Key takeaways
- OpenRouter dates the tencent/hy4-flash listing to November 16, 2026. On November 17 it exposes one Tencent-hosted endpoint with text and image input and text output.
- The registry specifies 1,048,576 tokens of context and up to 65,536 completion tokens. These are serving limits, not evidence that long-context answers are reliable.
- The displayed 50% promotion gives rates of $0.035 input, $0.1405 output and $0.0015 cache read per million tokens. No promotion end date was verified.
- The endpoint allows automatic tool choice but reports required and named-function tool choice as unsupported. Test existing agent scaffolds before switching models.
What is available now
Tencent’s Hy4 Flash is listed on OpenRouter with a November 16, 2026 release date and the API identifier tencent/hy4-flash. When checked on November 17, OpenRouter exposed a single Tencent endpoint, labeled Tencent Cloud on the model page. It accepts text and images and returns text. This is a newly listed paid API option, not a free stealth model.
OpenRouter describes Flash as a cost-efficient Hy4 variant served with roughly a quarter of the flagship’s parameter count, aimed at code assistance, agent pipelines, document and chart comprehension, and routine automation. That size comparison and the implied latency benefits are platform descriptions, not AccessAllGPT measurements. The sources reviewed do not establish an exact total or active parameter count.
The date here is OpenRouter’s listing date. We did not verify a separate Tencent announcement establishing a worldwide release date or the same offer on Tencent’s direct API. Builders can evaluate this OpenRouter route without assuming identical access, prices or terms elsewhere.
A million-token window, with a separate output ceiling
The public registry and endpoint record specify a 1,048,576-token context window and a maximum of 65,536 completion tokens. The endpoint does not publish a separate maximum prompt-token value in the reviewed response. Do not interpret the context number as permission to submit that many input tokens and then add the entire output allowance on top.
The endpoint is tagged FP8. That is a property of the hosted serving route, not a downloadable artifact or an independently verified model architecture. Text-and-image input also does not mean image generation, audio input, video processing or speech output; those modalities are not listed for this model.
Our analysis: the interesting use case is inexpensive document and chart processing with enough context for substantial source material. But a large advertised window does not establish evidence recall, image grounding or instruction-following near its limit. Start with representative documents and known-answer questions rather than buying a context-length claim as a quality guarantee.
The low prices are a promotion
OpenRouter’s provider table shows a 50% discount. At review time, the resulting per-million-token rates were $0.035 for input, $0.1405 for output and $0.0015 for cache reads. The same table shows undiscounted rates of $0.07, $0.281 and $0.003 respectively. These are the observed OpenRouter route prices, not a verified Tencent direct-price schedule.
The public endpoint record independently exposes a discount value of 0.5 alongside the effective per-token rates. We converted those values to per-million-token units to check the page. No promotion end date was verified. Budget for a rate change instead of assuming this launch price is permanent.
OpenRouter also displays weighted-average prices derived from actual usage. Those are not the quoted price for a fresh request: cache mix and discounts can change the average. The endpoint reports implicit caching support, but we did not test cache eligibility, cache hits or billing. Compare actual cost per accepted task, including rejected answers and retries, before declaring this the cheaper production choice.
Agent builders should check tool-choice modes
The model registry lists tools, tool_choice, structured_outputs, response_format, reasoning and include_reasoning among supported parameters. The model-specific endpoint record adds an important qualification: tool-choice modes none and auto are supported, while required and function are false. A parameter appearing in a supported list does not mean every value of that parameter works.
That matters for agents which require a tool call at a particular step or force one named function. Such a scaffold may not be a drop-in fit even if ordinary automatic tool use works. We did not send a request to determine whether an unsupported mode is rejected, transformed or otherwise handled. Do not assume a particular failure behavior without a test.
The model registry describes reasoning as enabled by default but not mandatory. The Flash endpoint’s supported list does not include reasoning_effort, unlike the flagship Hy4 listing inspected alongside it. Avoid copying the flagship’s reasoning-effort settings into Flash merely because the names share a family. Verify the exact request contract your application needs.
