agents v1.1 (pass@1) | 63.1 | 75.2 | 72.9 | 44.7 |
| AutomationBench v1.0.6 | 52.2 | 56.2 | 54.3 | 38.1 |
| HLE-Full (with tools) | 55.2 | 59.1 | 57.1 | 49.7 |
| OSWorld 2.1 (full, partial) | 68.1 | 78.9 | 68.4 | 55.7 |
| MMMU-Pro | 75.8 | 78.6 | 76.7 | 77.5 |
| ZeroBench-main (tools, pass@5) | 45.8 | 55.4 | 16.5 | 33.6 |
Honest reading of Tencent's own table: Hy4 flash sits close to the bigger models on office automation and computer use (AutomationBench 52.2 vs Hy4 56.2; OSWorld 68.1 vs DeepSeek 68.4). It is far behind on terminal and long-horizon coding (Terminal-Bench 4.0 19.2 vs Hy4 44.4 and DeepSeek 40.0; ProgramBench 16.0 vs 30.0 and 29.0). It is clearly ahead of GPT-6 Luna on most agent rows. Tallies with a gap under 1.0 counted as a tie: vs Hy4, 0 wins and 9 losses; vs DeepSeek, 1 win, 2 ties, 6 losses; vs Luna, 7 wins and 2 losses; vs GLM-5.3-Flash, 6 wins and 3 losses; vs Qwen3.8-Flash-Next, 5 wins, 1 tie, 3 losses. Average of the seven public agent rows (all except MMMU-Pro and ZeroBench): Hy4 flash 48.3, Hy4 59.8, DeepSeek 56.5, Qwen 48.1, GLM 47.5, Luna 39.2.
- DeepSWE v1.1 and ProgramBench: mini-swe-agent harness.
- Terminal-Bench 4.0: Claude Code harness, up to 500 turns, 13-hour timeout per trial.
- APEX-Agents v1.1: official reference harness, max 250 steps.
- AutomationBench v1.0.6: 600 public tasks released by Zapier.
- ZeroBench-main: network-isolated Jupyter sandbox, max 80 tool calls, pass@5 over 10 runs.
WorkIndex is a separate, Tencent-built suite run in WorkBuddy CLI. Average scores: Hy4 flash 42.1, Hy4 51.4, Hy4 preview 39.7. Tencent's chart puts average WorkIndex cost per task at ¥5.7 for Hy4, ¥3.2 for Hy4 preview, and ¥0.2 for Hy4 flash (about 24.5x cheaper than Hy4 by Tencent's arithmetic from list prices at 7.2 CNY per USD). Treat WorkIndex as an internal product metric, not an independent public benchmark.
Human evaluation
Tencent also reports a WorkBuddy blind test: 85 internal WorkBuddy users, 259 everyday tasks in 27 categories, external models at "high." These are Tencent product-team claims.
| Model | Satisfaction (out of 4) | Median end-to-end time | Cost per task (Tencent estimate) |
|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 | 3.09 | 10.3 min | ¥1.20 |
| Claude Haiku 5.5 | 2.81 | 6.8 min | ¥4.73 |
| Hy4 flash | 2.79 | 6.7 min | ¥0.26 |
| GLM-5.3-Flash | 2.53 | 7.6 min | ¥0.81 |
Honest reading: Hy4 flash ties Haiku 5.5 on satisfaction and time at a fraction of the cost. DeepSeek-V4.1-Flash-1017 scored highest on satisfaction but took longest. Tencent also says that in WorkBuddy, versus Hy4, average task time fell 70% and average cost was about 1/15, with roughly equal results on daily work tasks and a small drop on long coding tasks. WeCom reports output speed 88% faster than Hy4 and about 60% fewer thinking-mode output tokens. Label all of that as Tencent product-team claims.
Pricing
Lead with list price. Hy4 flash on Tencent Cloud TokenHub is $0.07 / $0.281 / $0.003 per million tokens (input / output / cached input). A 50% TokenHub promo brings that to $0.035 / $0.1405 / $0.0015 until November 29, 2026, 11:00 PM ET (12:00 noon Beijing time on November 30). After that, usage bills at list. No batch price is listed. WorkBuddy access is free through November 30 as a consumer product promotion, not an API price.
Versus the shared $0.10 / $0.50 of Claude Haiku 5.5, GPT-6 Luna, and Muse Ember: Hy4 flash input is about 30% lower and output about 44% lower. Cached input at $0.003 is less than a third of Haiku's and Luna's $0.01. Versus DeepSeek-V4.1-Flash-1017: about a quarter of the peak rate and about half off-peak. Versus Hy3: about half. Versus Hy4 list: about 1/12 of input and 1/9 of output.
Versus Xiaomi's MiMo-V2.6-Flash ($0.14 / $0.28 / $0.0028): Hy4 flash has half the input price, essentially the same output price, and a slightly higher cached price. MiMo also offers a Batch tier at half price; Hy4 flash lists none. That is why Hy4 flash is one of the two cheapest in this set, not an unqualified cheapest model.
