AI Model Daily - 2026-06-15 | QName AI

30-Second Summary

The most important model-library addition from the past 24 hours is Kimi-K2.7-Code: Moonshot/Kimi's official X account has announced and open-sourced it, the Hugging Face model card confirms Modified MIT, 1T MoE, 32B active parameters, 256K context, and MoonViT vision encoder, and Cloudflare Workers AI added @cf/moonshotai/kimi-k2.7-code on June 12.
Gemini-SQL2 is today's strongest specialized-capability signal: Google Research announced on X that it is powered by Gemini 3.1 Pro and ranks No. 1 on BIRD Single Trained Model Track with Test EX 80.04%; for now it looks like a closed text-to-SQL system/capability, not a new general Gemini model.
XiaoHu AI Daily amplified SkillOpt: Microsoft's project page says it trains Markdown skill files to improve frozen target-model agent performance, with average gains of +23.5 for GPT-5.5 direct chat, +21.8 for Codex harness, and +18.6 for Claude Code harness. This is an agent-skill optimization method, not a new model.
Anthropic Fable 5 / Mythos 5 availability remains a high-priority risk: @AnthropicAI and @ClaudeDevs still show access suspension and Build Day switching to Opus 4.8; Fable 5 remains strong on LMArena and Artificial Analysis, but production availability cannot be inferred from leaderboard position.
Arena leaderboards did not show a new date jump: Text remains Jun 11, 2026, Agent remains Jun 12, 2026, Image remains Jun 5, 2026, and Video remains Jun 10, 2026; Agent sessions increased to 670,363, with Fable 5 High still No. 1.

Must Watch

1. Kimi-K2.7-Code: Moonshot releases an open coding/agent model, now available on Workers AI

Event type: official release + X first signal + availability.
First/official/availability time: Kimi's official X release was on 2026-06-12; Cloudflare Workers AI's changelog added it on the same day; XiaoHu AI Daily's 2026-06-14 issue listed it as a high-scoring item.
Model/function name: Kimi-K2.7-Code / kimi-k2.7-code.
Capability change: Kimi's official X account says it improves over K2.6 by +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite, while reducing reasoning-token use by about 30%. The Hugging Face model card describes it as a K2.6-based coding-focused agentic model with MoE architecture, 1T total parameters, 32B activated parameters, 256K context, MoonViT vision encoder, image/video input, forced thinking / preserve thinking, and multi-step tool calls.
Availability: Hugging Face moonshotai/Kimi-K2.7-Code, Moonshot/Kimi API, Cloudflare Workers AI @cf/moonshotai/kimi-k2.7-code, Fireworks and other third-party inference providers, plus vLLM / SGLang / KTransformers deployment.
Impact: this is currently the most important new open coding/agent model to track. Compared with closed Fable/GPT/Opus options, Kimi's key advantages are open weights, self-hosting, and lower inference cost. However, the benchmark claims are first-party, and the model card shows Kimi K2.7 Code still trails GPT-5.5 / Opus 4.8 on some agentic metrics, so independent benchmarks are still needed.
Sources: @Kimi_Moonshot, Hugging Face model card, Cloudflare Workers AI changelog

2. Gemini-SQL2: Google Research text-to-SQL capability reaches No. 1 on BIRD

Event type: X first signal + public leaderboard / benchmark.
First/official/availability time: Google Research announced it on X on 2026-06-13; the BIRD leaderboard entry is dated 2026-06-03.
Model/function name: Gemini-SQL2, a text-to-SQL capability powered by Gemini 3.1 Pro.
Capability change: BIRD Single Trained Model Track lists Gemini-SQL2 as a Google Research & Cloud submission, size UNK, self-consistency Many, Dev EX 74.12%, and Test EX 80.04%; human performance is 92.96%.
Availability: currently confirmed through Google Research X and the BIRD leaderboard only; no standalone model card, public API, weights, or pricing were found.
Impact: this is an important specialized capability for database, BI, and enterprise-data agents, but it should not be treated as a general foundation-model release. It shows Gemini 3.1 Pro can reach SOTA through a vertical system/training path; product teams should watch whether Google Cloud, Vertex, or BigQuery exposes Gemini-SQL2 as a callable capability.
Sources: @GoogleResearch, BIRD leaderboard

3. SkillOpt: Microsoft turns Markdown skills into trainable external agent state

Event type: official research / open-source project + media report.
First/official/availability time: the Microsoft SkillOpt project page and GitHub repo are public; XiaoHu AI Daily's 2026-06-14 issue amplified it as a high-scoring item.
Model/function name: SkillOpt: Executive Strategy for Self-Evolving Agent Skills.
Capability change: SkillOpt uses an optimizer model to make bounded edits to Markdown skills, then selects updates through a held-out validation gate. Microsoft's project page reports average gains of +23.5 for GPT-5.5 direct chat, +21.8 for GPT-5.5 Codex harness, and +18.6 for GPT-5.5 Claude Code harness. Evaluations cover SearchQA, Spreadsheet, Office, DocVQA, LiveMath, ALFWorld, and related tasks.
Availability: Microsoft GitHub Pages, GitHub repo, and arXiv preprint.
Impact: this is not a new model, but it is very practical for agent products. Many workflows may not need fine-tuning or a model switch; they can encode working knowledge into skill files that are trainable, verifiable, and portable. Track whether this pattern enters mainstream Codex, Claude Code, Cursor, and other agent workflows.
Sources: SkillOpt project page, GitHub repo

