AI Model Daily - 2026-07-25

30-Second Summary

• Anthropic officially released Claude Opus 5: it is available the same day on paid Claude plans and the Claude API, positioned as near Fable 5 frontier intelligence at roughly half the price. Fast mode is about 2.5x faster than default mode.
• Artificial Analysis put Claude Opus 5 at the center of the day's benchmark coverage: Opus 5 narrowly leads its Intelligence Index and becomes the new AA-Briefcase leader, nearly 150 Elo ahead of Claude Fable 5 with about 20% lower Cost per Task.
• LMArena / Arena has made Claude Opus 5 available for testing in Agent Mode and Battle Mode, but public leaderboard rankings still need to stabilize. Arena also emphasized that Kimi K3 has passed Fable 5 in Code Arena and ranks first in 6 of 7 frontend domains.
• Sakana AI released Fugu-Ultra v1.1, using dynamic orchestration of the latest frontier models to improve complex coding/reasoning performance. This is an agent/orchestration model-system path, not a single base-model release.
• Grok, Runway, Gemini, and NVIDIA were mostly about availability, workflows, and infrastructure: Grok entered Google Workspace and surfaced STT through OpenRouter; Runway Agent added Workflows; Gemini Spark expanded to U.S. AI Pro users; NVIDIA ModelExpress reduced DeepSeek-V4 Pro startup time.

Must Watch

1. Claude Opus 5 officially launches

• Event type: Official release / availability update.
• First / official / available time: Anthropic's official blog and @claudeai launched it on 2026-07-24; Anthropic says it is available the same day.
• Capability change: Anthropic positions it as a new Opus version that is close to Fable 5 frontier intelligence at about half the price. It is stronger than Opus 4.8 on cybersecurity tasks, while Anthropic also says its exploit-development capability remains below Mythos 5. Fast mode is about 2.5x faster than default mode.
• Availability: Claude paid plans and Claude API. Claude Max uses Opus 5 by default; Claude Pro users get access to its strongest model tier.
• Impact: This is today's only clear top-tier official model launch. Short term, watch whether the Opus line can regain high-end reasoning and agent use cases through lower price and Fast mode. Mid term, the key issues are API limits, tool-use reliability, cost curves, and enterprise security policy.
• Sources: Anthropic - Claude Opus 5, Claude X launch, Claude X availability.

2. Claude Opus 5 becomes a key Artificial Analysis benchmark signal

• Event type: Arena / benchmark verification.
• First / official / available time: @ArtificialAnlys published multiple evaluation updates on 2026-07-24.
• Capability change: Artificial Analysis says Claude Opus 5 narrowly becomes the most intelligent model on its Intelligence Index, with intelligence close to Fable 5 but about 26% lower Cost per Task. It also becomes the new AA-Briefcase leader, nearly 150 Elo ahead of Claude Fable 5 with about 20% lower Cost per Task.
• Availability: Artificial Analysis model page and AA-Briefcase leaderboard.
• Impact: Opus 5's core pitch is not absolute domination. It is high-end capability plus lower task cost. However, AA also reports that high-effort settings can average 25.7-36.2 minutes per AA-Briefcase task, so production use should choose effort level by task value rather than maxing it by default.
• Sources: Artificial Analysis - Claude Opus 5, AA X Intelligence Index, AA X AA-Briefcase, AA-Briefcase.

3. Claude Opus 5 lands in Arena Agent Mode / Battle Mode

• Event type: Arena appearance / testable surface update.
• First / official / available time: @arena announced on 2026-07-24 that Claude Opus 5 had entered Arena.
• Capability change: Arena says Claude Opus 5 can now be tested in Agent Mode and Battle Mode, with static benchmark intelligence near Fable 5.
• Availability: Arena Agent Mode and Battle Mode; public leaderboard rankings still need stable samples.
• Impact: Arena availability means real battle signals can start accumulating, but today's update should not be read as a finalized leaderboard rank. The next thing to watch is whether Agent Arena, Text Arena, or Code/WebDev Arena shows a stable ranking shift.
• Sources: Arena X - Opus 5 landed, Arena Agent, Arena Text, Arena Code WebDev.

