Tencent announced Hy4 flash on November 16, 2026 as an economical option for everyday work. Its stated standard API prices are approximately $0.070 per million input tokens and $0.281 per million output tokens. The model accepts multimodal input and supports up to one million tokens of context. Tencent describes 224 billion total parameters, with 12 billion active. Open distribution remains a future plan rather than a verified downloadable release.

Read the price as a completed-work question

The useful comparison is the cost of an accepted result. If two models need different numbers of attempts to produce a correct spreadsheet, cheaper tokens can still produce a more expensive task. A tariff-only comparison leaves retries, corrections and human review outside the calculation. This is our editorial evaluation framework, not a measurement SIExpo has performed on Hy4 flash.

Tencent presents WorkIndex as an internal evaluation of practical workplace tasks. Its numbers should therefore remain attributed vendor results. They do not automatically establish performance on another organization’s documents, particularly when language, file structure and acceptance criteria change. We have not reproduced the evaluation and do not infer general leadership from it.

A sensible pilot would give each candidate the same anonymized assignment and define the required final artifact in advance. For example, a report might need correct formulas, preserved units and explicit identification of missing data. Count success after checking the delivered file, rather than when the agent says it has finished. That connects token expenditure to an actual useful outcome.

The pilot should also retain unsuccessful attempts. Excluding retries can make a fragile workflow appear economical. Time spent reviewing an answer matters as well: a fast draft that requires extensive repair may be less useful than a slower result that meets the specification. These are suggested assessment steps, not claims about observed model behavior.

Context capacity and availability are different properties

A large context window specifies how much material a request can accommodate. It does not by itself establish that every number in a long document will be retrieved and applied correctly. Evaluation can include questions about the beginning, middle and end, together with a question whose answer is absent. The last case tests whether the system acknowledges missing evidence.

Parameter counts are not a direct memory budget for a local computer either. The active fraction describes the computational path, while access to the complete architecture and its storage format are separate conditions. This announcement consequently cannot support a promise that the model runs on a particular consumer graphics card.

The concrete development is an additional choice for routine AI work. Its practical value depends on the quality of the final artifact, the permissions granted to an agent and the real bill for local workloads. The published prices are a snapshot of the primary announcement, not a guarantee of future commercial terms.

Editorial verification: 2026-11-16T17:16:38.474339+00:00.