AI Model Daily - 2026-06-26 | QName AI

30-Second Summary

The clearest official model-capability update today comes from Google: Gemini 3.5 Flash now exposes Computer Use as a built-in public preview tool. It is no longer only a separate Gemini 2.5 computer-use preview route; it is part of the main Flash model and can be used through the Gemini API / Gemini Enterprise Agent Platform to build browser, mobile, and desktop action agents.
Artificial Analysis added strong multimodal leaderboard signals today: HappyHorse-1.0 reached #1 on the new AI Video Editing Leaderboard; MAI-Image-2.5 is #2 in Text-to-Image and #3 in Image Editing in the Image Arena, behind only OpenAI image models.
OpenAI did not launch a new base model today, but Codex and agent workflows moved materially: OpenAI published "How agents are transforming work", while @OpenAIDevs announced Codex mobile app GA, one-to-one remote-control pairing, the DigitalOcean plugin, and related workflow updates.
Runway released Agent 2.0, moving beyond "generate a video" toward an integrated creative agent for marketing briefs, ad assets, campaign assets, and performance-data analysis. This is a product/workflow upgrade, not a new video model ID.
Midjourney's official X account exposed an early --preview signal for V8.2 aesthetics and personalization, plus large-batch draft mode with --sref random; treat this as an X-first preview signal until a formal V8.2 model card or release note appears.

Must Watch

1. Gemini 3.5 Flash: Computer Use enters the main model as public preview

Event type: official release + model capability update.
First/official/availability time: Google's official blog is dated 2026-06-24; @GoogleDeepMind posted the X signal again at 2026-06-25T16:21Z; Gemini API docs mark the capability as preview.
Model/function name: Gemini 3.5 Flash, Computer Use tool, gemini-3.5-flash.
Capability change: Google made Computer Use a built-in tool for Gemini 3.5 Flash. Developers can build browser, mobile, and desktop control agents from screenshots and UI actions. The docs emphasize multi-environment support, intent actions, configurable safety policies, and prompt-injection detection.
Availability: Gemini API and Gemini Enterprise Agent Platform. Gemini API examples use model: "gemini-3.5-flash" with tools: [{ type: "computer_use", environment: "browser" }].
Impact: this is the official capability update to test first today. It moves computer-use from a specialized-model path back into the main Flash model, directly affecting web automation, enterprise back-office operations, long-horizon agents, RPA replacement, and safety-policy design. It is still preview, so production use needs separate checks around action confirmation, sensitive-action blocking, and injection defenses.
Sources: Google Blog: Introducing computer use in Gemini 3.5 Flash, Gemini API changelog, Gemini API Computer Use docs, Gemini 3.5 Flash model card, @GoogleDeepMind

2. HappyHorse-1.0: #1 on Artificial Analysis' new Video Editing board

Event type: leaderboard appearance + ranking change.
First/official/availability time: @ArtificialAnlys posted the new AI Video Editing Leaderboard signal at 2026-06-25T18:13Z; Artificial Analysis' methodology states that video editing / video editing with audio is part of its video evaluation suite, with quality Elo continuously recomputed from user votes.
Model/function name: HappyHorse-1.0, Seedance 2.0, Wan 2.7, AI Video Editing Leaderboard.
Capability change: Artificial Analysis says HappyHorse-1.0 ranks #1 in AI video editing, followed by Seedance 2.0 and Wan 2.7. Its thread also highlights complex-editing capability and sound/speech editing as the most important separators, with HappyHorse-1.0 and Wan 2.7 clearly ahead on sound and speech editing.
Availability: Artificial Analysis leaderboard; model API/product availability still needs confirmation from each model provider or aggregator. Runway API previously listed happyhorse_1_0 as a callable video-generation model, but "#1 on a video editing leaderboard" does not automatically mean a full official editing API is available.
Impact: video competition is shifting from T2V/I2V toward editability. For ads, short-drama remixes, product demos, localized revisions, and voice/sound-synchronized edits, video-editing Elo is closer to production reality than pure generation Elo.
Sources: @ArtificialAnlys HappyHorse Video Editing, Artificial Analysis video methodology, Runway API changelog