4. Sakana AI releases Fugu-Ultra v1.1

• Event type: Official release / agent orchestration / evaluation pending.
• First / official / available time: @SakanaAILabs announced Fugu-Ultra v1.1 on 2026-07-24 and shared release notes.
• Capability change: Fugu-Ultra v1.1 dynamically orchestrates the latest frontier models. Signals from the Sakana team say performance improved by 7.9 points and surpassed Fable 5 on complex coding/reasoning, while Fable 5 was not included in the agent pool.
• Availability: Sakana Fugu release notes and Fugu product page. External API access, pricing, and usage restrictions still need verification.
• Impact: This is important, but the category matters. Fugu-Ultra v1.1 is closer to a multi-model orchestration/agent model system than to a single base model. It reinforces the trend that high-end tasks may increasingly be solved by a combination of model pools, routing, tools, and long-horizon execution.
• Sources: Sakana Fugu-Ultra v1.1, Sakana Fugu, Sakana AI X, Hardmaru X.

5. Grok enters Google Workspace and brings STT to OpenRouter

• Event type: Official release / availability update / API distribution.
• First / official / available time: @grok, @SpaceXAI, and OpenRouter-related X signals appeared on 2026-07-24.
• Capability change: Grok launched a Google Workspace add-on that can turn research into reports in Sheets, create decks from outlines in Slides, and draft content inside Docs. Grok STT is also live through OpenRouter with 25 languages, word-level timestamps, speaker diarization, and OpenRouter pricing at $0.10/audio hour.
• Availability: x.ai/grok/workspace, Google Workspace add-on, OpenRouter, and xAI's STT/TTS API page.
• Impact: Grok is moving from chat/model release into office surfaces and speech API distribution. The short-term items to test are Workspace add-on permissions, enterprise-domain install restrictions, real STT language quality, and diarization accuracy.
• Sources: Grok Workspace X, Grok Workspace, Grok STT on OpenRouter X, xAI STT/TTS APIs.

6. Runway Agent adds Workflows

• Event type: Official release / product capability update.
• First / official / available time: @runwayml announced it on 2026-07-24.
• Capability change: Runway Agent now includes Workflows. Users can create, run, and edit node-based workflows through natural language; the entry point is the / Workflow skill.
• Availability: Runway Agent workflow surface; plan coverage and rollout scope still need verification.
• Impact: This is not a new video model, but it matters for generative media operations. Runway is moving from one-off generation toward editable multi-node generation processes. The next thing to watch is whether it can reliably manage assets, branches, reuse, and batch processing.
• Sources: Runway Workflows X.

7. NVIDIA ModelExpress shortens DeepSeek-V4 Pro startup time

• Event type: Official release / inference infrastructure.
• First / official / available time: @NVIDIAAI posted the update on 2026-07-24 and pointed to an NVIDIA developer blog post.
• Capability change: NVIDIA says ModelExpress in Dynamo uses a GPU-to-GPU RDMA path to send model artifacts directly into GPU memory, reducing DeepSeek-V4 Pro startup from about 8 minutes to under 2 minutes.
• Availability: NVIDIA Dynamo / ModelExpress stack.
• Impact: This is not a model launch, but it matters for large-model serving cost. Multi-model routing, cold starts, and elastic scaling will increasingly depend on this kind of artifact-distribution infrastructure. Watch supported models, deployment complexity, and real cloud-environment gains.
• Sources: NVIDIA AI X, NVIDIA Developer Blog - ModelExpress.

8. Gemini Spark expands to U.S. Google AI Pro users

• Event type: X-first signal / product availability update.
• First / official / available time: @GeminiApp reposted a Google product-lead signal on 2026-07-24 saying Spark is live for U.S. Google AI Pro subscribers and will expand globally next.
• Capability change: The example shows Spark reading a school calendar PDF and adding no-school days to Google Calendar. Google's AI plans page positions Spark as a personal AI agent that can take action across a user's digital life.
• Availability: U.S. Google AI Pro users; Google AI Ultra and additional-country coverage need account-level verification.
• Impact: Spark is a key consumer-agent entry point for Google. Today's new signal is Pro-user coverage, not a new model launch. The real issues are permission confirmation, depth across Google Workspace/Calendar, failure rollback, and cross-region availability.
• Sources: Josh Woodward X, Gemini App X, Google AI plans.

