AI Model Daily - 2026-06-28 | QName AI

30-Second Summary

OpenAI officially previewed the GPT-5.6 family: Sol is the new flagship, Terra targets balanced everyday work, and Luna targets fast low-cost usage. OpenAI says this is a limited preview for trusted partners and organizations through API / Codex first, with broader ChatGPT, Codex, and API rollout in the coming weeks; the first @OpenAI X post was at 2026-06-26T17:10Z.
Anthropic's official X account says the US government has allowed Claude Mythos 5 to be redeployed to a set of US organizations that operate and defend critical infrastructure. There is still no clear broad restoration timeline for Fable 5. Treat this as a high-priority model-availability and policy signal.
HappyHorse 1.1 is now in Arena's Video Arena for Text-to-Video, Image-to-Video, and Video Edit; Artificial Analysis has also added HappyHorse-1.1 to its video boards. In T2V without audio, HappyHorse-1.0 is #1 and HappyHorse-1.1 is #2; in T2V with audio, Dreamina Seedance 2.0 720p leads.
Microsoft MAI-Code-1-Flash is now GA for GitHub Copilot Business / Enterprise; Runway API added seedance2_mini; Sakana Fugu Ultra is available on Vercel AI Gateway. These are callable-entry changes, not just leaderboard rumors.
@xiaohu highlighted Alibaba Wan team's Wan Streamer, described as a live video-call human AI. For now, classify it as an X/media signal until an official Wan model card, API page, or paper confirms it.

Must Watch

1. GPT-5.6 Sol / Terra / Luna: OpenAI's limited preview of the next model family

Event type: official release + X-first signal + availability change.
First/official/availability time: the first @OpenAI post was at 2026-06-26T17:10Z; OpenAI's official page is dated 2026-06-26; X displays it as Jun 27 in HKT-side views.
Model/function names: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, max reasoning effort, ultra mode.
Capability change: OpenAI positions Sol as the GPT-5.6 flagship, emphasizing gains in coding, biology, cybersecurity, and agentic workflows; Terra targets balanced everyday work and is described as near GPT-5.5-level performance at roughly 2x lower cost; Luna targets the fastest, lowest-cost use cases. OpenAI also introduced a new max reasoning-effort level and an ultra mode that uses subagents to speed up complex work.
Availability: during the limited preview, GPT-5.6 is first available through API and Codex for a small set of trusted partners and organizations; OpenAI says it will expand to ChatGPT, Codex, and API later. Official pricing per 1M tokens is Sol $5 input / $30 output, Terra $2.50 input / $15 output, and Luna $1 input / $6 output; explicit cache breakpoints and a 30-minute minimum cache life were also added.
Safety/governance signal: OpenAI's system card treats Sol, Terra, and Luna as High capability for both Cybersecurity and Biological/Chemical categories, while not meeting the High threshold for AI Self-Improvement. OpenAI also emphasizes layered real-time monitoring, account-level enforcement, and trusted access. METR's preliminary evaluation warns that Sol's software-engineering automation potential may be understated and recommends extra caution for delegation involving sensitive assets.
Impact: this is the most important model release in this issue. The immediate tests should not stop at benchmark rank: verify whether max/ultra changes long-horizon agent success, what real Codex/API quotas and latency look like, and whether safety enforcement affects legitimate defensive cyber or coding workflows.
Sources: OpenAI: Previewing GPT-5.6 Sol, GPT-5.6 Preview System Card, METR preliminary evaluation, @OpenAI GPT-5.6, @OpenAI cybersecurity

2. Claude Mythos 5: limited restored access for US critical-infrastructure organizations

Event type: official release + availability change + policy/compliance event.
First/official/availability time: @AnthropicAI posted at 2026-06-27T00:29Z; Anthropic's Claude Fable 5 / Mythos 5 official page was previously updated on 2026-06-12 to say access had been paused.
Model/function names: Claude Mythos 5, Claude Fable 5.
Capability change: Anthropic says it has been working with the US government since June 12 to restore access to Mythos 5 and Fable 5; the current government notification allows Mythos 5 to be redeployed to a set of US organizations that operate and defend critical infrastructure. Third-party coverage generally interprets this as limited restoration, not broad reopening.
Availability: limited to an approved set of US organizations; ordinary users, overseas customers, and Fable 5 restoration timing remain unclear.
Impact: Mythos 5 and GPT-5.6 Sol are now both part of the same governance question: how to phase access to frontier cyber-capable models. For enterprise users, access rights, geography, and trusted-access review can matter as much as raw model capability.
Sources: @AnthropicAI Mythos 5, Anthropic: Claude Fable 5 and Claude Mythos 5, Engadget coverage, Axios coverage

