Tabbit
ResourcesBlogModels
Tabbit LogoTabbit

Tabbit — The AI Browser that Works for You

Topics

  • Best Browsers 2026: Rankings and Top Lists
  • AI Browser Guides and Resources
  • Browser Comparisons: Head-to-Head Guides
  • Browser Reviews and Explainers
  • Browser Alternatives
  • Browser Directory: Guides by Browser
  • Agentic Browser Resources
  • Browser Downloads and Install Guides
  • Browser Productivity Resources
  • SillyTavern Model Guides and Presets

Popular Guides

  • Best Browsers 2026: Top 10 Ranked & Compared
  • Best Free AI Browsers 2026: Plans & Prices Compared
  • Rabbit Browser: What It Is & Why Tabbit Is Better
  • Firefox vs Chrome 2026: Privacy, RAM & Speed Compared
  • Best Lightweight Browsers 2026: 7 Tested for RAM & Speed
  • AI Browser List 2026: 10 Browsers Compared
  • Neobrowser: Norton Neo, Download & Alternatives
  • Edge vs Chrome 2026: Speed, RAM & Battery Tested
  • Best Internet Browser for Android 2026
  • Best AI Browser 2026: Top 9 Tested & Ranked

Events

  • Tabbit Skill Competition
  • KPOP SBTI Fandom Personality Test
  • Tabbit Campus Creator Program
  • fifi's Picks: AI Skills for Research Papers
  • User Survey

About

  • Tabbit Blog
  • Press & Media
English
简体中文English
Review
CommunityHy4

Artificial Analysis: Hy4 Intelligence Index v5 and Evaluation Costs

Original source

X; Artificial Analysis model and evaluation methodology pages

AuthorArtificial Analysis (@ArtificialAnlys)

Source date2026-11-05

Tabbit curation2026-11-05

Read original

One-sentence takeaway

On November 5, Artificial Analysis reported that Hy4 scored 35 on Intelligence Index v5, alongside results for agentic coding, knowledge question answering, and evaluation costs. The post did not specify a more precise v5 subversion. As of November 16, the methodology and model pages showed v5.0.1; the model page still listed a score of 35, but showed a cost of $1.03 per Index task, versus $0.78 in the original post. These are snapshots from different pages and should not be combined into a single cost figure.

Use cases

  • Tasks these results can help assess: Hy4's relative performance on Artificial Analysis's mainly English composite benchmark and the model costs calculated for that platform's Index tasks.

  • Tasks these results should not be extrapolated to: Predicting success rates on Chinese-language tasks or across all real-world workflows, or treating the score as a measure of overall capability and value in arbitrary product environments.

  • Applicable model version: The original post names Hy4 but does not specify the server-side build; it was a proprietary model at the time. As of the collection date, the model page identified it as Tencent Hy4, released in November 2026.

  • Test environment or client: Artificial Analysis's own benchmark and execution workflow; the Index includes agent-style and tool-use tasks. The original post does not specify Hy4's service provider or deployment build.

  • Reasoning level and parameters: Hy4's reasoning level and sampling parameters are not specified. Comparator settings are retained only where explicitly noted in the original post, such as Grok 4.7 (xhigh) and Claude Haiku 5.5 (max).

Evaluation method

The Hy4 model page provides a snapshot from the collection date. The Intelligence Index methodology explains how the composite score is calculated.

The methodology page currently states that Intelligence Index v5.0.1 comprises nine evaluations across four equally weighted categories, each accounting for 25%: knowledge work, coding, science, and general. The nine evaluations and their weights in the total score are: AA-Briefcase v1.1 (15%), AutomationBench-AA (5%), GDP.pdf (5%), FrontierCode-AA 2.0 (15%), Terminal-Bench 4.0 (10%), Terminal-Bench-Science 0.1 (15%), CritPt v2 Hard Subset (10%), AA-Omniscience v1.1 (15%), and HLE-Diamond (10%). The methodology page says the Index primarily uses English-language text evaluations; some evaluations also give image questions to models that support image input. Multilingual, audio, and vision capabilities are evaluated separately.

The task counts and number of runs currently disclosed on the methodology page are: AA-Briefcase, 91 tasks and 1 run; AutomationBench-AA, 657 tasks and 1 run; GDP.pdf, 100 questions and 5 runs; FrontierCode-AA 2.0, 132 questions and 3 runs; Terminal-Bench 4.0, 66 questions and 3 runs; Terminal-Bench-Science, 70 questions and 3 runs; CritPt v2 Hard Subset, 40 questions and 5 runs; AA-Omniscience v1.1, 3,000 questions and 1 run; and HLE-Diamond, 1,000 questions and 1 run. Scoring methods vary by evaluation and include pairwise comparisons, rubrics, automated tests, and answer verification. These are the methodology page's v5.0.1 details at the time of collection. The original post only identifies v5, so it is not possible to confirm that its exact subversion and per-evaluation execution parameters were identical.

Key results

  • Composite score and comparisons: Hy4 scored 35, matching Grok 4.7 (xhigh) and trailing Claude Haiku 5.5 (max) by 1 point, at 36. The original post called it the second-ranked model from a Chinese AI lab, behind Kimi K3.1 Pro at 46.

  • Coding: Hy4 scored 31% on FrontierCode 2.0, compared with 32% for Claude Fable 5.1 (max), 28% for DeepSeek V4.1 Flash 1017 (max), and 25% for Kimi K3.1 Pro. It scored 37% on Terminal-Bench 4.0; the original post said this ranked fourth among models from Chinese AI labs.

