AI Model Daily - 2026-08-05 | QName AI

30-Second Summary

I did not find a verified new foundation-model launch from OpenAI, Anthropic, or Google DeepMind today; the main story moved to secondary validation for already released models entering leaderboards and provider benchmarks.
Arena's 2026-08-04 changelog added three relevant entries: qwen-image-3.0-pro entered Text-to-Image Arena, qwen3.8-max entered Image-to-WebDev, and Deepseek V4 Flash (High) (20260731) entered Agent Arena.
Qwen-Image-3.0-Pro ranks #5 in Text-to-Image Arena at 1263+-11 Preliminary, only one point behind gemini-3.1-flash-image (nano-banana-2) at #6 with 1262+-5; image-model competition is shifting from "good-looking images" toward small text, multilingual rendering, UI, and layout productivity.
Artificial Analysis surfaced the Endpoint Accuracy Index with GLM-5.2, gpt-oss-120b, and DeepSeek V4 Pro as the first covered models. The core finding is that the same open-weight model can show large accuracy, output length, speed, and price differences across serverless API endpoints.
Tencent Hunyuan officially released Hy ASR 3.0 preview, now available through Tencent Cloud APIs and Yuanbao. This is a speech-recognition model release rather than an LLM leaderboard event, but it belongs in multimodal and voice-model tracking.

Must Watch

1. Arena August 4 Update: Qwen Image, Qwen3.8-Max, and DeepSeek V4 Flash High Enter New Tracks

Event type: Arena appearance / leaderboard change.
First/official/availability time: Arena Leaderboard Changelog records the update date as 2026-08-04.
Capability change: qwen-image-3.0-pro was added to Text-to-Image Arena; qwen3.8-max was added to Image-to-WebDev; Deepseek V4 Flash (High) (20260731) was added to Agent Arena.
Access: Arena changelog, Text-to-Image leaderboard, Image-to-WebDev leaderboard, and Agent Arena.
Impact: The important signal today is not another vendor launch but third-party placement into narrower tracks. Qwen3.8-Max is extending from WebDev/Text/Vision into Image-to-WebDev, making it a stronger candidate for vision-to-frontend-code agentic coding tests. DeepSeek V4 Flash High entering Agent Arena needs follow-up to see whether it becomes a low-cost agent contender.
Sources: Arena changelog, Text-to-Image Arena, Image-to-WebDev leaderboard, Agent Arena.

2. Qwen-Image-3.0-Pro: #5 in Text-to-Image, Almost Tied with Nano Banana 2

Event type: Arena appearance / X first signal / access update.
First/official/availability time: Qwen's official blog was published on 2026-07-20; Arena and @Alibaba_Qwen promoted its Text-to-Image Arena entry on 2026-08-04.
Capability change: Arena lists qwen-image-3.0-pro at #5 with 1263+-11 Preliminary and about 2,801 votes; the next model, gemini-3.1-flash-image (nano-banana-2), has 1262+-5. The QwenCloud model page emphasizes 4.5k token input, dense layouts, 10px small text, 12 languages, 20+ fonts, and UI simulations for websites, games, and live rooms.
Access: QwenCloud qwen-image-3.0-pro and Arena Text-to-Image.
Impact: Qwen's image model now has a much clearer comparison point. It is not yet the gpt-image-2 top tier, but its gains in text rendering, commercial design, and cartoon categories make it worth testing for Chinese ecommerce, posters, infographics, and UI mockups.
Sources: QwenCloud model page, Qwen official blog, Arena X, Alibaba_Qwen X, Text-to-Image Arena.

3. Artificial Analysis Endpoint Accuracy Index: Same Open Model, Different API Endpoints Can Lose Capability

Event type: Public benchmark update / X first signal.
First/official/availability time: @ArtificialAnlys posted the Endpoint Accuracy Index within this window, covering GLM-5.2, gpt-oss-120b, and DeepSeek V4 Pro, with Kimi K3 announced as coming soon.
Capability change: AA defines the metric as a compound of BFCL v4-500, HLE-250, and AA-LCR-25, expressed as a percentage of a reference endpoint. Low scores usually indicate that quantization, sampling defaults, output limits, reasoning effort, or other endpoint-side settings lost accuracy. The GLM-5.2 page lists 15 providers with Endpoint Accuracy ranging from Blackbox AI 52% to Fireworks/Nebius/FriendliAI 100%; gpt-oss-120b lists 20 providers with Cloudflare 70%, Google Vertex 72%, and Amazon Bedrock 101%; DeepSeek V4 Pro lists 9 providers, mostly around 97% to 107%.
Access: Artificial Analysis provider pages for each model.
Impact: This is the most useful engineering signal today. For open-weight model APIs, the model name alone is not enough. Provider, quantization format, context window, reasoning tokens, output cap, and default sampling have to be treated as one product spec.
Sources: AA GLM-5.2 providers, AA gpt-oss-120b providers, AA DeepSeek V4 Pro providers, AA methodology link, Artificial Analysis X.

