AI Model Daily - 2026-06-14 | QName AI

30-Second Summary

Today's highest-priority item is Anthropic's availability reversal. The company stated on June 12 that a U.S. government export-control directive forced it to suspend access to Fable 5 and Mythos 5; visible X posts from @AnthropicAI and @ClaudeDevs show the practical impact has expanded to paused Fable 5 sessions for all customers, with new sessions falling back to the default model or Opus 4.8.
Leaderboards and availability are now split. LMArena Text still ranks claude-fable-5 No. 1 overall, and Agent Arena still shows Claude Fable 5 (High) at No. 1, but official production availability is suspended. Treat it as both capability-leading and high availability risk.
Artificial Analysis' key X signal in the last 24 hours was that the "Intelligence Frontier" moved backward for the first time, alongside the launch of AA-AgentPerf. This is an agentic-inference hardware and serving benchmark, not a new foundation model, but it matters for agent product capacity and cost planning.
XiaoHu AI Daily's latest issue led with Perplexity Search as Code. Perplexity's official research page shows SaC is a programmable search architecture for Agent API / Computer, not a new model, but it reports strong performance and cost-frontier gains on DSQA, BrowseComp, WideSearch, and WANDR.
The video-model watchlist has no new official model card. Text-to-Video Arena still has gemini-omni-flash at No. 1 and happyhorse-1.0 at No. 3; the nano-banana image line also showed no new rank or availability change.

Must Watch

1. Anthropic suspends Fable 5 / Mythos 5 access; the leaderboard winner becomes an availability risk

Event type: official release + X first signal + availability change.
First/official/availability time: Anthropic's official statement is dated 2026-06-12; the X posts remained high-visibility during this window. Fable 5 / Mythos 5 originally launched on 2026-06-09.
Model/function name: Claude Fable 5, Claude Mythos 5.
Capability change: there is no official capability regression. The change is access suspension. Anthropic says the U.S. government directive requires suspending access to Fable 5 / Mythos 5 by any foreign national; the compliance effect is that Anthropic must disable the two models for all customers. Other Anthropic models are not affected.
Availability: Fable 5 / Mythos 5 are currently unavailable under Anthropic's official position; @ClaudeDevs says new sessions across Claude products run on the user's default model or Opus 4.8, and Build Day is also switching to Opus 4.8.
Impact: this is today's most important model item. Fable 5 still looks strong on LMArena and Artificial Analysis, but short-term product decisions cannot rely on leaderboards alone. Integrators need to confirm fallback behavior, billing, and whether API, Bedrock, Vertex, Foundry, and Claude product access differ by region or identity.
Sources: Anthropic statement, Claude Fable 5 and Claude Mythos 5, Claude API docs, @AnthropicAI, @ClaudeDevs

2. Agent Arena: Fable 5 remains No. 1, but must now be read together with the access suspension

Event type: arena appearance + leaderboard change.
First/official/availability time: Agent Arena is updated to 2026-06-12, with 648,037 sessions and 25 models.
Model/function name: Claude Fable 5 (High), GPT 5.5 (xHigh), Claude Opus 4.8 (Thinking).
Capability change: Claude Fable 5 (High) is No. 1 with 13.68%±1.59% Net Improvement, 17.21%±2.59% Confirmed Success, and 16,283 sessions. GPT 5.5 (xHigh) is No. 2 with 11.03%±2.52% Net Improvement. Claude Opus 4.8 (Thinking) is No. 3 with 9.05%±1.34%.
Availability: Arena Agent leaderboard; production access must be checked through each model's official channel.
Impact: Fable 5's agentic-task ranking continues to strengthen, but today's real conclusion is capability lead plus availability breakage. Teams selecting models for agent or coding products should re-evaluate Opus 4.8 Thinking and GPT 5.5 xHigh / High as substitutes.
Source: Arena Agent

