AI Model Daily - 2026-06-24 | QName AI

30-Second Summary

The clearest official model release today is Mistral OCR 4: Mistral released a new document-understanding/OCR model on 2026-06-23, with support for 170 languages, bounding boxes, block classification, inline confidence scores, and availability through the API, Mistral Studio Document AI, Amazon SageMaker, Microsoft Foundry, and other entry points.
Anthropic released Claude Tag: a Slack-based team collaboration agent for Claude Enterprise and Team customers in beta. It uses Opus 4.8 and supports channel-scoped memory, ambient behavior, asynchronous long-running tasks, and administrator permission controls. This is a product/agent update, not a new foundation model.
ByteDance video models showed strong X/media signals: @xiaohu posted that the Seedance 2.0 upgrade and Seedance 2.5 appeared at a ByteDance/Volcengine event, with 30-second video generation, native 4K, 50 multimodal references, 3D white-model support, and an AI copyright commercialization platform. I did not find an official ByteDance/Volcengine Seedance 2.5 model page, so this is classified as X first signal/media coverage pending official documentation.
Artificial Analysis launched the Speech to Speech Index, bringing native speech-to-speech models into a unified evaluation: GPT-Realtime-2 (High) ranks first at 77.2%, followed by Grok Voice Think Fast 1.0 at 75.7%, GPT-Realtime-1.5 at 72.0%, and Gemini 3.1 Flash Live Preview (High) at 69.5%.
Price Per Token shows Last 24h 3: besides a Latent Space article, the new items are GLM-5.2 provider entries through GMICloud and Venice. X also showed a signal that GLM-5.2 is available in the Perplexity Agent API. There was no new general-purpose LLM foundation model ID launch today.

Must Watch

1. Mistral OCR 4: an official new document-understanding model

Event type: official release + model availability update.
First/official/availability time: Mistral's official article is dated 2026-06-23; @MistralAI posted the X thread on 2026-06-23T14:00Z.
Model/function name: Mistral OCR 4, mistral-ocr-latest, Document AI.
Capability change: OCR 4 does more than extract text. It also returns bounding boxes, typed-block classification, and per-page/per-word confidence scores. Mistral says it supports 170 languages across 10 language groups, handles enterprise formats including PDF, DOC, PPT, and OpenDocument, and can be used as a structured entry point for RAG, enterprise search, form/invoice processing, compliance review, and agentic document workflows.
Availability: Mistral API, Mistral Studio Document AI, Amazon SageMaker, and Microsoft Foundry; Snowflake Parse Document support is coming; enterprise customers can self-host with a single-container deployment. Mistral lists pricing at API $4 / 1000 pages, Batch API $2 / 1000 pages, and Document AI $5 / 1000 pages.
Impact: this is today's most directly testable callable model. It is not a general-purpose LLM, but it is immediately useful for enterprise RAG, knowledge bases, finance/legal document pipelines, and review workflows. Bounding boxes plus block types plus confidence scores should make citations, redaction, human review loops, and structured extraction more stable.
Sources: Mistral OCR 4, @MistralAI OCR 4

2. Claude Tag: a Slack team agent powered by Opus 4.8

Event type: official release + X first signal + product/agent availability update.
First/official/availability time: Anthropic's official article is dated 2026-06-23; @claudeai posted the Claude Tag thread on 2026-06-23T17:12Z.
Model/function name: Claude Tag, @Claude, Opus 4.8.
Capability change: Claude can join selected Slack channels as a team member, with access to administrator-approved tools, data, and codebases; team members delegate tasks by tagging @Claude. Anthropic emphasizes multiplayer context, channel-scoped memory, ambient behavior, asynchronous tasks, future task planning, and returning work inside Slack threads.
Availability: Slack beta for Claude Enterprise and Team customers; it replaces the old Claude in Slack app. Administrators have a 30-day opt-in migration window, and Anthropic is providing launch credits to eligible organizations.
Impact: this is not a new model release, but it matters for agent product shape. Anthropic is moving Claude Code/Cowork-style long-running task ability into the team collaboration surface. The important part is not "a chatbot in Slack"; it is a team-grade agent with permission boundaries, memory boundaries, spending limits, and audit logs.
Sources: Anthropic: Introducing Claude Tag, @claudeai Claude Tag