One provider and no verified self-hosting package
The model page says requests go to one provider without routing. At review time the endpoint record likewise contained only Tencent. OpenRouter availability therefore does not provide same-model redundancy across multiple independent hosts for this listing. An application still needs an explicit fallback plan if that route is unavailable.
The registry has no Hugging Face identifier for Flash. We did not verify a downloadable weight package, license, model card, training recipe or safety report. A candidate repository API lookup returned HTTP 401, which is inconclusive; it is not proof that weights do or do not exist. Do not inherit the flagship’s artifact claims, license or deployment requirements.
No quality benchmark or performance comparison was reproduced for this briefing. OpenRouter’s description is useful for understanding intended use, but it does not establish that Flash matches Hy4 on coding, vision, long-context work or multi-step tool reliability. Lower token prices and a smaller reported model are reasons to evaluate, not evidence to migrate automatically.
What to do next
Our recommendation is a bounded trial for cost-sensitive text-and-image workloads that can use automatic tool selection. Keep the current model as a baseline. Send only non-sensitive inputs until processing terms are approved, and test exact tool schemas, output formats, image types and context lengths used by your application.
Record answer acceptance, unsupported claims, chart-reading mistakes, tool success, schema validity, response time, retries and actual billed tokens. If the workflow requires forced tool calls, verified self-hosting artifacts or multi-provider same-model failover, resolve that requirement before expanding traffic. AccessAllGPT has not tested inference or paid for model calls for this report.
A short Hy4 Flash trial checklist
Use the exact OpenRouter route rather than treating the Hy4 family name as an API contract.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Select a text-and-image workload with known-answer examples and a current-model baseline.
Check whether auto or none is sufficient. Do not assume required or named-function selection is supported.
Snapshot promotional rates; inspect actual usage and cache hits; include retries and a possible end to the discount.
Set a small traffic ceiling, an alternate model and measurable acceptance criteria. Wait if a license or self-hosting package is required.
Primary sources
Browse the publication-wide evidence index →
- Tencent: Hy4 flash model pageOpenRouter · Reviewed: November 16 release date; full model description; provider count; price table and 50% discount · Retrieved · Supports: Primary evidence for the OpenRouter listing, advertised use cases, single Tencent Cloud route and displayed effective and undiscounted prices; not independent validation of quality.
- OpenRouter public models registryOpenRouter · Reviewed: tencent/hy4-flash record; modalities; context; output limit; reasoning defaults; parameter list; null Hugging Face ID · Retrieved · Supports: Machine-readable model ID, text-and-image input, text output, serving limits, effective pricing and reasoning defaults; registry metadata does not prove successful inference or downloadable artifacts.
- Hy4 Flash provider endpoint recordOpenRouter · Reviewed: Tencent endpoint; FP8; discount; tool-choice mode flags; implicit caching; output ceiling · Retrieved · Supports: Current single-provider route, effective per-token rates and explicit unsupported required/function tool-choice modes; availability metadata is not an authenticated hands-on test.
Limitations
Public-source availability briefing, not hands-on testing. No authenticated request, tool invocation, image task, long-context evaluation, cache hit, bill, latency measurement or benchmark was performed. OpenRouter is authoritative for its own listing and route metadata, not independent validation of Tencent capability claims. A separate Tencent launch announcement, promotion expiry, direct API terms, downloadable weights, license, safety report and self-hosting requirements were not verified. A candidate Hugging Face API returned HTTP 401 and is inconclusive. Prices and endpoint behavior may change after retrieval.
Disclosures
AccessAllGPT received no payment, API credits, private access or vendor review for this briefing. Tencent and OpenRouter did not review or endorse it. AccessAllGPT Research is operated by NeuralArc and is independent of these vendors. Publication-wide relationships are listed on the disclosures page.



