AI Model Daily - 2026-07-24 | QName AI

30-Second Summary

OpenAI brought ChatGPT Voice into the desktop Chat, Work, and Codex experience: GPT-Live powers voice-based starting, checking, and steering of multiple Work/Codex tasks, iOS remote is available, Android is coming soon; Health in ChatGPT also started rolling out to logged-in U.S. users aged 18+.
Claude Voice was upgraded on the same day: it now moves beyond Haiku to Opus/Sonnet/Haiku, can reach connected tools such as Gmail, Slack, email, and calendars during voice conversations, supports more languages, and is in public beta on mobile, desktop, and web.
Qwen officially promoted Qwen-Audio-3.0-TTS: Flash and Plus variants, with Flash for real-time interaction and Plus for higher-quality generation; 16 languages, natural-language style control, fine-grained tags, and an Artificial Analysis TTS leaderboard signal showing Plus in the top position.
Grok 4.5 completed its consumer rollout across grok.com, X, iOS, and Android; Artificial Analysis records Intelligence Index 54, rank #9/186, price $2/$6, and 500K context.
Google, Runway, and NVIDIA are pushing model capability into narrower workflows: Gemini 3.5 Flash Cyber is in limited defensive-security access, Gemini Spark is rolling out to U.S. AI Pro users, and Runway Media Router chooses generative media models by cost, quality, or latency preference.

Must Watch

1. ChatGPT Voice reaches desktop Chat, Work, and Codex

Event type: Official release / availability update.
First / official / available time: OpenAI explained the rollout in X posts and ChatGPT Learn updates on 2026-07-23; desktop and iOS remote rollout started.
Capability change: ChatGPT Voice is powered by GPT-Live and can start, check, and steer ChatGPT Work and Codex tasks inside one voice conversation. OpenAI Devs also highlighted voice-based steering of Codex and Work.
Availability: ChatGPT desktop app; iOS app paired remote access; Android support coming soon.
Impact: This is not just a voice UI. It puts multi-task agent control into a low-friction surface. The workflow impact is larger than a normal chat-experience update, especially for engineering work. The next items to watch are permission boundaries, task-state visibility, error recovery, and enterprise controls.
Sources: OpenAI X, OpenAI Devs X, ChatGPT Learn - What is new.

2. Claude Voice adds Opus/Sonnet/Haiku and connected tools

Event type: Official release / availability update.
First / official / available time: Anthropic published the official blog post and @claudeai X thread on 2026-07-23; public beta is rolling out on mobile, desktop, and web.
Capability change: Voice mode no longer only runs on Haiku. It can use Opus, Sonnet, and Haiku, can reach connected tools such as Gmail, Slack, Google Calendar, Canva, and email, and asks for permission before using connected tools. It also supports more languages.
Availability: Claude mobile, desktop, and web. Free users get Haiku and one connected tool; paid users get broader model and tool coverage.
Impact: Claude and ChatGPT both upgraded voice from an input method into an agent operation layer on the same day. Claude's differentiation is connected tools plus model switching. It is worth testing on real office workflows such as longer voice conversations, calendar tasks, and email handling.
Sources: Anthropic Blog, Claude X, Claude X thread.

3. Qwen-Audio-3.0-TTS launches; TTS competition shifts toward multilingual controllability

Event type: Official release / arena appearance.
First / official / available time: Tongyi Lab published the technical article on 2026-07-20; @Alibaba_Qwen promoted it heavily on X on 2026-07-23.
Capability change: The release has two variants: Flash for real-time interaction, with official first-packet latency around 300ms, and Plus for higher-quality generation. It supports 16 languages plus natural-language style control and fine-grained tags such as [whisper], [angry], [breaths], and [laughs].
Availability: Qwen/Tongyi ecosystem and Alibaba Cloud Model Studio access still need direct verification; Artificial Analysis TTS Arena has listed Qwen-Audio-3.0-TTS-Plus.
Impact: TTS is moving from "does the voice sound similar" toward multilingual coverage, low latency, controllable expression, and comparable leaderboard signals. If the Plus leaderboard position remains stable, Qwen's usability in voice-generation pipelines will improve materially.
Sources: Qwen X, Tongyi Lab Substack, Artificial Analysis Text to Speech Arena.