3. HappyHorse 1.1: in Video Arena, with early AA rankings visible

Event type: leaderboard appearance + ranking change.
First/official/availability time: @arena said at 2026-06-26T21:56Z that HappyHorse 1.1 is in Video Arena; Artificial Analysis lists HappyHorse-1.1 among video models added in the last month.
Model/function names: HappyHorse 1.1, HappyHorse 1.0, Video Arena, Artificial Analysis Video Arena.
Capability change: Arena says HappyHorse 1.1 covers Text-to-Video, Image-to-Video, and Video Edit, with scores still accumulating. Artificial Analysis currently shows HappyHorse-1.0 leading T2V without audio at Elo 1290, HappyHorse-1.1 at Elo 1285 in #2; in T2V with audio, Dreamina Seedance 2.0 720p leads at Elo 1219 and HappyHorse-1.1 is #2 at Elo 1152; in I2V with audio, Dreamina Seedance 2.0 720p leads at Elo 1194, with HappyHorse-1.1 around Elo 1118/1119 in #2.
Availability: Arena / Artificial Analysis voting and comparison are available; Runway API already has happyhorse_1_0, but the official Runway API changelog does not yet show a callable happyhorse_1_1 entry.
Impact: HappyHorse 1.1 has reached the observable-arena stage, but not yet the stable-API stage. Keep watching three signals: Video Edit rank, audio-dimension advantage, and whether a happyhorse_1_1 API ID appears.
Sources: @arena HappyHorse 1.1, Arena Text-to-Video, Artificial Analysis Text to Video, Artificial Analysis Image to Video, Runway API changelog

4. MAI-Code-1-Flash: GA for GitHub Copilot Business / Enterprise

Event type: official release + availability change.
First/official/availability time: the GitHub Changelog entry is dated 2026-06-26; @MicrosoftAI posted about it at 2026-06-26T17:30Z and 2026-06-26T19:24Z.
Model/function name: MAI-Code-1-Flash.
Capability change: GitHub describes MAI-Code-1-Flash as a Microsoft AI-built coding model that is now GA for Copilot Business and Copilot Enterprise; it is optimized for GitHub Copilot, with emphasis on low latency, high-frequency iteration, and agentic coding workflows.
Availability: Copilot Business / Enterprise admins must enable the MAI-Code-1-Flash policy in Copilot settings; usage follows provider list pricing and usage-based billing.
Impact: this is a practical signal that Microsoft's own model is entering enterprise coding workflows. The near-term evaluation should focus on latency, token cost, long-task stability, and differences against Claude / GPT coding models across VS Code Copilot Chat, code completions, and asynchronous coding agent flows.
Sources: GitHub Changelog: MAI-Code-1-Flash, @MicrosoftAI GA, @MicrosoftAI coding model

5. Runway API: Seedance 2.0 Mini and ad-localization Recipe

Event type: official release + API availability change.
First/official/availability time: Runway API changelog is dated 2026-06-26; @runwayml posted at 2026-06-27T13:02Z that Localize Ads is now available as a Runway API Recipe.
Model/function names: seedance2_mini, Runway Recipes, Ad Localization / Localize Ads.
Capability change: seedance2_mini supports text-to-video, image-to-video, and video-to-video, with 4-15 second outputs, 480p/720p, 16 credits/second, and a 64-credit minimum. Runway's developer portal also shows an Ad Localization Recipe: input an ad image, output versions in target languages while preserving visual creative and layout.
Availability: Runway API text-to-video, image-to-video, and video-to-video endpoints; Runway Developer Portal / Recipes.
Impact: Runway is productizing at two layers: bringing third-party video models into the lower-level API, and packaging ad/product-asset workflows into single-call Recipes above it. For AI video products, the Recipe layer may be closer to customer budget than raw model calls.
Sources: Runway API changelog, Runway Developer Portal, @runwayml Localize Ads