9. Open-weight model ecosystem becomes a policy and platform theme

• Event type: Official statement / ecosystem signal / non-model release.
• First / official / available time: Microsoft, NVIDIA, Hugging Face, Arena, Sam Altman, and others shared or responded to open-weight model ecosystem statements on 2026-07-24.
• Capability change: The statement emphasizes the importance of open-weight models for U.S. AI leadership, innovation ecosystems, and infrastructure. It does not directly release a new model.
• Availability: Microsoft open-weight statement.
• Impact: This should not be categorized as a model launch, but it belongs in the daily. It may influence future model openness, platform default support, government procurement, and enterprise usage boundaries. Watch whether OpenAI, Meta, Microsoft, or NVIDIA follows with actual model, license, hosting, or procurement actions.
• Sources: Microsoft open-weight statement, Arena X, Sam Altman X, Hugging Face X.

Arena Watch

• LMArena / Arena: Claude Opus 5 is now in Arena Agent Mode and Battle Mode, so testing can begin, but today's report does not record it as a stable public leaderboard rank. Sources: Arena X, Arena Agent.
• LMArena / Code Arena: @arena says Kimi K3 has passed Fable 5 in Code Arena and ranks first in 6 of 7 frontend domains. The public Code WebDev leaderboard still needs follow-up snapshots, so the X launch signal and stable leaderboard state should not be conflated. Sources: Arena X - Kimi K3, Arena Code WebDev.
• Artificial Analysis / Intelligence Index: AA says Claude Opus 5 narrowly becomes the most intelligent model, close to Fable 5 while reducing Cost per Task by about 26%. Sources: AA X Intelligence Index, Claude Opus 5 model page.
• Artificial Analysis / AA-Briefcase: Claude Opus 5 becomes the new AA-Briefcase leader, nearly 150 Elo ahead of Claude Fable 5, with about 20% lower Cost per Task. Sources: AA X AA-Briefcase, AA-Briefcase.
• Artificial Analysis / effort tradeoff: AA reports that high-effort Opus 5 settings average about 36.2, 34.3, and 25.7 minutes per AA-Briefcase task. On GDPval-AA v2, different effort settings span 407 Elo, and output-token usage differs by about 8x. This means Opus 5's practical value depends on matching effort level to task complexity and budget. Sources: AA X task time, AA X effort spread.
• Conclusion: Today's leaderboard theme is Claude Opus 5 combining high-end capability with cost efficiency, plus Kimi K3 remaining strong in Code Arena. Stable Arena public rankings need the next leaderboard snapshot.