3. MAI-Image-2.5: Microsoft image model stays near the front of AA Image Arena

Event type: leaderboard appearance + ranking change.
First/official/availability time: @ArtificialAnlys posted at 2026-06-25T16:11Z that Microsoft's MAI-Image-2.5 ranks #2 in Text-to-Image and #3 in Image Editing on AA Image Arena; the leaderboard pages list it among models added in the last month.
Model/function name: MAI-Image-2.5, MAI-Image-2.5-Flash.
Capability change: AA Text to Image places MAI-Image-2.5 near the top, behind GPT Image 2 (high). AA Image Editing places it behind GPT Image 1.5 high and GPT Image 2 high, ahead of or very close to the Nano Banana family.
Availability: Artificial Analysis Image Arena; Microsoft's or platform providers' official API, pricing, region, and commercial terms still need follow-up.
Impact: this is a clear signal that Microsoft's image model is competing in the first tier. The useful near-term tests are editing consistency, text rendering, branded visuals, posters/product shots, localized edits, and batch-generation cost, not only subjective single-image quality.
Sources: @ArtificialAnlys MAI-Image-2.5, AA Text to Image Leaderboard, AA Image Editing Leaderboard

4. OpenAI Codex / agents: no new model, but the agent product chain keeps accelerating

Event type: official release + product/research update.
First/official/availability time: OpenAI's "How agents are transforming work" article is dated 2026-06-25; @OpenAI posted it at 2026-06-25T17:23Z. @OpenAIDevs posted at 2026-06-25T21:15Z that Codex in the ChatGPT mobile app is generally available, alongside remote-control pairing, notifications, goals, side chat, file previews, and inline review comments.
Model/function name: Codex, Codex mobile, Record & Replay, DigitalOcean plugin, OpenAI agents.
Capability change: OpenAI's research article frames agentic AI as a shift from single interactions to delegated long-horizon tasks. The Codex changelog is practical: mobile GA, one-to-one device pairing, better long-thread handling, more reliable connections, improved diff handling, the DigitalOcean plugin, and remote workspaces.
Availability: Codex in the ChatGPT mobile app, Codex App, OpenAI Codex docs/changelog, and the DigitalOcean plugin. Record & Replay is a macOS Codex feature described by official docs as demonstrating a workflow once and turning it into a reusable skill, with availability limits by region and Computer Use access.
Impact: OpenAI's signal today is not a stronger model ID; it is agent-workbench maturity. For developers, mobile handoff, remote environments, workflow recording, long-thread robustness, and diff reliability can matter more in daily production than a small benchmark move.
Sources: OpenAI: How agents are transforming work, Codex changelog, Codex Record and Replay docs, @OpenAI agents, @OpenAIDevs Codex mobile, @OpenAIDevs DigitalOcean plugin

5. Runway Agent 2.0: marketing creative-agent upgrade

Event type: official release + product/agent workflow update.
First/official/availability time: Runway's official news page shows Agent 2.0 as released; @runwayml posted at 2026-06-25T18:40Z.
Model/function name: Runway Agent 2.0.
Capability change: Runway positions Agent 2.0 as a creative agent for marketing teams: it can take a prompt, campaign brief, product launch, or audience input and create marketing briefs, ad assets, videos, and campaign assets, while also analyzing performance data and scaling across platforms, formats, and markets.
Availability: Runway product Agent; not a new API model ID. On the Runway API side, the recent model entries still worth separate tracking are seedance2_fast, aleph2, happyhorse_1_0, and similar model endpoints.
Impact: Runway is moving from "model tool" toward "creative production system". For AI video SaaS products, the lesson is that brief handling, asset management, localization, campaign feedback, and batch versioning may become the paid layer above raw model capability.
Sources: Runway: Introducing Agent 2.0, @runwayml Agent 2.0, Runway API changelog

6. Midjourney V8.2 preview: X-first preview, not a formal model launch

Event type: X-first signal + preview pending verification.
First/official/availability time: @midjourney posted at 2026-06-25T19:11Z that --preview can expose an early look at V8.2 aesthetics and personalization, combined with large-batch draft mode and --sref random for 24x style exploration.
Model/function name: V8.2 preview, --preview, draft mode, --sref random.
Capability change: the official X language reads like a preview channel, not a default-model switch. Current Midjourney docs still list V8.1 as the default version, and the draft-mode docs describe V8.1 draft mode as generating 24 lower-resolution drafts that can later be regenerated at higher resolution with Vary/Remix.
Availability: Midjourney web/Discord parameters; confirm against the official docs and actual product UI.
Impact: this is a useful exploration signal for visual creators today, but it should not be written as a formal V8.2 launch. Wait for the model-version page, announcement, or release notes to confirm V8.2 default status, availability, GPU cost, and parameter compatibility.
Sources: @midjourney V8.2 preview, Midjourney Version docs, Midjourney Draft Mode docs