3. Seedance 2.5: strong first signal, but the official model page still needs verification

Event type: X first signal + media/conference coverage + pending official documentation.
First/official/availability time: @xiaohu posted on 2026-06-23T03:52Z; several Chinese media/blog reports say the announcement came from the 2026-06-23 Volcengine FORCE conference. I did not find an official ByteDance Seed or Volcengine Seedance 2.5 model page.
Model/function name: Seedance 2.5, Seedance 2.0 upgrade.
Capability change: visible signals say Seedance 2.5 supports one-shot 30-second short video generation, up to 50 multimodal reference inputs, and 3D white-model support. Seedance 2.0 is also said to receive native 4K support. There is also an AI copyright commercialization platform signal, described as enabling creation with officially licensed IP templates and revenue sharing.
Availability: media reports say Seedance 2.5 is in global enterprise internal testing and is expected to open publicly in early 2026-07, including API access; Seedance 2.0 already has official API docs. Runway also posted on the same day that Seedance 4K, Seedance Mini, and Kling 3.0 Turbo are available on Runway, but that is a platform-listing signal and still needs cross-checking against official model-provider pages.
Impact: if the 30-second native generation and 50 multimodal reference claims are accurate, this is today's strongest video-model capability signal. The evidence level still matters: the report should not treat X/media claims as a completed public API launch until official documentation appears.
Sources: @xiaohu Seedance 2.5, CSDN conference coverage, Volcengine Seedance 2.0 API docs, @runwayml platform listing signal

4. Artificial Analysis Speech to Speech Index: native speech models enter a unified leaderboard

Event type: public leaderboard/evaluation release.
First/official/availability time: Artificial Analysis's article is dated 2026-06-23; @ArtificialAnlys posted the X thread on 2026-06-23T15:03Z.
Model/function name: Artificial Analysis Speech to Speech Index, GPT-Realtime-2, Grok Voice Think Fast 1.0, Gemini 3.1 Flash Live Preview.
Capability change: the new index combines Speech Reasoning from Big Bench Audio, Conversational Dynamics from a Full Duplex Bench subset, and Agentic Performance from tau-Voice. The current leaderboard has GPT-Realtime-2 (High) first at 77.2%, Grok Voice Think Fast 1.0 second at 75.7%, GPT-Realtime-1.5 third at 72.0%, and Gemini 3.1 Flash Live Preview (High) fourth at 69.5%.
Availability: Artificial Analysis public leaderboard and methodology pages; direct availability for each model still depends on OpenAI, xAI, and Google API/product permissions.
Impact: this is today's most important evaluation-infrastructure change. Speech models are no longer being compared only by latency or isolated ASR/TTS ability; they are beginning to be evaluated across reasoning, conversational timing, and tool/customer-service task completion. This is directly useful for voice-agent model selection.
Sources: AA Speech to Speech Index announcement, AA Speech to Speech leaderboard, @ArtificialAnlys thread

5. GLM-5.2: new Perplexity Agent API and provider availability

Event type: provider availability update + X signal + aggregator verification.
First/official/availability time: the Perplexity Agent API signal reposted by @Zai_org appeared on 2026-06-23T03:43Z; Price Per Token showed GLM-5.2 as newly available through GMICloud and Venice on 2026-06-23.
Model/function name: GLM-5.2.
Capability change: the model itself is not new today; the change is distribution expansion. The Perplexity Agent API docs page shows third-party model comparison, and search results explicitly mention GLM 5.2; the X signal says it pairs with the Search SDK and sandbox for wide and deep research. Price Per Token lists new provider pricing: GMICloud at $0.98 in / $3.08 out, Venice at $1.40 in / $4.40 out, with a 1M context window.
Availability: Perplexity Agent API, GMICloud, Venice; earlier signals already included Z.ai, OpenRouter, AWS Marketplace, and other entries.
Impact: GLM-5.2's key change is not another benchmark result; it is whether teams can call it reliably. If Perplexity Agent API exposes GLM-5.2 together with Search as Code, the model moves further from "strong open model" into agent runtime infrastructure.
Sources: Perplexity Agent API models, @Zai_org / Perplexity signal, Price Per Token model releases