X Watch

• @xiaohu: High-value Chinese-language leads included an interpretation of Jensen Huang's first X post supporting open/open-weight model ecosystems, plus Hyper3D/Rodin BANG to Parts, a 3D model decomposition capability. The latter decomposes a complete 3D model into independent usable parts for material editing, animation, 3D printing, and asset reuse. For now, treat it as an X lead until Hyper3D publishes an official page or technical documentation. Sources: xiaohu open models, xiaohu BANG to Parts, xiaohu BANG explanation.
• @arena: Two important updates: Claude Opus 5 entered Arena Agent Mode / Battle Mode, and Kimi K3 passed Fable 5 in Code Arena while leading most frontend domains. Sources: Opus 5 in Arena, Kimi K3 Code Arena.
• @ArtificialAnlys: The densest third-party benchmark source today. The core signal is that Opus 5 is strong across the Intelligence Index, AA-Briefcase, and GDPval-AA v2, while effort/cost/time differences are large. Sources: AA Intelligence, AA-Briefcase, AA task time.
• @claudeai / @AnthropicAI: The official Claude Opus 5 launch thread gave capability, price, safety, and availability details. Sources: Claude launch, Claude safety, Claude availability.
• @GeminiApp: Reposted Josh Woodward saying Gemini Spark is live for U.S. Google AI Pro users and will expand globally. This is a product-lead/X rollout signal and still needs account-level availability verification. Source: Josh Woodward Spark.
• @OpenAIDevs: Reposted two developer-workflow signals: one showing ChatGPT Voice with visual UI and tool calls, and another about a Codex Multi-Agent V2 article describing Sol/Terra delegation, shared updates, and coordination. No new official OpenAI model launch was seen today. Sources: Voice workflow, Codex multi-agent article.
• @grok / @SpaceXAI: Grok Workspace add-on, Grok STT on OpenRouter, and a Grok 4.5 invoice-processing benchmark claim. The invoice claim comes through a Ramp test signal and needs public methodology before it should be treated as a general benchmark. Sources: Grok Workspace, Grok invoice benchmark, Grok STT OpenRouter.
• @runwayml: Runway Agent Workflows launched, emphasizing natural-language creation, execution, and editing of node-based workflows. Source: Runway Workflows.
• @NVIDIAAI: ModelExpress / Dynamo inference-infrastructure signal, focused on DeepSeek-V4 Pro cold-start optimization. Other updates included Nemotron Reasoning Challenge and Kaggle agent-competition ecosystem items. Source: ModelExpress.
• @SakanaAILabs: Fugu-Ultra v1.1 is the most important model-system update today after Opus 5. Sources: Fugu announcement, Fugu release notes.
• @HuggingFace: Main signals were the open-weight model ecosystem statement, Google's gmn/gnm 3D head model on Hugging Face Spaces, and AMD MI455X hardware. The Google 3D head model needs verification of naming, license, and official model card. Source: HuggingApps gmn/gnm.
• @GoogleDeepMind, @GoogleAI, @OpenAI, @Alibaba_Qwen, @Kimi_Moonshot, @AIatMeta, @MistralAI, @cohere, @LumaLabsAI, @Kling_ai, @Alibaba_Wan, @TencentHunyuan, @deepseek_ai, @karpathy: No high-confidence new official model release was visible in this window. Luma and Kling were mostly use-case demonstrations and were not promoted into Must Watch.
• XiaoHu AI Daily site: Chrome opened the xiaohu.ai homepage and could see the XiaoHu.AI Daily entry point and site structure, but the visible page did not expose the full current-window daily article content. This run uses the logged-in @xiaohu X timeline as the XiaoHu signal source.

To Confirm

• Claude Opus 5: Verify API model id, context length, tool-use limits, Fast mode coverage, batch/cache pricing, and real agent-task differences versus Fable 5, Opus 4.8, and Mythos 5.
• Claude Opus 5 leaderboards: Wait for stable Arena public leaderboard snapshots. For AA's Intelligence Index, AA-Briefcase, and GDPval-AA v2, continue watching sample updates, Cost per Task methodology, and default effort settings.
• Fugu-Ultra v1.1: Confirm the agent pool, list of frontier models used, benchmark set, external API availability, pricing, and license. Do not treat "surpassed Fable 5" as an independent third-party conclusion yet.
• Grok Workspace: Confirm Google Workspace add-on permission scope, enterprise-domain install restrictions, data retention policy, whether Drive/Calendar/Gmail are supported, and how it routes between Grok 4.5 and other models.
• Grok STT on OpenRouter: Verify OpenRouter model name, live billing, the 25-language list, speaker-diarization quality, word-level timestamp accuracy, and whether it is equivalent to xAI's official STT API.
• Grok 4.5 invoice benchmark: Ramp's 150K-bill test needs public methodology, sample categories, field definitions, error metrics, and comparison-model settings. Treat it as a high-visibility X benchmark claim for now.
• Gemini Spark: Use real accounts to verify U.S. Google AI Pro coverage, global expansion timing, Calendar/PDF permission prompts, failure rollback, and action boundaries.
• Runway Workflows: Confirm rollout scope, whether it is available to all Runway Agent users, workflow node types, export/reuse support, and API triggering.
• NVIDIA ModelExpress: Verify supported models beyond DeepSeek-V4 Pro, Dynamo version requirements, RDMA/network dependencies, cloud-provider availability, and cost impact.
• Hyper3D/Rodin BANG to Parts: Current evidence mainly comes from @xiaohu X leads. Wait for Hyper3D official technical pages, demos, model cards, or papers before raising confidence.
• Google gmn/gnm 3D head model: The Hugging Face/X signal needs verification of official name, Space URL, license, training data, and whether it is an official Google release.
• Open-weight model statement: Today's item is an ecosystem/policy signal, not a model release. Raise priority if it leads to actual models, licenses, hosting, or government-procurement actions.