6. Fugu Ultra: Sakana's composition model reaches Vercel AI Gateway

Event type: official/platform availability change.
First/official/availability time: Vercel changelog is dated 2026-06-22; @SakanaAILabs posted again at 2026-06-27T11:47Z.
Model/function names: Sakana Fugu Ultra, sakana/fugu-ultra.
Capability change: Vercel describes Fugu Ultra not as a single model, but as a system built on multiple publicly accessible frontier models, dynamically routing each prompt to 1-3 agents and merging the results. Vercel says it approaches Claude Mythos Preview and Fable 5 on reasoning / scientific benchmarks.
Availability: Vercel AI Gateway, with AI SDK model set to model: 'sakana/fugu-ultra'.
Impact: Fugu Ultra matters less as "one more model ID" and more as a commercialized model-router / multi-agent composition endpoint. It is suitable to test for complex reasoning, code review, and scientific Q&A, while tracking underlying model pool, data retention, cost, and reproducibility.
Sources: Vercel Changelog: Sakana Fugu Ultra, @SakanaAILabs

7. Wan Streamer: an Alibaba real-time video-call model signal surfaced by Xiaohu

Event type: X-first/media signal + pending verification.
First/official/availability time: @xiaohu published the article at 2026-06-27T09:10Z.
Model/function name: Wan Streamer.
Capability change: Xiaohu says Alibaba Tongyi Lab's Wan team released Wan Streamer, a human AI that can hold real-time video calls. This is highly worth tracking, but this run did not find an official Wan blog, model card, API doc, or paper page.
Availability: Xiaohu's X article; official product/API access is still pending confirmation.
Impact: if officially confirmed, this would push video generation from offline clips toward real-time avatar / video-call interaction. Until an official source appears, keep it as a high-priority unverified signal.
Sources: @xiaohu Wan Streamer, Xiaohu AI Daily

Arena Watch

LMArena / Arena Video: HappyHorse 1.1 is now in Text-to-Video, Image-to-Video, and Video Edit. Arena's official X account says scores are coming, so mark it as a leaderboard appearance rather than a settled ranking.
LMArena Text-to-Video: public pages still show gemini-omni-flash at #1, around 1527±14; Dreamina Seedance 2.0 720p is #2 and HappyHorse-1.0 is #3. gemini-omni-flash still lacks an official Google model card/API ID, so continue to classify it as an arena-only appearance.
Artificial Analysis Text-to-Video: with audio is led by Dreamina Seedance 2.0 720p at Elo 1219, followed by HappyHorse-1.1 at Elo 1152 and HappyHorse-1.0 at Elo 1123; without audio is led by HappyHorse-1.0 at Elo 1290, followed by HappyHorse-1.1 at Elo 1285 and Dreamina Seedance 2.0 720p at Elo 1274.
Artificial Analysis Image-to-Video: with audio is led by Dreamina Seedance 2.0 720p around Elo 1194, followed by HappyHorse-1.1 around Elo 1118/1119 and grok-imagine-video-1.5-preview; without audio is led by Dreamina Seedance 2.0 720p at Elo 1343, with grok-imagine-video, grok-imagine-video-1.5-preview, and PixVerse V6 close behind, while HappyHorse-1.1 sits at Elo 1312 in #5.
Artificial Analysis Text-to-Image: GPT Image 2 (high) remains #1 at Elo 1338; MAI-Image-2.5 is #2 at Elo 1273; HiDream-O1-Image-1.5 is #3 at Elo 1263; GPT Image 1.5 (high) is #4 at Elo 1262; Nano Banana 2 is #5 at Elo 1254.
Arena Code / WebDev: in the previous window, @arena said GLM-5.2 Max ranked above Claude Opus 4.8 Thinking in Code Arena and was approaching Claude Fable 5. This window did not show a new Code Arena leader change, but GLM-5.2 remains worth tracking.
Artificial Analysis product signal: @ArtificialAnlys posted at 2026-06-26T19:21Z that Model Sets are live, allowing saved custom model sets across charts. This is a product feature, not a model ranking change.
Sources: Arena Text-to-Video, Arena Video Editing, @arena HappyHorse 1.1, @arena GLM-5.2, Artificial Analysis Text to Video, Artificial Analysis Image to Video, Artificial Analysis Text to Image, @ArtificialAnlys Model Sets