Arena Watch

LMArena / Arena Agent: the overall top tier remains Claude and GPT-5.5 heavy. Public leaderboards show Claude Fable 5 High near the top, with Claude Opus 4.8 Thinking, GPT 5.5 xHigh, Claude Opus 4.7 Thinking, and GPT 5.5 High forming the first group; GLM 5.2 Max is especially strong in Confirmed Success and Praise vs Complaint submetrics.
LMArena WebDev: Claude Fable 5 remains WebDev #1 at about 1654 Elo; GLM-5.2 (max) is about 1593 Elo at #2; Claude Opus 4.8 Thinking is #3. @arena posted GLM's Code Arena: Frontend trajectory today, from GLM-4.6 at 1408 to GLM-5.2 (Max) at 1595, saying it has passed Opus 4.8 and is closing in on Claude Fable 5.
LMArena Text: Claude Fable 5 remains near the top, with the public board around 1508+/-9; Claude Opus, GPT-5.5, Gemini, and GLM models continue to cluster near the front. No new text-model debut appeared today, but Fable 5 availability should stay on the watchlist.
LMArena Text-to-Video: gemini-omni-flash remains #1, publicly shown around 1527+/-14; Dreamina Seedance 2.0 720p is #2; HappyHorse-1.0 #3; Veo 3.1 audio 1080p, Wan2.7-t2v, and Sora-2-pro follow. gemini-omni-flash still lacks a corresponding official Gemini model card, so treat it as a leaderboard appearance.
LMArena Image-to-Video: Dreamina Seedance 2.0 720p and gemini-omni-flash remain near the front; grok-imagine-video-1.5-preview, HappyHorse-1.0, and Wan2.7-i2v are in the chasing group. @arena today reposted the Wan-2.7 I2V #5 signal at about 1434.
LMArena Video Edit: the public page still shows Dreamina Seedance 2.0 720p #1, HappyHorse-1.0 #2, and grok-imagine-video #3. This differs from Artificial Analysis' "HappyHorse-1.0 #1 video editing" signal, likely because the task design, voting pool, and audio dimensions differ. Do not collapse the two boards into one ranking.
Artificial Analysis Text to Image: GPT Image 2 (high) remains #1, with MAI-Image-2.5 near the front alongside HiDream-O1-Image-1.5, GPT Image 1.5 high, and Nano Banana 2. Today's important signal is MAI-Image-2.5 holding the front group, not a new image-model launch.
Artificial Analysis Image Editing: GPT Image 1.5 high and GPT Image 2 high are the top two, MAI-Image-2.5 is #3, and Nano Banana 2 / Nano Banana Pro follow closely. For product images and branded visuals, the editing board matters more than the T2I board.
Artificial Analysis Video: T2V with audio is led by Dreamina Seedance 2.0 720p, HappyHorse-1.1, HappyHorse-1.0, Kling 3.0 1080p Pro, and SkyReels V4; T2V without audio is led by HappyHorse-1.0, HappyHorse-1.1, Dreamina Seedance 2.0, Kling 3.0 1080p Pro, and Kling 3.0 Omni 1080p Pro. I2V with audio is led by Dreamina Seedance 2.0 720p, HappyHorse-1.1, grok-imagine-video-1.5-preview, Wan 2.7, and HappyHorse-1.0.
Artificial Analysis LLM / Intelligence: the front group remains Claude Fable 5 (with fallback), Claude Opus 4.8 (max), GPT-5.5 (xhigh), Claude Opus 4.7 (max), and similar models. There was no new language-model evaluation debut today; use this board mainly as context.
Price Per Token: the page shows no new model-release entries in the last 24 hours; the past week includes Sakana Fugu Ultra, Nex N2 Pro, Grok 4.3 via Amazon Bedrock, and other callable entries. Today it adds less value for new first releases than X, official sources, and leaderboards.
Sources: Arena overview, Arena WebDev, Arena Text-to-Video, Arena Image-to-Video, Arena Video Edit, @arena GLM-5.2, @arena Wan-2.7 I2V, Artificial Analysis, AA Text to Image, AA Image Editing, AA Text to Video, AA Image to Video, Price Per Token

