A generated hero card comparing Hy4 flash and Claude Haiku 5.5, with a large figure '2.79 vs 2.81' labelled 'satisfaction out of 4 in Tencent's own 85-user blind test', and a footer reading 'Both satisfaction scores are Tencent's measurement. Hy4 flash weights published 17 November 2026.'
Guides & Insights

Hy4 flash vs Claude Haiku 5.5: Tencent Ran This Comparison Itself, and Published the Loss

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Somebody at Tencent decided that the right opponent for Hy4 flash was Claude Haiku 5.5, put both models in front of 85 of the company's own WorkBuddy users across 259 tasks, and published the result even though it loses. Hy4 flash scored 2.79 out of 4 on user satisfaction. Claude Haiku 5.5 scored 2.81. That is a two-hundredths-of-a-point gap on a four-point scale, measured by the challenger, on the challenger's own tasks — and it is the single most interesting number in the Hy4 flash release, because it tells you what Tencent actually believes it is selling. Not a better model. A cheaper one that is close enough that the difference does not survive contact with a satisfaction survey.

The rest of the release supports that reading. Hy4 flash was published as open weights on 17 November 2026, two days after it went onto Tencent's API on 16 November: 220B total parameters, 12B active per token, a 1M-token context, image input with text output. Claude Haiku 5.5 has been generally available since 7 October 2026 at $0.10 and $0.50 per million tokens. Neither model is on the OrcaRouter catalogue, and this piece will not pretend otherwise; what is routable is discussed near the end, and it is a different pair of models.

The blind test, read properly

Tencent's write-up of the evaluation is three sentences and one chart, and the chart is doing something specific: it plots satisfaction against median end-to-end completion time and cost per task, so that the model which wins on two axes is visible at a glance. The numbers are 2.79/4 for Hy4 flash against 2.81/4 for Claude Haiku 5.5, a median completion time of 6.7 minutes for Hy4 flash, and RMB 0.26 per task. Tencent's claim is that Hy4 flash recorded both the lowest cost per task and the shortest median completion time among the evaluated models.

What the write-up does not give you is the full set of evaluated models, the per-category breakdown, or the rubric. It does say the population: 85 internal users, 27 categories, 259 tasks, all drawn from WorkBuddy, Tencent's own productivity assistant. That is a coherent evaluation and it is not a neutral one. Employees evaluating their employer's model know which model is theirs, and even in a blind interface the tasks are the ones Tencent chose because they resemble Tencent's product.

The useful thing about the 0.02 gap is what it does to the argument either way. If you are inclined to trust the test, you have Tencent's word that the capability difference against Claude Haiku 5.5 is negligible for office work, which is a claim against interest. If you are inclined to distrust it, the same test gives you no reason to believe Hy4 flash is better, and you are left with the cost column, which is where Tencent wants your attention anyway.

Where the two models actually differ

The spec sheets diverge in ways the satisfaction score cannot show. Every figure below is the vendor's own published specification unless it is marked as an Artificial Analysis measurement.

• Total and active parameters — Hy4 flash 220B total with 64B of that in Engram tables, 12B active per token; Claude Haiku 5.5's parameter counts are undisclosed, which is normal for Anthropic.

• Context window — 1,048,576 tokens for Hy4 flash; 1,000,000 for Claude Haiku 5.5.

• Maximum output — 64,000 tokens on Hy4 flash per the serving config; 128,000 tokens on Claude Haiku 5.5, twice as much.

• Input modality — Hy4 flash takes text and images through a SigLIP2-derived 27-layer encoder; Claude Haiku 5.5's release material centres on text and tool use, and its published capability set is narrower on the vision side.

• Licence — Hy4 flash ships under the Tencent Hy License with a Section 10.2 that requires a separate deal above $50 million in MaaS revenue or 100 million MAU; Claude Haiku 5.5 is proprietary and has no weights at all.