  • Knowledge and general reasoning: On AA-Omniscience v1.1, Hy4 had 50% accuracy, 1 percentage point above Kimi K3.1 Pro; Hy4's hallucination rate was 89%, compared with Kimi's 37%. Hy4 scored 33% on HLE-Diamond, versus 37% for Kimi. The accuracy advantage does not mean its answers are more reliable; the hallucination rate should also be considered.

  • Cost in the original post: Cost per Intelligence Index task was $0.78. In the post's comparison, Hy4 scored 35 and cost 10 times less than Grok 4.7 (xhigh), which also scored 35; it cost 13 times less than Qwen3.8 Max (0902), which scored 34; but cost nearly twice as much as Claude Haiku 5.5 (max). The post also reported input/output prices of $0.834/$2.501 per million tokens and a cached input price of $0.042.

  • Specifications: 1M-token context window; text and image input, text output.

  • Model-page snapshot on the collection date: The model page showed v5.0.1, an Index score of 35, and a cost of $1.03 per Index task. It listed input/output prices of about $0.83/$2.50 per million tokens. The page's current cost figure is higher than the value reported in the November 5 post; the page did not provide a historical cost snapshot aligned with the post or explain the difference in the content reviewed.

Raw data

MetricHy4ComparatorSource and basis
Intelligence Index35 (the post says v5)Grok 4.7 (xhigh) 35; Claude Haiku 5.5 (max) 36; Kimi K3.1 Pro 46X post, 2026-11-05
FrontierCode 2.031%Claude Fable 5.1 (max) 32%; DeepSeek V4.1 Flash 1017 (max) 28%; Kimi K3.1 Pro 25%X post, 2026-11-05
Terminal-Bench 4.037%4th among models from Chinese AI labsX post, 2026-11-05
AA-Omniscience v1.150% accuracy; 89% hallucination rateKimi K3.1 Pro: 49% accuracy; 37% hallucination rateX post, 2026-11-05
HLE-Diamond33%Kimi K3.1 Pro 37%X post, 2026-11-05
Cost per Index task$0.78Grok 4.7 (xhigh), same score, about 10 times Hy4's cost; Qwen3.8 Max (0902), 13 times higher; Claude Haiku 5.5 (max), about half Hy4's costX post, 2026-11-05; comparisons as stated in the post
Cost per Index task (later page snapshot)$1.03—Artificial Analysis model page, collected 2026-11-16
API pricesInput $0.834/M; output $2.501/M; cached input $0.042/M—X post, 2026-11-05; M means one million tokens

Replies in the X thread reiterated the FrontierCode 2.0 score, AA-Omniscience and HLE-Diamond data, and cost per task, and pointed to the Hy4 model page for a breakdown of individual evaluation scores. The original post also mentioned the 1-million-token context window and multimodal specifications. As of the collection date, the model page listed Index v5.0.1 and a score that remained 35; its cost figure differs from the original post and should be cited with its own date and page basis.

Conclusions and limitations

These data describe an Artificial Analysis composite-benchmark snapshot. The Index is primarily based on English-language tasks and combines nine heterogeneous evaluations using weights; its score cannot represent all languages, applications, or agent environments. Although Hy4's accuracy on AA-Omniscience was higher than the listed Kimi comparator's, its 89% hallucination rate was substantially higher; the accuracy figure should not be cited on its own. Cost comparisons depend on the evaluation tasks, actual token usage, price list, and page snapshot. The original post's $0.78 and the model page's $1.03 on the collection date differ. The public pages do not explain the change between the two figures, so they do not establish that costs rose under the same conditions, and neither figure should be treated as a universal cost for current use. The public materials also do not provide per-question outputs or complete run logs. This note records the platform's reported results and does not constitute an independent recalculation.

Reproduction notes

Open the Artificial Analysis X post from 2026-11-05 in Tabbit and expand the thread to check the original post and replies. Then review the Hy4 model page and Intelligence Index methodology page, recording the version, score, cost, and evaluation structure shown on 2026-11-16. The model and methodology pages are dynamic, so record the date again when revisiting them. Artificial Analysis has not published the complete outputs needed to recalculate Hy4's overall score and cost question by question; this record does not claim to have rerun the evaluation.

Curated by Tabbit

This is a third-party source navigator. Model versions, test environments, and personal experience vary; consult the original source.

Hy4

Use and compare models in Tabbit

Hy4

Related reviews

OfficialTencent Hunyuan official website, “Research”2026-11-04

Tencent Hunyuan: Hy4 Stable Release and Benchmark Results

MediaMercor APEX

Mercor APEX-Agents Professional Task Benchmark: Hy4 High Score Snapshot

MediaVals AI

Vals AI: Hy4's Cross-domain Results on Vals Index

MediaPlotly official blog2026-11-04

Plotly: Tencent Hunyuan Hy4 Data Analytics Benchmark

Hy4

Related showcases & prompts

OfficialTencent Cloud Model-as-a-Service TokenHub documentation

Tencent Cloud TokenHub: Hy4 GA Model ID, Reasoning, and Multimodal Parameters

OfficialOpenRouter model page and API configuration panel

OpenRouter: Hy4 Production Model API and Routing Configuration

OfficialOpenClaw documentation

Explicitly Select the Tencent Hunyuan Hy4 GA in OpenClaw

OfficialTencent Cloud TokenHub documentation

Connect Tencent Cloud Hy4 to OpenCode