X Watch

@GoogleDeepMind: posted at 2026-06-25T16:21Z that Gemini 3.5 Flash now supports native computer use. Confidence is high, cross-checked against the Google blog, Gemini API changelog, and model card. Classification: official release.
@ArtificialAnlys: posted MAI-Image-2.5 Image Arena rankings at 2026-06-25T16:11Z; posted HappyHorse-1.0 as #1 on AI Video Editing Leaderboard at 2026-06-25T18:13Z. Confidence is high, but rankings can move with votes. Classification: leaderboard appearance.
@arena: posted the GLM Code Arena: Frontend trajectory at 2026-06-25T15:56Z, emphasizing GLM-5.2 (Max) around 1595 and closing in on Claude Fable 5; also reposted Wan-2.7 I2V #5. Classification: leaderboard appearance.
@OpenAI: posted the agents research article at 2026-06-25T17:23Z. This is an official research/product direction signal, not a new model. Classification: official release.
@OpenAIDevs: posted Codex mobile GA, remote-control pairing, notifications, goals, side chat, file preview, inline review comments, and related fixes at 2026-06-25T21:15Z; posted the DigitalOcean plugin at 2026-06-25T21:43Z. Classification: official release.
@runwayml: posted Agent 2.0 at 2026-06-25T18:40Z. Confidence is high, verified against Runway's news page. Classification: official release.
@midjourney: posted the --preview V8.2 aesthetics and personalization signal at 2026-06-25T19:11Z, plus large-batch draft mode and --sref random. Classification: X-first / preview pending verification.
@xiaohu: posted at 2026-06-25T11:17Z that "Fable 5 may return tonight." There is no official confirmation, so this stays in the pending-confirmation list and should not be written as available. The latest daily.xiaohu.ai issue still highlights Claude Tag, Seedance 2.5, Mistral OCR 4, OpenAI Record and Replay, and Gemini 3.5 Flash Computer Use as recent priorities; use it as an auxiliary lead source, not a replacement for official sources. Classification: unverified rumor + media/personal source.
@xai / @grok: @xai today mainly said SuperGrok and X subscriptions can be used in T3code; @grok reposted a Vercel AI Gateway signal about Grok Imagine Video usage. No new Grok model ID appeared. Classification: product/ecosystem signal.
@AnthropicAI / @claudeai: the window mostly contained partnerships and product content, with no official new Claude model. Claude Tag is a XiaoHu Daily priority, not today's new model.
@GoogleAI / @GeminiApp: beyond Gemini 3.5 Flash Computer Use, there was no official Gemini Omni or Gemini image-model release in this window. gemini-omni-flash remains a LMArena leaderboard model until official docs appear.
@Alibaba_Qwen: no new Qwen model release in this window; Qwen-AgentWorld was the previous window's major item and remains worth testing.
@MistralAI: OCR 4 was the previous window's major item; no new model release appeared in this window.
@AIatMeta, @cohere, @LumaLabsAI, @Kling_ai, @SakanaAILabs: no official signal in this window rose to a model-first-release level; most posts were product, event, case-study, or hiring content.
Sources: @GoogleDeepMind, @ArtificialAnlys MAI-Image-2.5, @ArtificialAnlys HappyHorse, @arena GLM, @OpenAI agents, @OpenAIDevs Codex mobile, @OpenAIDevs DigitalOcean, @runwayml Agent 2.0, @midjourney preview, @xiaohu Fable 5 rumor, XiaoHu AI Daily 2026-06-25

Pending Confirmation

Claude Fable 5 return rumor: @xiaohu said it "may return tonight", but there is no Anthropic/Claude/LMArena official confirmation of availability changes. Keep it as an unverified rumor.
gemini-omni-flash: still #1 in LMArena T2V, but no official Google Gemini model card, API ID, price, or availability scope has appeared. Keep it as a leaderboard appearance.
HappyHorse-1.0: Artificial Analysis video editing #1 and LMArena video edit #2 can both be true, but the boards use different methodologies. Confirm the official model provider, API availability, edit-task definition, audio-editing capability, and commercial terms.
MAI-Image-2.5: the AA image-board signal is strong, but Microsoft's official model card, API pricing, regions, enterprise data policy, and MAI-Image-2.5-Flash capability differences still need verification.
Midjourney V8.2: currently an early --preview signal. Wait for Midjourney docs/release notes to confirm formal version status, default behavior, GPU cost, parameter compatibility, and whether Discord and web support match.
Runway Agent 2.0: the product update is official, but it is not a new model ID. Confirm underlying model selection, API access, asset/data boundaries, team permissions, and ad-performance analysis capabilities.
OpenAI Record & Replay: official docs exist, but availability depends on macOS, Computer Use, and regional constraints. Verify whether the current account can use it and whether generated skills can be reused reliably inside Codex automation.
Gemini 3.5 Flash Computer Use: official, but preview. Test browser/mobile/desktop environments, sensitive-action confirmation, prompt-injection stop behavior, latency, cost, and failure recovery.
Nex N2 Pro, Grok 4.3 via Amazon Bedrock: Price Per Token listed these in the past week, but they are not official first releases in this window. Confirm model cards, context lengths, regions, and prices against provider docs.