What happened

Tencent has released Hy4 Flash, a cut-price AI model built for coding, office chores and AI agents, and the pitch is blunt: in the company's own blind tests it matches Claude Haiku 5.5 (high) on everyday work, while individual office tasks cost more than 16 times less than Tencent's flagship Hy4. Set against Claude Haiku, Tencent puts the gap at roughly 18 times cheaper.

A quick jargon decoder before the numbers, because most coverage assumes you already speak model-nerd:

• Parameters are the dials inside a model. Hy4 Flash has 224 billion of them. More dials usually means more capability, but also a bigger bill to run.

• Mixture of experts (MoE) is the trick that keeps it affordable. Picture a huge firm where only the relevant specialists show up for each question. Of those 224 billion parameters, just 12 billion switch on for any given task.

• Engram is Tencent's memory layer: 64 billion extra parameters that store frequently seen patterns so the model can pull them up instead of working them out from scratch. Think of a well-stocked pantry rather than re-shopping for every meal.

• Context window is how much you can hand over in one go. Here it's up to 1 million tokens, which is roughly a stack of full-length novels.

The model also reads images and generates at more than 150 tokens per second — comfortably faster than you can read.

Then there's the part most people actually care about, the API price per million tokens:

• Cached input: $0.003

• Input tokens: $0.070

• Output tokens: $0.281

Cached input is the killer line. If you keep reusing the same instructions or documents — a style guide, a product manual, a rulebook for an agent — that content gets billed at three-tenths of a cent per million tokens. That's the kind of number where you stop counting.

Hy4 Flash is available now through Tencent Cloud TokenHub, and Tencent says it plans to open the model's weights, which would let people run it on their own hardware. The technical write-up is on Tencent's research page, and Pandaily sums up the headline price as 2 yuan per million output tokens.

What it means for you

You don't need to be a developer for this to matter. Cheap models change what feels worth asking. When a query costs a fraction of a cent, you stop saving questions for later.

At home: the boring admin pile

Think about the little jobs nobody wants: decoding a utility bill, drafting a polite email to a landlord, making sense of a warranty receipt you photographed. Hy4 Flash takes images, so you can snap a photo of a confusing letter and ask what it actually means. At this pricing, a household could run hundreds of those questions a month for less than the price of a coffee.

At work: the long-document problem

A 1 million token context window is the real story for office work. That's a quarterly report, a full contract, a year of meeting notes or an entire product manual in one go. Instead of chopping a document into pieces and hoping the model remembers the beginning, you paste the whole thing and ask for a summary, a list of risks, or the three decisions buried on page 88. Because cached input is so cheap, a team can reuse the same reference document all day without watching the meter.

For a small business: cheap experiments

The old maths of AI for a small company was simple — every experiment cost real money, so you only tried the ideas you were confident about. At $0.070 per million input tokens, testing ten different chatbot scripts for your shop's FAQ costs less than the paper you'd print them on. If you want to sketch out prompts and campaign copy before you touch an API at all, the free tools at MyKreaTool are a sensible first stop.

For studying: a patient tutor

Students get the most from the image support. Photograph a textbook diagram or a page of handwritten notes and ask for a plain-English explanation, then ask it to quiz you. Generating practice questions is cheap on the output side, so you can pull fifty of them without thinking about it. A million output tokens is a small library of explanations.

For creative projects: the idea machine

Writing, naming, outlining, rewriting in a different tone — this is the grind where cheap tokens shine. Ask for twenty product names, hate all of them, ask for twenty more. Ten rounds of that costs pennies. The 150+ tokens per second speed matters here too: brainstorming feels like a conversation rather than waiting for a dial-up page to load.

For side income: thin margins that work

If you're freelancing or building a small tool, the output price is what decides whether a project makes money. At $0.281 per million output tokens, a client-facing assistant that answers a few thousand questions a month stays comfortably profitable even if you charge very little. That's the world where a solo operator can undercut an agency and still eat.

How to try it right now

You don't have to write code to look at this sensibly, and you can start for free.

1. Read the model card first (free). Open Tencent Hy4 Flash and skim the specs and the benchmark notes. No account, no card. It tells you straight away whether the model fits your kind of work.

2. Create a Tencent Cloud account. Sign up at Tencent Cloud and go to TokenHub, the marketplace where Hy4 Flash is listed. You'll need to add a payment method, but there's no minimum spend commitment mentioned in the announcement.

3. Generate an API key. In TokenHub, open the Hy4 Flash entry and create a key. Treat it like a password — anyone who has it can spend your credit.

4. Plug it into a chat client. Most desktop and browser chat apps that accept a custom API endpoint will take a Tencent Cloud base URL and your key. Paste it in, pick Hy4 Flash as the model, and you're talking to it.

5. Test with a real job, not a toy. Upload a document you genuinely need to understand and ask three questions about it. That's a better signal than any benchmark.

6. Watch the cached-input line. If you're building anything repetitive, put your standing instructions at the start of every request so they get cached at $0.003 instead of billed as fresh input.

Upsides and what changes

The obvious win is price. A 224 billion parameter model with vision, a 1 million token context and 150+ tokens per second, priced where a hobbyist can afford to leave it running, moves a lot of use cases from "someday" to "why not."

The bigger shift is psychological. Expensive models make people cautious — you write shorter prompts, you batch your questions, you skip the experiment. Cheap ones make people curious. Expect to see a wave of small, unglamorous tools built on this: inbox sorters, receipt readers, meeting-note summarisers, the assistant for a single niche nobody thought was worth serving.

And the Engram design is genuinely interesting. Storing recurring patterns so the model doesn't recompute them is the same logic as caching in web software, applied to thinking. If that idea spreads, the whole price curve bends down.

Limitations

Be clear-eyed about what this is and isn't. Every comparison here comes from Tencent's own blind tests, which is not the same as independent evaluation — and the company itself says Hy4 Flash falls behind its flagship Hy4 on hard programming problems and long, multi-step agent runs, so if you're shipping complex autonomous workflows, this isn't the model you reach for. It's the budget option for everyday work, not a frontier replacement. The weights aren't open yet (only planned), access runs through Tencent Cloud rather than the developer platforms you may already use, and there's an obvious data-residency question if you're handling client or personal information. Also worth noting: cheap output tokens don't fix a bad prompt. A model that answers instantly and wrongly just helps you make mistakes faster.

Conclusion

Hy4 Flash is one of those releases that matters less for what it does than for what it makes affordable. A 224 billion parameter model, a million tokens of context, image support, and output at $0.281 per million tokens — that combination pushes AI from a line item into loose change.

Here's your one action for today: open the Hy4 Flash research page, find the one repetitive task that eats twenty minutes of your week, and write down what you'd ask the model to do about it. That note is your first experiment, and it's ready the moment you get an API key.