Collected in Hong Kong time
AI Model Daily - 2026-07-05
2026-07-05 AI model news: no major official model launch in the past 24 hours; Claude Fable 5 returned to Arena and leads Agent Arena, Doubao and Qwen agent apps are reportedly shutting down on July 15, and Sakana posted sparse LLM and black-box optimization research.
30-Second Summary
- No new official model launch appeared in the past 24 hours from OpenAI, Google, Anthropic, Qwen, Mistral, xAI, Meta, Cohere, Runway, Luma, Kling, Midjourney, or Microsoft AI. Today's main thread is model re-availability, leaderboard re-entry, and product-entry changes.
Claude Fable 5is the most important Arena signal in this window:@arenasays Fable 5 is back in Battle Mode and Agent Mode, while the official leaderboard showsClaude Fable 5 (High)at #1 overall in Agent Arena, with overall13.34% +/- 1.55%, Confirmed Success16.12% +/- 2.68%, and Praise vs Complaint30.63% +/- 5.69%.- Anthropic's Fable 5 / Mythos 5 redeployment is a confirmed
2026-07-01event, not a new launch today. The fresh Arena and X activity shows Fable 5 re-entering real user testing and long-horizon agent evaluation. @xiaohuand several Chinese media reports point to an important product-entry change: Doubao and Tongyi Qianwen agent features are reportedly scheduled to shut down on2026-07-15. This is not a model release, but it affects consumer-facing agent app access and history preservation in China.- Xiaohu AI Daily's public page is still stuck on
2026-06-27, so it is only useful as historical background in this run. The Chinese-side signal today mainly comes from@xiaohuon X cross-checked with public media reports.
Must Watch
1. Fable 5 returns to Arena: Agent Arena #1 overall, with the official leaderboard reflecting re-availability
- Event type: arena appearance + leaderboard movement + continued availability watch.
- First/official/availability time: Anthropic's official redeployment update says Fable 5 became available again on
2026-07-01;@arenasaid at2026-07-04T17:26Zthat Fable 5 was back in Arena and testable in Battle Mode and Agent Mode; at2026-07-04T19:11Z, Arena again highlighted its post-return test performance. - Model/function names:
Claude Fable 5,Claude Fable 5 (High), Battle Mode, Agent Mode, Agent Arena. - Leaderboard signal: Arena Agent Leaderboard shows
Claude Fable 5 (High)#1 overall, with overall13.34% +/- 1.55%and16,082sessions; Confirmed Success #1 at16.12% +/- 2.68%; Praise vs Complaint #1 at30.63% +/- 5.69%; Steerability #3 at9.21% +/- 2.92%; Tool Hallucination tied with several models at1.31% +/- 0.13%. - Availability: Anthropic says Fable 5 has been globally restored since
2026-07-01across Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Pro, Max, Team, and some Enterprise plans can use up to 50% of weekly included usage for Fable 5 until2026-07-07; after that, usage credits are required. Arena users can test it in Battle Mode and Agent Mode. - Impact: today is not Fable 5's official launch day, but it is this run's most important observable model signal. After returning, Fable 5 already leads Agent Arena on both user-confirmed task completion and the praise-versus-complaint metric, which is closer to real agent quality than a single demo. Continue watching whether leaderboard scores stabilize and whether Anthropic's new classifier causes false positives for legitimate coding or security work.
- Sources: Anthropic: Redeploying Fable 5, Arena Agent Leaderboard, Arena Text Leaderboard, @arena Fable 5 back, @arena Battle/Agent Mode
2. Doubao / Tongyi Qianwen agent features reportedly shutting down: not a model launch, but it changes agent access
- Event type: media report + X-first personal signal + availability change.
- First/official/availability time:
@xiaohusaid at2026-07-04T04:10Zthat Doubao and Qianwen agent apps had been removed; multiple Chinese media outlets reported on2026-07-04that Doubao and Tongyi Qianwen agent features will officially shut down on2026-07-15. - Model/function names: Doubao agents, Qianwen agents, Tongyi Qianwen agent service.
- Capability/access change: reports say that after Doubao's agent feature shuts down, users can still view and save agent information and historical conversations for a period; after
2026-10-15, related data will be handled under the privacy policy. Qianwen also reportedly says its agent features and service will shut down on2026-07-15, after which users will no longer be able to access related agent configurations or historical conversation records. - Availability: this is a product feature shutdown notice, not a model API or model-card change. Reports say Doubao points users toward ByteDance's Maoxiang app for creating agents; Qianwen recommends saving important content in advance.
- Impact: this directly affects consumer users in China who build or run agents through Doubao or Qianwen. In the short term, users should back up configurations and conversations. It does not mean Qwen model capability has regressed, nor does it prove Qwen API or enterprise agent products are shutting down. The exact boundary still needs confirmation from ByteDance / Alibaba official product notices and developer docs.