4. Tencent Hunyuan Hy ASR 3.0 Preview: Speech Recognition Moves Toward Context Understanding

Event type: Official release / X first signal / access update.
First/official/availability time: Tencent Cloud's official account published it on 2026-08-04; @TencentHunyuan posted on X the same day.
Capability change: Tencent says Hy ASR 3.0 preview is based on Hy3 language understanding, combining a MoE architecture, an in-house unsupervised speech encoder, tens of millions of hours of speech data, SFT, and multi-stage reinforcement learning. It covers Mandarin, Cantonese, Wu, and other dialect groups, reports about 3% WER on open evaluation sets, including Mandarin 3.34%, English 2.62%, and Cantonese 3.12%, and supports contextual correction, hotword injection, noisy audio, and whisper scenarios.
Access: Tencent Cloud API, Yuanbao voice input first experience, and later integration into WorkBuddy and other products.
Impact: This is not a general LLM flagship, but it is practical for meeting notes, customer-service QA, voice search, and multi-dialect products. The next verification items are Tencent Cloud's specific engine ID, pricing, streaming latency, long-audio stability, and custom-vocabulary support.
Sources: Tencent Cloud official release, TencentHunyuan X, CLS brief.

5. Celeris-1: Strong AA Speed Signal, But Not Frontier Intelligence

Event type: Public benchmark update / media report / vendor-claim to verify.
First/official/availability time: Artificial Analysis lists Celeris-1 as released July 24, 2026; AA and social posts circulated its fastest-output signal within this window.
Capability change: The AA model page reports Intelligence Index 12, text/image input, text output, 131k context, output speed around 2102 tokens/s, TTFT 0.67s, input/output price $0.20/$0.70 per 1M tokens, and 7:2:1 blended price $0.25/M.
Access: Celeris provider API and Artificial Analysis.
Impact: This is a speed-specialized model, not a frontier general model. It may fit real-time UI, search reranking, and low-latency voice/customer-service pipelines, but #1 fastest should not be read as #1 overall capability.
Sources: AA Celeris-1 model page, AA Celeris-1 providers, Celeris social signal.

6. Video Leaderboards Continue: Gemini Omni Leads T2V, MiniMax H3 Sits Near the Front in I2V/T2V

Event type: Arena appearance / leaderboard change / carryover.
First/official/availability time: Arena changelog shows MiniMax H3 was added to Text-to-Video and Image-to-Video on 2026-08-03; the leaderboard date is 2026-08-02.
Capability change: In Text-to-Video Arena, gemini-omni-flash is #1 at 1513+-12, minimax-h3 is #4 at 1455+-19, and happyhorse-1.0 is #5 at 1428+-13. In Image-to-Video Arena, minimax-h3 is #2 at 1476+-19, gemini-omni-flash is #4 at 1462+-7 Preliminary, and happyhorse-1.0 is #5 at 1442+-10.
Access: Arena video leaderboards, Gemini Omni official page, and MiniMax H3 access as linked from Arena's Notion/community license.
Impact: This is not a fresh launch date for video models, but the leaderboard positions are worth recording. MiniMax H3 is strong in I2V, and the next checks are weights, license, API access, and whether the tested version matches public availability. HappyHorse1.1 appeared in Alibaba Token Plan, but Arena currently lists happyhorse-1.0; do not merge the versions.
Sources: Text-to-Video Arena, Image-to-Video Arena, Gemini Omni, Arena changelog.

Arena Watch

Text-to-Image: qwen-image-3.0-pro is #5 with 1263+-11 Preliminary and about 2,801 votes; gemini-3.1-flash-image (nano-banana-2) is #6 with 1262+-5 and about 26,659 votes. In Labs view, Alibaba is #4, while OpenAI gpt-image-2 (medium) remains #1 at 1380+-5. Sources: Text-to-Image Arena, Labs view.
Image-to-WebDev: qwen3.8-max is #2 with 1631+20/-20 Preliminary, 1,045 votes, $2/$6 pricing, and 1M context. It sits below only claude-opus-5-max and above claude-fable-5 and gpt-5.6-sol-xhigh. Source: Image-to-WebDev leaderboard.
Agent Arena: the page date is 2026-08-03, with the visible front pack still led by Claude Opus/Fable, Kimi K3, and GPT-5.6 Sol. Arena changelog added Deepseek V4 Flash (High) (20260731), but its exact row was not stable in the public first-screen table and needs follow-up. Sources: Agent Arena, Arena changelog.
Text-to-Video: gemini-omni-flash is #1 at 1513+-12 Preliminary; minimax-h3 is #4 at 1455+-19; happyhorse-1.0 is #5 at 1428+-13. Source: Text-to-Video Arena.
Image-to-Video: dreamina-seedance-2.0-720p is #1 at 1478+-10; minimax-h3 is #2 at 1476+-19; gemini-omni-flash is #4 at 1462+-7 Preliminary; happyhorse-1.0 is #5 at 1442+-10. Source: Image-to-Video Arena.
Artificial Analysis: Endpoint Accuracy Index is today's most useful new engineering dimension. GLM-5.2 providers range from 52% to 100%; gpt-oss-120b ranges from 70% to 101%; DeepSeek V4 Pro mostly stays near the reference. Sources: AA GLM-5.2, AA gpt-oss-120b, AA DeepSeek V4 Pro.