• Independent score — Hy4 flash has none, and its Artificial Analysis URL returns 404; Claude Haiku 5.5 scores 36 on Intelligence Index v5.0.1 at Max effort.

• Measured cost per Index task — $0.41 for Claude Haiku 5.5 per Artificial Analysis; not measured for Hy4 flash, which is not on the board.

The output ceiling is the row most likely to bite in practice. A model with 1M of context and 64k of output is configured for reading a great deal and writing a moderate amount; a document-drafting workflow that expects to emit more than 64,000 tokens in one response cannot use it. Tencent's own recommended use — deep search, document processing, slide creation, simple frontend demos, computer use — is consistent with that ceiling. Long-form generation is not on the list.

The price comparison is not one number against another

Tencent's mainland rate for Hy4 flash is ¥0.5 per million input tokens, ¥2 per million output, and ¥0.02 per cached input token — about $0.070, $0.281 and $0.003. Claude Haiku 5.5's published card starts at $0.10 per million input and $0.50 per million output, with cache writes at $0.125 and $0.625 for five- and one-hour caches and cache reads at $0.01.

On those headline rates Hy4 flash is roughly 30% cheaper on input, 44% cheaper on output and 70% cheaper on a cache hit. Both are flat rates on the first hundred thousand tokens. The difference is what happens after that: Anthropic's card steps Claude Haiku 5.5's input price from $0.10 to $0.50 per million and its output price from $0.50 to $2.50 once a request crosses 100,000 input tokens. On a 200,000-token document, Claude Haiku 5.5 is billed at five times its headline input rate, which puts that comparison at $0.50 against Tencent's $0.070 — a 7x gap rather than a 30% one.

That is a design choice, not a trick, and it has a defensible logic: long inputs cost more to prefill, so the vendor charges more for them. But it means the headline number is the wrong number for exactly the workload both models claim to be good at. Document processing at 1M context is a >100,000-token workload by definition. If your use case is long-document work, price both models on the rows that apply to your prompt length, not on the top of the card.

Latency: the metric Tencent measured is not the one Anthropic publishes

Tencent reports a median end-to-end completion time of 6.7 minutes for Hy4 flash on a WorkBuddy task. Artificial Analysis reports Claude Haiku 5.5 at 14.7 minutes per Intelligence Index task at Max effort, with a 192-second time to first answer token — and the board flags the model as "notably slow per task and very verbose," at 298M output tokens for the index run.

Those two numbers are not comparable and should not be subtracted. Different tasks, different harnesses, different definitions of "done." What they do establish is that Claude Haiku 5.5's effort setting is a latency dial with a very wide range, and that the top of it is expensive in wall-clock time:

• Max — Index 36, 14.7 minutes per task, 192-second first token, 298M output tokens.

• Xhigh — Index 35, 9.6 minutes, 44-second first token, 151M tokens.

• High — Index 30, 6.4 minutes, 21-second first token, 89.8M tokens.

• Medium — Index 25, 3.9 minutes, 10-second first token, 52.7M tokens.

• Low — Index 19, 2.2 minutes, 5.8-second first token, 29.9M tokens.

A generated reference card headed 'The effort ladder Artificial Analysis measured', with five horizontal bars for Claude Haiku 5.5 at Max, Xhigh, High, Medium and Low effort, each carrying its Intelligence Index score, minutes per task and seconds to first token, and a vertical marker at the High row labelled 'the setting most products would ship'.

All five rows are Artificial Analysis measurements on v5.0.1. The point of printing them is that "Claude Haiku 5.5 scores 36" is a statement about the most expensive configuration Anthropic offers. At its high setting it scores 30 and finishes in less than half the time. Anyone comparing a latency-sensitive product against the Max number is comparing against a setting they will not ship.

Hy4 flash has the same dial, in a coarser form: reasoning_effort at high, which is the default, or no_think. Two positions against five. If your workload needs a middle setting — thinking on, but not maximally — Claude Haiku 5.5 gives you the knob and Hy4 flash mostly does not.