4. Grok 4.5 expands from developer/build surfaces to consumer apps

Event type: Official release / availability update / leaderboard verification.
First / official / available time: xAI launched Grok 4.5 on 2026-07-16; on 2026-07-22 it announced availability across grok.com, X, iOS, and Android; @grok reconfirmed availability on 2026-07-23.
Capability change: xAI positions Grok 4.5 as a main model for coding, agentic tasks, and knowledge work. xAI also says it used about 15,954 output tokens per SWE Bench Pro task, roughly 4.2x fewer than Opus 4.8 max.
Availability: grok.com, X, iOS, Android, Grok Build, Cursor, xAI API / SpaceXAI console.
Impact: Grok 4.5's focus has shifted from launch to distribution. The items to watch now are real usage limits, API stability, Cursor/Grok Build workflow performance, and whether the $2/M input and $6/M output pricing creates a durable cost advantage.
Sources: Grok 4.5 official, Grok 4.5 everywhere, Grok X, Artificial Analysis - Grok 4.5.

5. Gemini 3.5 Flash Cyber enters limited defensive-security pilot

Event type: Official release / specialized model / restricted availability.
First / official / available time: Google DeepMind had published the official blog post, and @GoogleDeepMind added test results and access details in a 2026-07-23 X thread.
Capability change: The model is fine-tuned from Gemini 3.5 Flash for finding, validating, and patching vulnerabilities. Google says it caught complex vulnerabilities missed by standard models in internal codebases such as Chrome and Android, as well as V8 scenarios.
Availability: Limited-access pilot for governments and trusted partners, through CodeMender-related workflows.
Impact: This is a "lightweight model plus specialized security toolchain" signal, not an open general-purpose model. The key follow-ups are whether it becomes an externally purchasable security agent platform and whether Google publishes reproducible benchmarks, false-positive rates, and safety boundaries.
Sources: Google DeepMind Blog, Google DeepMind X, Google DeepMind X thread.

6. Health in ChatGPT starts U.S. rollout

Event type: Official release / product availability / high-risk domain.
First / official / available time: OpenAI published the official post on 2026-07-23 and started rolling it out to logged-in U.S. users aged 18+ on web and iOS.
Capability change: Users can connect Apple Health and supported medical records. OpenAI says GPT-5.5 Instant brought frontier health intelligence to free users, and GPT-5.6 Sol is stronger on complex health questions.
Availability: ChatGPT Free, Go, Plus, and Pro on web/iOS; OpenAI says it is not available in Codex.
Impact: This belongs in the model daily, but should be treated as a health product capability rather than a new model release. The risk areas are privacy, explainability, wrong advice, and the boundary with professional medical care.
Sources: OpenAI - Health in ChatGPT, OpenAI X.

7. Runway Media Router: generative media moves into model routing

Event type: Official release / product capability update.
First / official / available time: @runwayml announced it on 2026-07-23; developer documentation is live.
Capability change: Users define preferences such as cost, quality, or latency. Runway's router filters eligible models inside video/image/audio endpoints, selects an appropriate model, and returns both result and metadata.
Availability: Runway developer API / Media Router config.
Impact: This is not a new video model, but it matters for production generative-media workloads. Model choice is moving from manual selection toward dynamic routing by business constraints, which affects cost control, latency SLAs, and quality measurement.
Sources: Runway X, Runway Docs - Model routers.

8. Gemini Spark rolls out to U.S. AI Pro users

Event type: X first signal / product availability update.
First / official / available time: @GeminiApp said on 2026-07-23 that Gemini Spark is rolling out to U.S. Google AI Pro subscribers, after earlier availability for AI Ultra.
Capability change: Google's AI plans page positions Spark as a 24/7 personal AI agent that can navigate digital life and take action under user direction. The same page also lists agentic entries such as AI Inbox and Chrome auto browse.
Availability: Rolling out to U.S. Google AI Pro users; Google AI Ultra has had Spark in selected countries.
Impact: Spark is one of Google's core consumer AI-agent surfaces. The highest-priority confirmations are Pro rollout scope by country/account, exact permissions, actionable app coverage, and failure handling.
Sources: Gemini App X, Google AI plans.

