Hy4 flash and GPT-6 Luna are both positioned for frequent, cost-sensitive work. The best direct evidence found is Plotly’s single-run data-analysis test: Hy4 flash scored 39 of 43 questions and GPT-6 Luna 38 of 43. The reported run cost was $0.13 for each; GPT-6 Luna had a shorter median per-question time, 18 seconds versus 29.
That one-point score gap is not a stable ranking. Plotly says repeated runs of other models have varied by as many as three points. The test questions and data are private, and model settings were not fully reported. Use it as a useful task-specific signal, not a claim that Hy4 flash is broadly smarter or GPT-6 Luna always faster.
Key takeaways
Plotly reports a close result on the same harness: 39/43 for Hy4 flash and 38/43 for GPT-6 Luna, one run each.
Both runs cost $0.13 in Plotly’s setup; this benchmark total is not a per-token price comparison.
Tencent lists lower base input/output rates than OpenAI, but task cost also depends on tokens, cache, provider and retries.
The official product docs for GPT-6 Luna describe API tools and model settings; Tencent positions Hy4 flash for everyday workflows and Hy4 for harder long-horizon engineering.
Neither the Plotly score nor vendor claims replace testing your own tasks.
At a glance
| Published detail | Hy4 flash | GPT-6 Luna |
|---|---|---|
| Provider model ID | hy4-flash | gpt-6-luna |
| Context window | Up to 1M tokens | 1.05M tokens |
| Maximum output | 64K tokens | 128K tokens |
| Input/output price per MTok | $0.070 / $0.281 | $0.10 / $0.50 |
| Plotly score, one run | 39/43 | 38/43 |
| Plotly run cost | $0.13 | $0.13 |
| Median time per question | 29 sec | 18 sec |
Hy4 flash’s context, parameter count, list prices and promotion expiry are in Tencent’s launch page and model page. GPT-6 Luna’s 1.05M context, 128K output and $0.10/$0.50 rates are in OpenAI API docs. Tencent was offering a time-limited 50% TokenHub discount through November 30, 2026; do not treat its promotional rate as the standard price.
What the matched test says
Plotly ran its private 43-question DataAnalyticsBench through Plotly Studio Embedded using the same exam, harness and scoring rules. Hy4 flash answered 39 correctly; GPT-6 Luna answered 38. Plotly reports one run for each model. Its benchmark covers standard analysis, ambiguity, impossible requests, data-quality issues, statistical traps and hidden patterns. Both systems struggled with some hidden-pattern questions.
The cost and timing comparison is bounded by the test. Plotly reports $0.13 for each run, median times of 29 and 18 seconds per question, and different average call/output-token counts. Differences in cache hits and model output affect the total. The private exam, missing complete run logs and undocumented settings limit reproduction. The result does not show expected cost or speed for a different prompt, tool loop or provider.
Published token rates are not the whole bill
At list prices, for a simple request with 1 million uncached input tokens and 100,000 output tokens, token arithmetic is $0.0981 for Hy4 flash and $0.15 for GPT-6 Luna, before tools or provider-specific fees. This calculation uses the published standard rates and ignores cache. It is illustrative arithmetic, not the cost of a tested task.
| Example rate arithmetic | Hy4 flash | GPT-6 Luna |
|---|---|---|
| 1M uncached input tokens | $0.070 | $0.100 |
| 100K output tokens | $0.0281 | $0.050 |
| Subtotal, no cache/tool charges | $0.0981 | $0.1500 |
Hy4 flash’s Tencent TokenHub promotion was scheduled to end November 30, 2026. GPT-6 Luna has separate cached-input prices and processing modes in its official model page. Recheck the current endpoint, region, cache policy and discounts before estimating a production workload.
How to decide for your workload
Start from the task, not the benchmark winner. For high-frequency simple steps, measure the cost of a completed task, including retries and human correction. For code changes and long workflows, check whether the model can follow your tools and constraints; a low token price does not compensate for repeated failures. For multimodal work, verify exact input formats and any image-specific limits in the endpoint you will use.
Tencent recommends Hy4 flash for everyday work such as document processing and simpler computer-use tasks, and the larger Hy4 for complex long-horizon software engineering. OpenAI positions GPT-6 Luna for focused, high-volume tasks and documents Responses API tools. These are provider recommendations, not independent proof of comparative quality.
A practical comparison workflow
Use the same prompt and source material when you test a model on your own work.
For adjacent topics, read our guides to AI browser comparison, agentic browsers, browser automation, agent browser research, deep research workflows, agent cost per success, browser agent failures, and browser use vs computer use.
Verdict
Plotly’s one-run result is close: Hy4 flash scored one point higher, while GPT-6 Luna’s median per-question time was 11 seconds lower. The benchmark also reports the same total cost for each run. Since Plotly notes up to three points of variation in repeated tests of other models, none of those differences is enough to make a general winner claim. Choose by current endpoint access, regional pricing and your own repeated workload tests.
Sources
Tabbit Browser offers Hy4 Flash. Download Tabbit, then select Hy4 Flash from the model list in Chat to start using it.
FAQ
Which model scored higher on Plotly’s data-analysis test?
Hy4 flash scored 39/43 and GPT-6 Luna 38/43 in one run each. Plotly warns that single-run scores may vary by up to three points.
Which model costs less per token?
Tencent’s published standard rates are $0.070 input and $0.281 output per million tokens; OpenAI lists $0.10 and $0.50 for GPT-6 Luna. Region, cache, promotions, provider fees and output volume change the total.
Was the Plotly run cheaper for Hy4 flash?
Plotly reported $0.13 for each model in its one-run benchmark. The test’s workload and cache pattern differ, so this does not establish equal cost on other tasks.
Is Hy4 flash open source now?
Tencent said it plans to open-source Hy4 flash. The launch announcement does not confirm that weights and a license are already available.
How should I choose?
Run repeated, representative tasks using fixed prompts, tools, endpoints, region and cache rules. Compare correctness, retries, time and the final invoice.