Arena Watch

LMArena / Arena Agent: the overall board did not show a major reshuffle. Claude Fable 5 (High) remains No. 1 with Net Improvement 14.05%+/-1.53%; Claude Opus 4.8 (Thinking) is No. 2, GPT 5.5 (xHigh) No. 3, Claude Opus 4.7 Thinking No. 4, GPT 5.5 High No. 5, and GLM 5.2 Max No. 10.
Arena Text: the top slots remain dominated by Claude models, with claude-fable-5 at 1508+/-9 No. 1, claude-opus-4-6-thinking at 1504+/-4 No. 2, and claude-opus-4-7-thinking at 1502+/-5 No. 3; muse-spark is No. 6, gemini-3.1-pro-preview No. 7, gemini-3-pro No. 8, and gpt-5.5-high No. 10.
Arena WebDev: claude-fable-5 is No. 1 at 1654+16/-16, glm-5.2 (max) No. 2 at 1593+15/-15, and claude-opus-4-8-thinking No. 3; seed-2.1-pro-preview remains No. 8 at 1539+15/-15, marked preliminary.
Arena Text-to-Image: gpt-image-2 (medium) is No. 1 at 1385+/-6, reve-2.0 No. 2, gemini-3.1-flash-image-preview (nano-banana-2) [web-search] No. 3, mai-image-2.5 No. 4, and gemini-3-pro-image-preview-2k (nano-banana-pro) No. 5.
New Arena X signal: @arena said on 2026-06-23T02:15Z that Cosmos 3 Super by NVIDIA ranks No. 8 and No. 11 among open models in Text-to-Image Arena, corresponding to No. 49 and No. 54 overall. NVIDIA Cosmos 3 itself was a May/earlier open world model release; today this should be treated as an arena ranking signal, not a new model launch.
Arena Text-to-Video: the board is still led by gemini-omni-flash at 1527+/-14, followed by dreamina-seedance-2.0-720p at No. 2, happyhorse-1.0 at No. 3, veo-3.1-audio-1080p at No. 4, and sora-2-pro at No. 7.
Arena Image-to-Video: gemini-omni-flash and dreamina-seedance-2.0-720p are both high at 1475, followed by grok-imagine-video-1.5-preview-720p No. 3, happyhorse-1.0 No. 4, and kling-v3-pro No. 12.
Arena Video Edit: dreamina-seedance-2.0-720p is No. 1, happyhorse-1.0 No. 2, grok-imagine-video No. 3, kling-o3-pro No. 4, and runway-gen4-aleph No. 6.
Artificial Analysis Intelligence: the homepage shows Claude Fable 5 (with fallback) at 60, but not currently available; GPT-5.5 (xhigh) at 55, GLM-5.2 (max) at 51, Gemini 3.5 Flash at 50, Claude Sonnet 4.6 (max) at 47, Gemini 3.1 Pro Preview at 46, and Qwen3.7 Max at 46.
Artificial Analysis Speech to Speech: the new voice leaderboard has GPT-Realtime-2 (High) leading at 77.2%, with Grok Voice Think Fast 1.0 second at 75.7%; for speed, Deepslate Opal has the fastest TTFA at 0.44s; for cost, Gemini 3.1 Flash Live Preview (Minimal) is lowest at $1.50 per input audio hour.
Price Per Token: Last 24h 3, This Week 36, This Month 128, All 156. Today's items include a Latent Space article and new GLM-5.2 entries through GMICloud and Venice; yesterday's items still include OpenAI Daybreak and other GLM-5.2 provider entries.
Sources: Arena overview, Arena WebDev, Arena Text-to-Video, Arena Image-to-Video, @arena Cosmos 3 Super, Artificial Analysis, AA Speech to Speech, Price Per Token