9. Open model infrastructure: Nemotron RL workflow and The Stack v3 signal

Event type: Ecosystem/media release / X first signal / needs verification.
First / official / available time: NVIDIA promoted a Nemotron 3 Nano customization workflow on Prime Intellect Lab through X and a developer blog on 2026-07-23; Hugging Face reshared The Stack v3 dataset signal on 2026-07-23.
Capability change: NVIDIA's case says a hosted RL run improved Nemotron 3 Nano accuracy on a single math task from 22% to 91% at low cost and produced a downloadable LoRA adapter. The Stack v3 X signal claims 5T tokens ready to train and 120TB raw data.
Availability: Prime Intellect Lab / NVIDIA developer workflow; The Stack v3 public dataset card and governance details still need verification.
Impact: This matters for open-model post-training and code-model data supply, but it is not a new base model launch today. The Stack v3 should stay in the confirmation queue until there is a dataset card, license, benchmark, or complete technical page.
Sources: NVIDIA AI X, NVIDIA Developer Blog, Hugging Face X.

Arena Watch

LMArena / Arena Code WebDev Overall: the leaderboard timestamp shows 2026-07-21, with 506,528 votes and 101 models. Top signals: Kimi K3 #1, Elo 1678; Claude Fable 5 #2; GPT 5.6 Sol xHigh #3; Grok 4.5 #7, Elo 1557, price $2/$6, 500K context; Muse Spark 1.1 #11; Gemini 3.6 Flash #12; Hy3 #16, 262.1K context.
LMArena / Arena Text Overall: the leaderboard timestamp shows 2026-07-21, with 7,430,560 votes and 378 models. Claude Fable 5 is #1, Muse Spark 1.1 is #5, and Kimi K3 is #10. No new @arena X post was visible in this window, so this is recorded as a leaderboard snapshot rather than a last-24-hour change.
LMArena / Agent Arena: the leaderboard timestamp shows 2026-07-21, with 1,242,857 sessions and 38 models. Claude Fable 5 (High) is #1, GPT 5.6 Sol (xHigh) is #2, Claude Opus 4.8 Thinking is #3, Kimi K3 is #4, and GLM 5.2 Max is #10.
Artificial Analysis / Grok 4.5: AA's model page lists Grok 4.5 as released in July 2026, Intelligence #9/186, Index 54, speed 68.8 output tokens/s, price $2/M input and $6/M output, text/image input, text output, and 500K context.
Artificial Analysis / TTS: Qwen-Audio-3.0-TTS-Plus appears at the top of the AA Text to Speech Arena, with Elo around 1234. This gives Qwen's higher-quality TTS tier an independent comparison signal, but sample size and leaderboard volatility still need monitoring.
Artificial Analysis X: @ArtificialAnlys emphasized cost per task and token efficiency on 2026-07-23, saying OpenAI still occupies much of the token-efficiency Pareto frontier in its Intelligence Index. No new general-model Index reshuffle was visible today, but cost efficiency is becoming a main model-comparison axis.

Sources: Arena Code WebDev, Arena Text, Arena Agent, Artificial Analysis - Grok 4.5, Artificial Analysis TTS Arena, Artificial Analysis X.