Output price rails · USD per 1M · log scale
$0.10 to $3.00
Hy4 flash
$0.28
out / 1M
in $0.070 · cache $0.003
MiMo-V2.6-Flash
$0.28
out / 1M
in $0.14 · cache $0.0028
Muse Ember
$0.50
out / 1M
in $0.10 · cache $0.004
Claude Haiku 5.5
$0.50
out / 1M
in $0.10 · cache $0.010
GPT-6 Luna
$0.50
out / 1M
in $0.10 · cache $0.010
Hy3
$0.53
out / 1M
in $0.13 · cache $0.033
DeepSeek-V4.1-Flash-1017
$1.20
out / 1M
in $0.30 · cache $0.006
Gemini 3.1 Flash-Lite
$1.50
out / 1M
in $0.25 · cache $0.025
Hy4
$2.50
out / 1M
in $0.83 · cache $0.042
A low sticker does not by itself make it the cheapest per task. Tencent's own per-task estimates favor Hy4 flash, including the WorkIndex "24.5x" figure vs Hy4, but those are Tencent's numbers computed from list prices. Official peer pages: TokenHub pricing, Anthropic, OpenAI, Meta, Google, DeepSeek, and Xiaomi MiMo.
Specs
| Detail | Value |
|---|---|
| API id | hy4-flash |
| Context | 1M tokens |
| Max input | 960K tokens |
| Max output | 64K tokens |
| Modalities | Text and image in, text out |
| Protocols | OpenAI Chat Completions, OpenAI Responses, Anthropic Messages |
| Default reasoning | Deep thinking; reasoning_effort defaults to high. Adaptive thinking in tool-call scenarios. |
| Image limits | PNG, JPEG, WEBP, static GIF; 10 MB per image; up to 600 images per request; recommended 128 to 2048 px; above 5500 x 5500 rejected |
| Parameters | 224B total, including 64B Engram; 12B active per token |
| Availability | Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, Marvis |
When to use it
- Pick Hy4 flash for high-volume office agents (slides, documents, spreadsheets), deep search and summarization, computer-use flows, image plus text tasks at 1M context, and anywhere output cost dominates.
- Pick Hy4 or DeepSeek-V4.1-Flash-1017for terminal-heavy coding, long-horizon software engineering, and anything where Terminal-Bench or ProgramBench matter. In Tencent's own runs those gaps are large.
- Pick Haiku 5.5, GPT-6 Luna, or Muse Emberif you need those vendors' ecosystems, independent evaluation history, or an OpenRouter route today.
- Wait if you need open weights. Tencent plans to open-source Hy4 flash, with no date and no license stated.
Availability
Hy4 flash is API-only on Tencent Cloud TokenHub today. It is also live in WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. Tencent says it will launch on OpenRouter soon; it is not listed as of November 16, 2026. Weights are planned with no date. No license is stated. No model card is published. For the research write-up, see hy.tencent.ai/research/hy4-flash.
Banner: Tencent Hy. Cover art from the Hy4 flash research article: a serif Hy wordmark under a hatched right triangle on drawn axes, with the word flash, on an off-white ground.
FAQ
- Tencent released Hy4 flash on November 16, 2026. The API id is
hy4-flash. It is live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. - List price on Tencent Cloud TokenHub: $0.07 input, $0.281 output, and $0.003 cached input per million tokens. A 50% TokenHub promo runs at $0.035 / $0.1405 / $0.0015 until November 29, 2026, 11:00 PM ET (12:00 noon Beijing time on November 30). No batch price is listed.
- Among the small 1M-context, image-input reasoning models from major labs we compared, Hy4 flash has the lowest input price ($0.07). Its output price ($0.281) is essentially tied with Xiaomi's MiMo-V2.6-Flash ($0.28), which also offers a Batch tier at half price. So call it one of the two cheapest in that set, not the cheapest without qualification.
- Hy4 flash is cheaper by about 12x on input and 9x on outputat list rates, with a KV cache one sixth the size of Hy4's. On Tencent's self-reported appendix, where Tencent ran every model, Hy4 flash trails Hy4 on all 9 public rows. Tencent says use Hy4 flash for deep search, documents, slides, and computer use, and use Hy4 for long-horizon software engineering.
- Hy4 flash undercuts their shared $0.10 / $0.50list sticker (about 30% lower input, 44% lower output). In Tencent's runs it leads GPT-6 Luna on 7 of 9public rows. In Tencent's WorkBuddy blind test it ties Haiku 5.5 on satisfaction (2.79 vs 2.81 out of 4) and median time at a fraction of Haiku's estimated cost. Those are Tencent's numbers, not Anthropic's or OpenAI's.
- Hy4 flash list price is about a quarter of DeepSeek's peak rateand about half off-peak. On Tencent's self-reported appendix it trails DeepSeek on 6 of 9public rows (1 win on ZeroBench, 2 ties). DeepSeek also scored highest on Tencent's blind-test satisfaction (3.09 vs 2.79), while taking longer per task.
- Context is 1M tokens: max input 960K, max output 64K, per the Tencent Cloud TokenHub API guide.
- Not yet. Tencent says it plans to open-source Hy4 flash and iterate frequently. No date and no license are stated. Weights are not public as of November 16, 2026.
- Not yet. Tencent says it will launch on OpenRouter soon. As of November 16, 2026 it is not listed.
- No.It is Tencent's cheapest Hy model, built for high-volume office, search, and computer-use work. Tencent's own score table shows it behind Hy4 on every public row and behind DeepSeek-V4.1-Flash-1017 on most.
Continue Reading
- Qwen4-Max: Near-Frontier Agent Work at $1.50 and $4.50
- Gemini 4 Hydrogen: Argon-Class Agent Work at $1.50 and $7.50, Flat to 1M
- Seedream 5.5: ByteDance's Editing Model Beats 5.0 Pro at Less Than Half the Price per Image
- Muse Ember: Meta's Volume Model at the Same $0.10 and $0.50 as Haiku 5.5
- Claude Mythos 5.2: Same Model as Fable, Different Guest List