- Sources: @xiaohu, Sina Finance report, Sohu repost
3. Anthropic Fable 5 safeguards: the return comes with a stronger cyber classifier
- Event type: official release continuation + safety/governance event.
- First/official/availability time: Anthropic's redeployment post is dated
2026-06-30and was updated to say Fable 5 / Mythos 5 returned on2026-07-01; Arena's return and user testing brought it back into focus in this window. - Model/function names:
Claude Fable 5,Claude Mythos 5, cyber safety classifiers, jailbreak severity framework. - Capability/governance change: Anthropic says export controls on Fable 5 and Mythos 5 have been lifted. Fable 5 returned with a new classifier intended to block the security bypass behavior described in Amazon's report. Anthropic says the new classifier blocks that specific technique in more than 99% of cases, while also acknowledging that routine benign coding/debugging requests may see more false positives.
- Availability: Fable 5 has been restored globally. Mythos 5 has first been restored for a set of US organizations and continues expanding to Glasswing partners.
- Impact: Fable 5's return is not just "a stronger model"; it is a production experiment in combining a strong model with heavier safety gating. For coding and security users, the key test is not only the leaderboard. The practical question is whether legitimate security research, vulnerability reproduction, code review, debugging, and DevOps requests are incorrectly blocked, and how good the fallback to Opus 4.8 feels.
- Sources: Anthropic: Redeploying Fable 5, Anthropic Fable/Mythos status page, @AnthropicAI export controls lifted
4. Sakana AI: sparse LLM and black-box optimization research enter the watchlist
- Event type: research post / paper update; not a model release.
- First/official/availability time:
@SakanaAILabspromoted "Sparser, Faster, Lighter Transformer Language Models" again at2026-07-04T13:08Z, then promoted "Bridging Spherical Black-Box Optimizers" at2026-07-04T21:36Z. - Model/function names: TwELL, Sparser/Faster/Lighter Transformer Language Models, Bridging Spherical Black-Box Optimizers.
- Capability change: TwELL is Sakana's work with NVIDIA on sparse packing and GPU-friendly sparse formats, aiming to make sparse Transformers faster, lighter, and better suited to tiled GPU kernels. The black-box optimization paper targets gradient-free optimization for simulators, external APIs, reinforcement learning, and similar settings.
- Availability: papers and blog posts; no callable model API.
- Impact: Sakana did not introduce a new model ID today, but the research direction complements Fugu's model-composition / orchestration path: one side improves model orchestration, while the other targets lower-level efficiency and optimization. Product impact is limited in the short term, but it is worth tracking whether the ideas flow into future Fugu, Marlin, or Sakana Chat inference-cost improvements.
- Sources: @SakanaAILabs sparse LLMs, @SakanaAILabs black-box optimizers, Sakana sparse LLM blog
5. Grok / Luma / Kling: this window is mostly templates, showcases, and events, not model launches
- Event type: official product/content activity; not a model release.
- First/official/availability time:
@grokreposted Grok Imagine America 250 templates at2026-07-04T14:31Z;@LumaLabsAIreposted a Ray 3.2 work at2026-07-04T00:48Z;@Kling_aiposted NEXTGEN Awards content at2026-07-04T03:00Z. - Model/function names: Grok Imagine templates, Luma Ray 3.2, Kling AI NEXTGEN Awards.
- Availability: Grok Imagine, Luma app, Kling AI event/showcase surfaces.
- Impact: these updates help track community distribution and creative use cases for image/video models, but they should not be written up as a Grok 4.5, Luma, or Kling new-model release. Grok 4.5 still lacks xAI blog, docs, API/model-card, or product-entry confirmation; Luma and Kling also showed no new model ID in this window.
- Sources: @grok / @imagine, @LumaLabsAI / Dream Lab, @Kling_ai
Arena Watch
- Agent Arena overall:
Claude Fable 5 (High)#1, overall13.34% +/- 1.55%,16,082sessions;Claude Opus 4.8 (Thinking)#2, overall9.37% +/- 1.29%;GPT 5.5 (xHigh)#3, overall8.21% +/- 1.02%. - Confirmed Success:
Claude Fable 5 (High)#1,16.12% +/- 2.68%;GLM 5.2 (Max)#2,9.13%;Claude Opus 4.8 (Thinking)#3,8.59%. - Praise vs Complaint:
Claude Fable 5 (High)#1,30.63% +/- 5.69%;Claude Opus 4.8 (Thinking)#2,17.48% +/- 4.79%;GLM 5.2 (Max)#3,15.45%. - Text Arena Overall:
claude-fable-5#1, score1509 +/- 9,4,350votes, price shown as$10 / $50per 1M tokens;claude-opus-4-6-thinking#2;claude-opus-4-7-thinking#3. The Text leaderboard date is2026-07-01, so today it should be treated as a continuing leaderboard signal, not a new listing. - Video / Image Arena: no new HappyHorse, Gemini Omni Flash, Nano Banana 2 Lite, Kling, or Luma leaderboard debut or significant movement was found in this 24-hour window. Continue carrying the previous Omni Flash / HappyHorse watch items.