X Watch

@xiaohu: multiple high-circulation Chinese tracking signals appeared on 2026-07-23. Key items included Chinese summaries of OpenAI Presence/Health/Voice, an unverified rumor about Codex quota reset and "Code taking over Siri/computer", and sponsored OpenSquilla 0.5.0 Preview content. For model releases, this is not an official source, but it is useful as an early-signal and Chinese-distribution source. Sources: Codex rumor, OpenSquilla, OpenAI Presence summary.
@OpenAI / @OpenAIDevs: ChatGPT Voice, Health in ChatGPT, Codex remote voice control, multi-folder Codex projects, and Sites Analytics public testing were the densest official product updates today. The model layer is not a newly numbered base model, but GPT-Live and GPT-5.6 Sol appear in important product descriptions. Sources: OpenAI Voice, OpenAI Health, OpenAI Devs Voice, Codex multi-folder.
@claudeai: Voice mode upgrade is the direct counterpart to OpenAI's voice update today. The core points are Opus/Sonnet/Haiku, connected tools, more languages, and public beta rollout across platforms. Source: Claude Voice.
@Alibaba_Qwen: Qwen-Audio-3.0-TTS is today's clearest new model / speech-model official signal. Flash/Plus segmentation, tag control, multilingual support, and the AA leaderboard position deserve continued tracking. Source: Qwen-Audio-3.0-TTS.
@grok / xAI: Grok 4.5 consumer availability was confirmed. Direct access to @xai in Chrome also resolved to @SpaceXAI-related content, so future collection should continue verifying account redirects through the logged-in Chrome session. Sources: Grok X, Grok 4.5 official.
@GoogleDeepMind / @GeminiApp: Flash Cyber and Spark respectively represent a specialized security model and consumer-agent rollout. Flash Cyber is the official deep technical line; Spark is currently mainly an X rollout signal and should be cross-checked against the Google AI plans page. Sources: Flash Cyber X, Gemini Spark X.
@runwayml: Media Router is a generative-media engineering signal, not a new model. It is directly relevant to cost and latency optimization for video, image, and audio generation. Source: Runway Media Router.
@NVIDIAAI / @HuggingFace: NVIDIA promoted a Prime Intellect Lab RL customization workflow for Nemotron 3 Nano; Hugging Face reshared The Stack v3. Both are closer to open-model infrastructure and should wait for dataset cards, licenses, benchmarks, or complete technical pages before their credibility is upgraded. Sources: NVIDIA AI, Hugging Face.
@AIatMeta, @MistralAI, @cohere, @Kimi_Moonshot, @LumaLabsAI, @Kling_ai, @Alibaba_Wan, @SakanaAILabs, @deepseek_ai, @karpathy: no sufficiently high-signal new model release or leaderboard change was visible in this window; Luma/Kling posts were mainly use-case videos and are not listed as must-watch items.

Confirmation Queue

OpenAI Voice / GPT-Live: confirm whether GPT-Live has a developer API, what enterprise permission controls exist, how voice-steered Codex/Work recovers from failed tasks, and the exact Android rollout date.
Health in ChatGPT: keep tracking medical-record integration scope, privacy settings, professional medical disclaimers, answer auditing, and whether the feature expands outside the U.S.
Claude Voice: test Free/paid tool counts, Opus/Sonnet availability limits, tool permission prompts, non-English stability, and real success rates after connecting Gmail/Slack/Calendar.
Qwen-Audio-3.0-TTS: verify Model Studio/API availability, pricing, commercial terms, measured Flash first-packet latency, Plus leaderboard sample size, and consistency of emotion/breath/laughter tags across languages.
Grok 4.5: confirm region and subscription limits across grok.com/X/iOS/Android, API quotas, the end date of the Cursor/Grok Build free period, SWE Bench Pro reproducibility details, and the @xai redirect/account-collection ambiguity.
Gemini 3.5 Flash Cyber: needs a model card, reproducible benchmarks, false-positive/false-negative data, pilot application criteria, CodeMender API/platform boundaries, and whether access expands beyond governments and trusted partners.
Gemini Spark: @GeminiApp says U.S. AI Pro rollout, while the Google AI plans page still emphasizes Ultra/selected countries in places; account-level testing and country coverage confirmation are required.
Runway Media Router: confirm the eligible model pool, router config options, metadata detail, cost/latency transparency, and whether custom business scoring is supported.
The Stack v3: current signal is mainly from X; wait for a Hugging Face dataset card, license, governance note, dedup/PII processing details, and code-training split information.
OpenSquilla 0.5.0 Preview: @xiaohu clearly marked this as sponsored/paid partnership. The claimed Chinese-model team collaboration and cost advantage need official docs, an open-source repository, benchmarks, and independent reproduction before credibility can be upgraded.