3. Artificial Analysis launches AA-AgentPerf, comparing "agents supported per megawatt" for the agent era

Event type: third-party benchmark + X first signal.
First/official/availability time: @ArtificialAnlys posted first AA-AgentPerf results in this window; the official article was updated around 2026-06-13.
Model/function name: AA-AgentPerf, initially covering DeepSeek V4 Pro across NVIDIA Blackwell / Hopper and AMD systems.
Capability change: AA-AgentPerf does not measure model intelligence. It uses real coding-agent trajectories to measure concurrent agent capacity, TTFT, output speed, and efficiency per megawatt. Artificial Analysis says the dataset includes multi-turn coding-agent trajectories, with sessions up to 200 turns and beyond 100K tokens.
Availability: Artificial Analysis AA-AgentPerf article, methodology, and hardware benchmark pages.
Impact: this is not a new model, but it is important for agent SaaS economics. As long-context, multi-turn tool use and KV-cache reuse become mainstream, whether a model can be served affordably and at stable concurrency will matter alongside intelligence scores.
Sources: @ArtificialAnlys, AA-AgentPerf announcement, AA-AgentPerf methodology, NVIDIA technical blog

4. XiaoHu Daily highlights Perplexity Search as Code: not a new model, but an agent retrieval architecture

Event type: official research + media / auxiliary source.
First/official/availability time: the Perplexity Research article is dated 2026-06-01; XiaoHu AI Daily's 2026-06-13 latest issue used it as the lead story.
Model/function name: Search as Code (SaC).
Capability change: Perplexity says SaC lets models generate Python code to orchestrate search primitives including retrieval, ranking, filtering, fan-out, and rendering. Its evaluation reports SaC at 0.871 on DSQA, 0.805 on BrowseComp, 0.651 on WideSearch, and 0.386 on WANDR, leading four of five benchmarks and essentially tying OpenAI Responses on HLE.
Availability: Perplexity Computer and Perplexity Agent API.
Impact: for model tracking, the point is that product performance from the same underlying model increasingly depends on tool, retrieval, and sandbox architecture. XiaoHu's "100% accuracy and 85% token-cost reduction" framing is a high-signal propagation point, but the exact CVE subtask and cost basis still need the original Perplexity or author context.
Sources: Perplexity Research, daily.xiaohu.ai

5. Video watchlist: Gemini Omni and HappyHorse remain high on the arena, with no new official rollout details

Event type: arena appearance.
First/official/availability time: Text-to-Video Arena is dated 2026-06-10, with 479,075 votes and 41 models.
Model/function name: gemini-omni-flash, happyhorse-1.0.
Capability change: gemini-omni-flash is No. 1 at 1527±14, with 2,649 votes and a preliminary label. happyhorse-1.0 is No. 3 at 1437±7, with 15,722 votes. No. 2 is dreamina-seedance-2.0-720p at 1466±8.
Availability: Arena Text-to-Video leaderboard.
Impact: these remain the two most important anonymous / semi-public video signals to track. Gemini Omni has a low vote count but a clear score lead. HappyHorse has more samples but weak official sourcing. No new Google / Alibaba / ATH model card, API entry point, or pricing note appeared today.
Source: Arena Text-to-Video

Arena Watch

LMArena Text Overall: dated 2026-06-10, with 6,820,793 votes and 366 models. The top five are claude-fable-5 1510±11, claude-opus-4-6-thinking 1504±4, claude-opus-4-7-thinking 1502±5, claude-opus-4-6 1498±4, and claude-opus-4-7 1492±5. Note that the No. 1 Fable 5 entry has only 2,883 votes and official access has been suspended.
Arena Agent: dated 2026-06-12, with 648,037 sessions and 25 models. The top five are Claude Fable 5 (High), GPT 5.5 (xHigh), Claude Opus 4.8 (Thinking), Claude Opus 4.7 (Thinking), and GPT 5.5 (High). Fable 5's Net Improvement is 13.68%±1.59%, and its sample size continues to grow.
Arena Text-to-Image: still dated 2026-06-05, with 5,391,418 votes and 70 models. gpt-image-2 (medium) is No. 1, 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. No new nano-banana movement was observed.
Arena Text-to-Video: dated 2026-06-10, with 479,075 votes and 41 models. gemini-omni-flash is No. 1, dreamina-seedance-2.0-720p No. 2, happyhorse-1.0 No. 3, veo-3.1-audio-1080p No. 4, and wan2.7-t2v No. 5. sora-2-pro is No. 7, runway-gen-4.5 No. 20, and kling-2.5-turbo-1080p No. 21.
Artificial Analysis: Fable 5 remains the central model-leaderboard signal. Artificial Analysis' article reports Claude Fable 5 at Intelligence Index 64.9, nearly 5 points ahead of the nearest non-Anthropic model, while its X account also says the Intelligence Frontier moved backward for the first time. Track both together with Anthropic's access suspension.
Sources: LMArena Text, Arena Agent, Arena Text-to-Image, Arena Text-to-Video, Artificial Analysis, Fable 5 on Artificial Analysis