Screenshot of the Artificial Analysis model page for Claude Haiku 5.5, showing the five effort rows from Max down to Low with their Intelligence Index scores, time per task and output token counts.

Three things that decide this and are not on either spec sheet

The API surface. Claude Haiku 5.5's request format is stricter than its predecessor's in ways that break code. Anthropic's documentation rejects temperature, top_p and top_k at any non-default value, drops assistant prefill, removes manual budget_tokens, and requires a specific computer-use toolset version. Hy4 flash accepts the full sampling set, recommends temperature 0.9 with top_p 1.0, and ships hy_v4 tool-call and reasoning parsers for vLLM. If you have a prompt library tuned with temperature sweeps, one of these models will run it unchanged and the other will 400.

The deployment shape. Claude Haiku 5.5 is an API and nothing else. Hy4 flash is an API, a Hugging Face repository, an FP8 build, a base checkpoint, a GGUF quantisation and a fine-tuning guide. That is not automatically an advantage — a model you self-host is a model you operate, patch and capacity-plan — but it is the difference between a vendor relationship and an asset, and for a team with a data-residency requirement or a fine-tuning need it settles the comparison before any benchmark is consulted.

The measurement gap. Claude Haiku 5.5 has been on Artificial Analysis since October and has been picked apart by every comparison blog on the internet, this one included. Hy4 flash has a 404 where its score should be. Not because the model is bad, but because it is two days old and nobody independent has run it. The satisfaction score above is a measurement of a real population, and it is also the only measurement of this model by anyone other than its maker.

Calling either of them, honestly

Neither Hy4 flash nor Claude Haiku 5.5 is on the OrcaRouter catalogue as this is written, so there is no version of this section where we tell you to call them here. What is there is the generation on each side: tencent/hy4-preview, the 770B/49B model Hy4 flash was distilled from, at $0.834 input and $2.501 output per million with a $0.042 cache-read rate; and anthropic/claude-haiku-4.5 at $1.00 input and $5.00 output per million with a $0.10 cache read — the previous generation, not the 5.5 in this comparison.

That pairing is more useful than it sounds. If you are deciding whether Hy4 flash's cost profile is worth restructuring a pipeline for, the preview model gives you the family's behaviour at a price where the experiment is cheap, on the same key and the same endpoint as everything else in the catalogue, with the provider's list price passed through at 0% markup so a Tencent price change reaches you the day it happens. The routing DSL is the part that matters once you have two candidates: send the long-document tier to whichever model wins your own eval and keep the other as the fallback, instead of choosing one vendor for everything because switching costs a contract.

And when Hy4 flash lands on a catalogue — any catalogue, ours included — the thing to test is not the satisfaction score. It is the output-token count on your own tasks. Tencent's post-training claims a 50% reduction in average end-to-end output tokens, and if that holds on your workload it is worth more than a 30% per-token discount, because output tokens are the expensive half of every bill on both models.

Screenshot of the anthropic/claude-haiku-4.5 model page, showing the per-million-token input, output and cache-read prices the previous generation is listed at.

What would settle it

One independent run on a shared harness. Put Hy4 flash and Claude Haiku 5.5 through the same nine-evaluation Intelligence Index and look at two rows: the score, and cost per completed task. Tencent has already told you what it expects those rows to say — that Hy4 flash lands near Claude Haiku 5.5 on capability and far below it on cost. Until somebody outside Tencent publishes the pair, that expectation is a hypothesis with a company's marketing budget behind it, not a result.

What you can act on today is narrower and more useful. If your workload is long documents, price the >100,000-token row before anything else, because that is where the two cards diverge most. If it is latency-sensitive, compare Claude Haiku 5.5 at high effort rather than Max. If it needs fine-tuning or on-premises inference, the decision is already made and no benchmark will change it. The satisfaction gap of 0.02 points is not a reason to switch in either direction, and Tencent publishing it is the most credible thing in the release.