- Sources: Arena Agent Leaderboard, Arena Text Leaderboard, @arena Fable 5
X Watch
@arena: the most important X source in this window. At2026-07-04T17:26Z, Arena said Fable 5 was back and testable in Battle Mode and Agent Mode; at2026-07-04T19:11Z, it called Fable 5 one of the most impressive models it had seen and showed 60+ difficult 3D-generation tests. Classification: arena appearance + continuing leaderboard watch.@xiaohu: at2026-07-04T04:10Z, said Doubao and Qianwen agent apps were removed; Chinese media then reported a2026-07-15shutdown date for the agent features. Classification: X-first/media signal + product-entry change, pending official-boundary confirmation.@SakanaAILabs: posted two research updates in this window, one on black-box optimization and one on TwELL / sparse LLMs for ICML. Classification: research post; not a new model.@grok: reposted Grok Imagine America 250 templates. Classification: product template; not Grok 4.5 official confirmation.@LumaLabsAI: reposted Ray 3.2 work/case material. Classification: showcase signal; not a new model.@Kling_ai: posted NEXTGEN Awards / 4K creators content. Classification: event/showcase signal; not a new model.@OpenAI/@OpenAIDevs: no new model release in this window; the latest important model remains GPT-5.6 and the latest benchmark item remains GeneBench-Pro. Classification: no new model launch.@AnthropicAI/@claudeai: no new official post in this window; Fable 5 redeployment and Sonnet 5 are already-confirmed events from around July 1, with today's focus revived by Arena activity. Classification: no new model launch; Fable 5 remains an availability/quality watch.@GoogleDeepMind/@GoogleAI/@GeminiApp: no new Gemini model after Nano Banana 2 Lite / Gemini Omni Flash was found in this window; Gemini App activity was mainly macOS / MCP / app integration, not a model launch. Classification: no new model launch.@Alibaba_Qwen/@Alibaba_Wan: no new Qwen/Wan official model release appeared in this window; recent Wan activity is around Wan Skills and Music to Dance feature demos. Classification: no new model launch.@MistralAI,@xai,@AIatMeta,@cohere,@runwayml,@midjourney,@MicrosoftAI: no official X signal in this window that should be promoted to a new model launch. Classification: no new model launch.- Sources: @arena Fable 5, @xiaohu, @SakanaAILabs, @grok / imagine, @LumaLabsAI, @Kling_ai
To Confirm
- Fable 5 leaderboard stability: Arena's current Agent Leaderboard shows Fable 5 leading multiple metrics, but Arena's X posts say leaderboard scores are still coming / continuing to accumulate. Watch votes, sessions, and error bars.
- Fable 5 safety false positives: Anthropic explicitly says the new classifier may increase benign false positives on routine coding/debugging. Test code review, vulnerability reproduction, CTF, debugging, and DevOps flows.
- Fable 5 usage credits: after
2026-07-07, included-quota behavior for Pro, Max, Team, and some Enterprise plans needs hands-on verification, especially for Claude Code and Claude Cowork consumption. - Mythos 5 expansion scope: Anthropic still says it is first restored for a set of US organizations and expanding to Glasswing partners; ordinary-user and international partner access remains unclear.
- Doubao / Qianwen agent shutdown boundary: confirm the original official notices and whether the change only affects consumer agents, or also affects APIs, enterprise agents, Qwen-AgentWorld, or Tongyi Qianwen enterprise services.
- Doubao / Qianwen user data: users should back up around the
2026-07-15shutdown and2026-10-15data-handling milestones; exact export methods and recoverable-history scope should follow in-product notices. - Xiaohu AI Daily source: the public page remains stuck on
2026-06-27; keep watching whether updates resume, but do not treat it as a current daily source today. - Grok 4.5 officialization: still no xAI blog, docs, API model list, Grok product-entry, model card, or reproducible third-party test. Grok Imagine templates are not a Grok 4.5 release.
- HappyHorse 1.1 API: no official
happyhorse_1_1API ID, pricing, model card, or commercial terms were found today. - Gemini Omni Flash / Nano Banana 2 Lite: no new change today. Continue watching Google Cloud preview/GA state, Runway integration stability, and real cost.