Key takeaways
- OpenRouter’s registry records tencent/hy4 as created on November 4, 2026. One active endpoint, tencent/hy4-20261104, is served directly by Tencent and was reporting available during review.
- The route accepts text and image input and returns text. Its registry advertises 1,048,576 tokens of context, up to 65,536 completion tokens, reasoning controls, structured output and tool parameters.
- OpenRouter displayed a 50% promotional price of $0.417 per million input tokens and $1.2505 per million output tokens, plus $0.021 per million cache-read tokens. The undiscounted fields were $0.834, $2.501 and $0.042.
- The endpoint is tagged FP4 and OpenRouter links to a Tencent Hugging Face model path, but that path returned 404. Do not treat the listing as proof of downloadable weights, architecture details, a license or self-hosting readiness.
- Trial Hy4 behind a reversible routing flag on representative multimodal, long-context and tool-use tasks. Wait for a durable Tencent model card before relying on vendor benchmark, training, safety or open-weight claims.
Hy4 became callable on November 4
OpenRouter added tencent/hy4 on November 4, 2026. Its live registry exposes one endpoint—tencent/hy4-20261104—served by Tencent, and the page labels the release as the official successor to Hy4 Preview. The route was available in the registry during AccessAllGPT’s November 5 review.
This is a real access change for builders: the model can be selected through OpenRouter’s API rather than merely appearing in an announcement. It is not, however, a complete release package. AccessAllGPT found no working first-party model card linked from the listing and did not find a separate Tencent launch document to establish architecture, training data, safety methods or benchmark methodology.
The registered contract combines long context, vision and tools
The registry identifies the model as text+image to text with a 1,048,576-token context window and a 65,536-token maximum completion. Supported request fields include reasoning, reasoning_effort, response_format, structured_outputs, tools and tool_choice, as well as temperature, stop and token limits.
The endpoint reports automatic tool choice and “none” as supported, while required tool choice and forcing a named function are marked unsupported. That distinction matters for agents: an API accepting a tools array does not mean every tool-selection control works. Builders should test both schema adherence and whether the model calls, skips and retries tools appropriately.
The launch price is explicitly discounted
During review, OpenRouter showed a 50% discount: $0.417 per million input tokens, $1.2505 per million output tokens and $0.021 per million cache-read tokens. The endpoint record exposes undiscounted values of $0.834, $2.501 and $0.042 respectively. The page’s effective weighted-average chart is a usage-derived display, not the contract rate for a new request.
Treat the discount as mutable rather than a permanent unit-cost promise. A one-million-token window also does not make million-token requests economical or reliable. Measure actual prompt, cache and output use per accepted task, and alert when the endpoint’s published rate changes.
FP4 is a serving fact, not an open-weight release
OpenRouter tags the Tencent endpoint as FP4. That describes the hosted route’s quantization, but the listing does not disclose total or active parameter count, architecture, training corpus, post-training recipe, hardware requirements or an artifact license. The model page’s “Model weights” link pointed to huggingface.co/tencent/Hy4, which returned 404 on November 5.
Accordingly, buyers can evaluate a managed API today but cannot infer self-hosting readiness. A dead repository link is not evidence that weights are downloadable, and a previous model’s license or serving recipe cannot be carried forward to Hy4. Wait for the actual files, model card and license before making open-weight, sovereignty or on-premises plans.
Performance claims still need workload evidence
OpenRouter’s description says Hy4 improves over Hy4 Preview on agentic and long-horizon productivity tasks and adds visual understanding. No benchmark table, harness, comparison policy or safety report accompanied the registry evidence reviewed for this article. Those statements should therefore be treated as vendor/platform claims, not independent findings.
The public page also displayed recent provider telemetry, but short-window latency, throughput and uptime are snapshots that can change with traffic. They do not establish output quality, long-context recall, image reasoning, tool reliability or production service levels. AccessAllGPT did not call the endpoint.