X Watch

@AnthropicAI: the highest-spread visible post in this window is the Fable 5 / Mythos 5 access suspension statement, linking to Anthropic's official page. Credibility: official X plus official webpage, highest confidence.
@ClaudeDevs: explained developer-facing impact from the Fable 5 suspension and said Build Day is moving to Opus 4.8. The account also posted Claude Managed Agents self-hosted sandbox guides. Credibility: official developer account.
@ArtificialAnlys: within 22 hours, said the Intelligence Frontier moved backward for the first time; also released first AA-AgentPerf results covering DeepSeek V4 Pro on Blackwell / Hopper / AMD for agentic inference. Credibility: third-party benchmark source; methodology matters for interpretation.
@xiaohu / XiaoHu AI Daily: the latest site issue is 2026-06-13, led by Perplexity Search as Code. It also lists Fable 5 launch, open-source model clustering, U.S. export controls, Google Cloud open knowledge format, and Azure agent sandbox as themes. Credibility: auxiliary aggregation source; core facts were cross-checked against official and leaderboard sources.
@OpenAI: no new foundation model release appeared in this window. Visible activity focused on Codex rate-limit reset banking and invites. @OpenAIDevs launched a developers.openai.com docs agent that can generate custom guides from OpenAI documentation. Credibility: official product / developer update, not a new model.
@GoogleDeepMind / @GeminiApp: Google DeepMind focused on Robotics Accelerator, TacticAI, and Project Genie access expansion. Gemini App focused on a Neural Expressive design-language event and Gemini in Chrome regional expansion. Credibility: official product / research activity, not a Gemini Omni model card.
@Alibaba_Qwen: the latest high-signal item remains the 2026-06-02 pinned Qwen3.7-Plus release. No new Qwen model launch appeared in this window.
@MistralAI: the latest model / agent signal remains the late-May Mistral Vibe / Medium 3.5 line. No new model appeared in this window.
@xai: Grok Build Plugin Marketplace remains the main line, with MongoDB / Vercel / Sentry / Cloudflare / Chrome DevTools plugin activity. This is not a new Grok model.
@AIatMeta: Muse Spark remains the pinned model signal. No new model appeared in this window; visible activity focused on CVPR / SAM 3D paper propagation.
@cohere: after releasing North Mini Code, Cohere continued open-source / sovereign-AI messaging and used the Fable 5 suspension to emphasize control. Not a new model.
@runwayml: main activity was Runway AI Festival and expanded Lionsgate partnership. Aleph 2.0 remains the May pinned item, not a new release.
@LumaLabsAI: Dream Lab showed a June 13 Luma Ray3.2 example with multi-keyframe and expressive facial control. No new model card.
@Kling_ai: activity focused on its two-year anniversary, creator cases, and a 9-hour demo post. No new model card.

Pending Confirmation

Fable 5 / Mythos 5 suspension scope: Anthropic's official statement says foreign national access is barred, while the compliance effect is disabling access for all customers. Continue checking whether U.S.-based enterprises, employees, cloud platforms, Project Glasswing, and government / security partners have any exception path.
Fable 5 leaderboard treatment: whether LMArena / Artificial Analysis remove, freeze, or mark unavailable models will directly affect future comparability.
Fable 5 fallback behavior: confirm whether Claude products and API calls fully fall back to Opus 4.8 or differ by product line / cloud platform. Billing, cache, refusal, and fallback-credit behavior also need re-checking.
gemini-omni-flash: still missing an official Google model card, release time, API / product entry point, pricing, audio capability, and safety-boundary details.
happyhorse-1.0: still missing an Alibaba-ATH official launch page, first-release time, product entry point, and capability description. Treat it as an arena signal only.
Perplexity Search as Code: official research confirms the SaC architecture and aggregate benchmarks; the XiaoHu-propagated CVE 100% and 85% token-cost claims still need the original Perplexity or author table / post.
AA-AgentPerf: the metric is valuable, but the first wave mainly covers DeepSeek V4 Pro and a small set of hardware configurations. Watch for gpt-oss-120b, Qwen, Kimi, GLM, and more hardware combinations.
"25 open-source models across all modalities in one week": XiaoHu gives this a high-score theme, but this window did not break out the complete model list. Each item still needs verification for weights, license, leaderboard status, and official source.