X Watch

@OpenAI: first posted GPT-5.6 Sol / Terra / Luna at 2026-06-26T17:10Z; the same thread stresses that Sol is OpenAI's strongest cybersecurity model and that it uses OpenAI's strongest safety stack. Classification: official release + X-first signal.
@OpenAIDevs: posted Codex quality-of-life updates after 2026-06-27T17:30Z, including long-thread scrolling, thread switching, local-history loading, archived chats, badge sync, Markdown copying, large text pastes, and image previews. Classification: official product update, not a new model.
@AnthropicAI: posted Mythos 5 limited restored access at 2026-06-27T00:29Z. Classification: official release + availability change.
@arena: posted HappyHorse 1.1 entering Video Arena at 2026-06-26T21:56Z; posted at 2026-06-26T17:41Z that GLM-5.2 Max ranks above Claude Opus 4.8 Thinking in Code Arena. Classification: leaderboard appearance.
@ArtificialAnlys: posted Model Sets at 2026-06-26T19:21Z; HappyHorse-1.1 ranking should be read from the AA leaderboard pages. Classification: leaderboard/product auxiliary source.
@xiaohu: posted the Wan Streamer article at 2026-06-27T09:10Z; reposted a real-time webcam identity-swap demo at 2026-06-27T02:16Z; mentioned Codex quota resets at 2026-06-27T01:27Z. Classification: media/personal source, with Wan Streamer requiring verification.
@runwayml: posted at 2026-06-27T13:02Z that Localize Ads is available as a Runway API Recipe. Classification: official release + API workflow update.
@SakanaAILabs: posted at 2026-06-27T11:47Z that Fugu-Ultra is available on Vercel AI Gateway; reposted Fugu Technical Report discussion at 2026-06-27T18:38Z. Classification: official/platform availability change.
@MicrosoftAI: posted at 2026-06-26T17:30Z and 19:24Z about MAI-Code-1-Flash in Copilot Business / Enterprise. Classification: official release + availability change.
@GoogleDeepMind / @GeminiApp: no new official Gemini Omni / Nano Banana model release in this window; Gemini 3.5 Flash Computer Use remains a previous-window highlight, while GeminiApp only surfaced monthly Gemini Drops / Thinking Levels product signals. Classification: no new model launch.
@Alibaba_Qwen: no new Qwen model in this window; Qwen-AgentWorld remains a previous-window item to test. Classification: no new model launch.
@MistralAI: no new model release in this window; Mistral OCR 4 was a previous-window highlight. Classification: no new model launch.
@xai / @grok: mostly ecosystem signals such as SuperGrok/X subscriptions in T3code and Grok Build plugins; no new Grok model ID. Classification: product/ecosystem signal.
@AIatMeta, @cohere: no official new model signal in this window. Classification: no new model launch.
@LumaLabsAI: mostly Ray 3.2 multi-keyframe work/case-study reposts. Classification: showcase signal, not a model launch.
@Kling_ai: mostly showcases and campaign posts, with no new Kling model ID. Classification: no new model launch.
Sources: @OpenAI, @OpenAIDevs, @AnthropicAI, @arena, @ArtificialAnlys, @xiaohu, @runwayml, @SakanaAILabs, @MicrosoftAI

Pending Verification

GPT-5.6 real rollout schedule: OpenAI says the limited preview will expand more broadly within weeks, but ChatGPT/Codex/API account batches, regions, rate limits, API model IDs, and Cerebras 750 tok/s access need real verification.
GPT-5.6 safety and production usability: the system card treats Bio/Chem and Cyber as High capability categories; METR also raises software-engineering automation and delegation risk. Track third-party evals, Codex real-world traces, and false positives from enterprise safety enforcement.
Claude Fable 5 restoration timing: Anthropic's official X post only clearly says Mythos 5 can be redeployed to some US critical-infrastructure organizations; whether, when, and under what constraints Fable 5 returns remains unconfirmed.
Mythos 5 access scope: third-party coverage mentions more than 100 US agencies or critical-infrastructure organizations, but Anthropic has not published a complete list. Avoid treating limited redeployment as broad availability.
HappyHorse 1.1 callable access: Arena / AA make it observable, but Runway API changelog still lists only happyhorse_1_0. Keep watching for a happyhorse_1_1 API ID, pricing, audio support, video-edit capability, and commercial terms.
Wan Streamer official source: currently the main signal is @xiaohu's article. It needs a Wan / Alibaba official blog, model card, paper, demo URL, or API doc to confirm real-time video-call capability and access.
gemini-omni-flash: still #1 in Arena T2V, but there is no Google official model card, API ID, pricing, or Gemini product entry. Continue to treat it as an arena-only appearance.
MAI-Code-1-Flash: GA for Copilot Business / Enterprise, but ordinary developer API access, standalone model calls, exact pricing, data boundaries, and license restrictions need further GitHub / Microsoft documentation.
Runway Localize Ads Recipe: X and the Developer Portal show the signal, but exact endpoint docs, input/output schema, batch limits, pricing, and brand-consistency boundaries still need verification.
Fugu Ultra: Vercel makes it callable, but the underlying model pool, routing policy, privacy/data retention, latency, and cost reproducibility still need testing.