What builders and buyers should do next
Trial tencent/hy4 against a frozen set of ordinary and long prompts, images, structured-output schemas and multi-step tool tasks. Record exact endpoint, reasoning effort, context length, cache use, tool-choice mode, completion length, latency, retries, schema validity, tool success, factual errors and human acceptance. Include adversarial images and malformed tool results where the use case permits.
Constrain the route behind a feature flag until pricing stability, data handling, quality and failure behavior are understood. Wait if the decision depends on downloadable weights, a documented license, independent benchmarks or a Tencent safety report. Reject automatic migration from Hy4 Preview unless representative outcomes improve enough to offset compatibility work and variable promotional pricing.
Copy-ready Hy4 API trial record
Use one record per workload so a registry launch becomes a measured adoption decision rather than an assumed upgrade.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Users, task, image inputs, context distribution, tools, volume, latency target and failure consequence.
tencent/hy4 and observed endpoint ID; OpenRouter API version, provider policy, retrieval date and fallback route.
Reasoning effort, temperature, max completion tokens, response schema, tools, tool-choice mode and cache behavior.
Representative prompts, images, long-context needles, tool successes and failures, invalid schemas and competitor baseline.
Acceptance, factual errors, image-grounding errors, schema validity, tool success, latency percentiles, retries and token use.
Published list and discount rates, input/output/cache tokens, cost per accepted task, promotion expiry assumption and spend alert.
Missing model card, weights, license, architecture, benchmark methodology, safety report, data terms or service commitment.
Trial, constrain, wait or reject; traffic ceiling, rollback trigger, owner, evidence needed and retest date.
Primary sources
Browse the publication-wide evidence index →
- Tencent: Hy4 model pageOpenRouter · Reviewed: Model identity; release date; description; modalities; context; listed price; provider; pricing history; latency, throughput and uptime displays; model-weights link · Retrieved · Supports: OpenRouter lists tencent/hy4 as released November 4, 2026, with text and image input, text output, a 1,048,576-token context window and one Tencent-hosted endpoint. The page links to a Tencent Hugging Face path that returned 404 during review.
- Tencent Hy4 endpoint registry recordOpenRouter API · Reviewed: Model and endpoint identifiers; creation timestamp; architecture; context and completion limits; provider; quantization; price fields; supported parameters; cache and tool flags; endpoint status · Retrieved · Supports: The live registry returned model tencent/hy4 and endpoint tencent/hy4-20261104, served by Tencent in FP4, with 1,048,576 context, a 65,536-token maximum completion and support flags for reasoning, structured output and tools.
- OpenRouter API reference overviewOpenRouter Documentation · Reviewed: API endpoint; authentication; OpenAI SDK compatibility; request headers; model selection; response and error shapes · Retrieved · Supports: OpenRouter documents its OpenAI-compatible chat-completions route and bearer-key authentication. This supports the access path, not any Hy4 capability claim.
Limitations
This is a public registry and documentation review, not hands-on model testing, an independent benchmark, a privacy assessment, a security audit or a procurement review. AccessAllGPT did not call tencent/hy4; spend API credits; verify authenticated billing; test text, images, reasoning, long context, tools, structured output or caching; measure quality, hallucination, latency, throughput, uptime or cost; compare Hy4 with Hy4 Preview or another model; inspect training data, architecture, system prompts, moderation or safety controls; verify data retention or regional processing; obtain an SLA; or inspect weights. The linked Hugging Face repository returned 404, so no weight presence, license, artifact identity or self-hosting claim was verified. OpenRouter and Tencent control the cited serving surfaces, and qualitative capability statements remain vendor/platform claims.
Disclosures
AccessAllGPT received no model access, API credits, private demo, briefing, benchmark data, review, payment or compensation for this article. Tencent and OpenRouter did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or certify Hy4. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Tencent or OpenRouter. Publication-wide relationships are listed on the disclosures page.