X Watch

@MistralAI: posted the Mistral OCR 4 thread on 2026-06-23T14:00Z; high confidence, verified by Mistral's official article. This can be written as an official model release.
@claudeai / @AnthropicAI: posted/reposted Claude Tag on 2026-06-23T17:12Z; high confidence, verified by Anthropic's official article. This should be written as a product/agent update, not as a new model.
@xiaohu: posted the Seedance 2.0 upgrade and Seedance 2.5 event signal on 2026-06-23T03:52Z, with meaningful distribution; because no official model page was found, write it as X first signal/media coverage and keep it in pending verification.
@ArtificialAnlys: posted the Speech to Speech Index thread on 2026-06-23T15:03Z; high confidence, verified by the Artificial Analysis article and leaderboard page.
@arena: posted on 2026-06-23T02:15Z that Cosmos 3 Super ranked among open models in Text-to-Image Arena; write this as arena appearance/ranking signal.
@Zai_org: reposted the signal that GLM-5.2 is available in the Perplexity Agent API; together with Perplexity docs and Price Per Token's new provider entries, this forms evidence of availability expansion.
@OpenAI: the official X signal in this window was DevDay 2026 applications opening at 2026-06-23T18:10Z; OpenAI's website also had a GPT-5 Pro immunology research case. Neither is a new model release.
@GeminiApp: this window included reposts about Music Generation in Gemini on Pixel and Nano Banana 2 template usage; there was no new Gemini foundation model or Gemini Omni official model card.
@GoogleDeepMind / @GoogleAI: this window was mostly Genie/Cannes, RL/world-model research, and older pinned Gemini 3.5 Live Translate information; no new model release.
@runwayml: said on 2026-06-23T21:36Z that Seedance 4K, Seedance Mini, and Kling 3.0 Turbo are available on Runway; this is a platform-listing signal and still needs official model-provider verification.
@xai / @grok: this window included the Firecrawl plugin for the Grok Build marketplace and Interactive Brokers connection for Grok; /goal was from the previous window. There was no new Grok model ID today.
@Kling_ai: this window contained WILD MV, creation campaigns, and prompt/showcase posts; no new official Kling model release.
@Alibaba_Qwen, @AIatMeta, @cohere, @LumaLabsAI, @midjourney: no new model first-release signal strong enough to move into today's must-watch section.

Pending Confirmation

Seedance 2.5: needs confirmation from ByteDance Seed, Volcengine, or Doubao/Jimeng official model pages for the 30-second generation, 50 multimodal references, 3D white-model support, release timing, API pricing, regions, and usage limits.
Seedance 4K / Seedance Mini / Kling 3.0 Turbo on Runway: Runway gave a platform-listing signal, but Seed/Kling official model docs, version numbers, resolution, duration, pricing, and API entry points still need verification.
Cosmos 3 Super Text2Image / Agentic: Arena X gave open-model ranking positions, but the exact leaderboard rows, overall ranks, model access, and relationship among NVIDIA/fal/Hugging Face versions still need continued verification.
GPT-Realtime-2 and Grok Voice Think Fast 1.0: Artificial Analysis voice rankings are clear, but broad availability, API names, regional restrictions, pricing, and tool-calling boundaries still need to be checked against each provider.
GLM-5.2: distribution continues to expand, but Perplexity Agent API calling limits, Search as Code integration behavior, provider stability, and price differences across providers need follow-up.
Claude Tag: beta availability is clear, but administrator permission model, channel memory isolation, default ambient behavior, token-spend controls, and the exact old Slack app migration date need continued verification.
Gemini Omni / gemini-omni-flash: it remains high on Arena video boards, but this window had no new Google official model card, API pricing, or availability update.
happyhorse-1.0: it remains high on video boards, but the official team, model card, release date, API/pricing, and commercial licensing are still unclear.
nano-banana-2 / nano-banana-pro: Arena and Google/X usage signals continue, but official product naming, API availability, pricing, and relationship to the Gemini image model line still need unified verification.