X Watch

Chrome login-state status: I connected to Chrome and successfully read a small X search over core accounts. I did not read cookies, local storage, passwords, or session files. A second larger search over official company accounts timed out in Chrome control, so this was not treated as full bulk coverage.
@arena: The visible post says Qwen-Image-3.0-Pro reached #5 in Text-to-Image Arena with 1263 pts, compared with Qwen-Image-2.0-Pro at #15 with 1191 pts; a follow-up lists category gains in Cartoon, Commercial Design, and Text Rendering. Credibility: high as an official leaderboard account, but scores will keep moving. Source: Arena X.
@Alibaba_Qwen: Posted or amplified Qwen-Image-3.0-Pro's Arena result, matching the Arena page. Credibility: high as an official source, but the capability narrative still needs real image/UI sample tests. Source: Alibaba_Qwen X.
@ArtificialAnlys: Announced the Endpoint Accuracy Index for measuring how much accuracy open-weight model endpoints preserve across serverless APIs. First coverage includes GLM-5.2, gpt-oss-120b, and DeepSeek V4 Pro, with Kimi K3 coming soon. Credibility: high as the metric owner; endpoint results are point-in-time and provider configs can change. Sources: Artificial Analysis X, AA provider pages.
@TencentHunyuan: Announced Hy ASR 3.0 preview, emphasizing a Hy3 language brain, precision ASR, and deep semantic understanding. Credibility: high as an official source; detailed specs should be grounded in Tencent Cloud and Tencent's official release. Sources: TencentHunyuan X, Tencent Cloud official release.
@xiaohu / XiaoHu AI Daily: I did not get a stable window-specific original post from Chrome-visible X search, and public web search did not surface a matching article page. Today's report does not use XiaoHu as an independent factual source and keeps it as a follow-up monitoring item.
@OpenAI, @OpenAIDevs, @AnthropicAI, @GoogleDeepMind, @GeminiApp, @xai, @grok, @AIatMeta, @MistralAI, @Kimi_Moonshot, @deepseek_ai, @runwayml, @LumaLabsAI, @Kling_ai, and @cohere: I did not find a verified new foundation-model launch in this window. Visible changes were mainly leaderboard movement, product access, ecosystem/security discussion, or non-model news.

Pending Confirmation

DeepSeek V4 Flash High: Arena changelog records it entering Agent Arena, but the public Agent table did not reliably expose the exact rank, score, or sessions in the first-screen read. Next run should locate the full row and confirm whether it is the 20260731 version.
Qwen-Image-3.0-Pro version mapping: Arena, QwenCloud, and the Qwen official blog need confirmation that they reference exactly the same model configuration. QwenCloud claims are not the same as leaderboard task capability, especially for small text, formulas, tables, and code screenshots.
Qwen-Image-3.0-Pro availability: Confirm whether international QwenCloud, DashScope, and Bailian use the same name and pricing, whether free/limited access is still active, and whether the API supports reference images, image editing, URL output, or base64 output.
Endpoint Accuracy Index: The AA metric should be part of provider selection, but each provider's default sampling, reasoning effort, max output tokens, quantization, and context truncation need verification. A low score may reflect service configuration, not only model degradation.
GLM-5.2 provider choice: Blackbox AI shows only 52% Endpoint Accuracy in AA's table but high output speed; DeepInfra is cheap but at 73%. Production integration cannot optimize only for price and speed.
gpt-oss-120b provider choice: Cerebras reaches around 1963 t/s but Endpoint Accuracy is 87%; Amazon Bedrock reaches 101% but around 101 t/s. Separate low-latency tasks from high-fidelity tasks.
Hy ASR 3.0 preview: Confirm Tencent Cloud's exact API docs, pricing, QPS, streaming latency, language coverage, hotword limits, and privacy/data-retention policy.
Celeris-1: Confirm whether vendor speed claims and AA's current benchmark use comparable loads. Its Intelligence Index is only 12, so it should not replace complex reasoning or coding models.
MiniMax H3: Arena gives strong video-leaderboard signals, but official release notes predate this window. Continue verifying H3 weights, community license, API access, and the boundary between HappyHorse1.0 and HappyHorse1.1.