4. Fable 5 / Mythos 5: availability risk remains unresolved while leaderboard signals stay strong

Event type: official release + X first signal + availability change + leaderboard watch.
First/official/availability time: Anthropic's access-suspension statement spread on X on 2026-06-13; the official statement page remains available.
Model/function name: Claude Fable 5, Claude Mythos 5.
Capability change: LMArena Text still shows claude-fable-5 at No. 1 with 1510±11; Agent Arena still shows Claude Fable 5 (High) at No. 1 with Net Improvement 13.68%±1.59%. Artificial Analysis still records Fable 5 at Intelligence Index 64.9.
Availability: official access to Fable 5 / Mythos 5 remains suspended; ClaudeDevs says new sessions run on the default model or Opus 4.8, and Build Day switched to Opus 4.8.
Impact: Fable 5 should still be labeled "strong capability, high production-availability risk." Any model-selection, API-routing, or benchmark-reproduction work needs explicit fallback and availability handling.
Sources: Anthropic statement, @AnthropicAI, @ClaudeDevs, Artificial Analysis Fable 5

5. AA-AgentPerf: agentic inference benchmark keeps spreading, bringing hardware capacity into model tracking

Event type: third-party benchmark + X first signal.
First/official/availability time: the Artificial Analysis article is dated 2026-06-12; X continued spreading it on 2026-06-13.
Model/function name: AA-AgentPerf.
Capability change: AA-AgentPerf replays real coding-agent trajectories and measures how many concurrent agents a system can support under production serving targets, with Agents per Megawatt as a core metric. The first results cover NVIDIA Blackwell/Hopper and AMD systems running DeepSeek V4 Pro; Artificial Analysis says future coverage will expand to models such as gpt-oss-120b.
Availability: Artificial Analysis AA-AgentPerf article, hardware benchmark, and methodology.
Impact: this is not a model release, but it changes agent model selection. In the agent era, "how many long-context agents this model can run on a given inference platform" directly affects cost and user experience.
Sources: AA-AgentPerf, @ArtificialAnlys

Arena Watch

LMArena Text Overall: still dated Jun 11, 2026, 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. gpt-5.5-high is No. 10, qwen3.7-max-preview No. 16, and kimi-k2.6 No. 31; Kimi-K2.7-Code was not visible near the top of the Text overall table.
Arena Agent: still dated Jun 12, 2026, with sessions up to 670,363 and 25 models. Claude Fable 5 (High) is No. 1, Net Improvement 13.68%±1.59%, sessions 16,262; GPT 5.5 (xHigh) is No. 2, sessions 8,434; Claude Opus 4.8 (Thinking) is No. 3, sessions 25,745. The leaderboard date did not change, but sample size kept growing.
Arena Text-to-Image: still dated Jun 5, 2026, 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: still dated Jun 10, 2026, with 479,075 votes and 41 models. gemini-omni-flash is No. 1, dreamina-seedance-2.0-720p No. 2, and happyhorse-1.0 No. 3; 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 strongest recorded Intelligence Index signal; AA-AgentPerf is the new hardware / agentic-inference signal. No new Artificial Analysis model-ranking date jump was observed today.
Sources: LMArena Text, Arena Agent, Arena Text-to-Image, Arena Text-to-Video, Artificial Analysis

X Watch

@Kimi_Moonshot: pinned Kimi-K2.7-Code, saying it is open source and showing coding/agent benchmark gains over K2.6 plus a 30% reasoning-token reduction; it later posted an API recharge bonus. Credibility: official account.
@GoogleResearch: posted Gemini-SQL2, saying it is powered by Gemini 3.1 Pro and achieved SOTA on BIRD. Credibility: official X plus BIRD leaderboard cross-check.
@ArtificialAnlys: continued spreading the first backward move of the Intelligence Frontier and the first AA-AgentPerf results. Credibility: third-party benchmark source.
@AnthropicAI / @ClaudeDevs: the Fable 5 / Mythos 5 access suspension remains the highest-priority signal; ClaudeDevs also highlighted Managed Agents self-hosted sandbox guides. Credibility: official accounts.
@xiaohu: the latest site issue is 2026-06-14, led by Fable/Mythos 5, with Apple Siri AI, Gemini-SQL2, SkillOpt, Fable 5 FrontierMath, and Kimi K2.7 Code among the key items. Credibility: auxiliary aggregation source; facts need official or leaderboard verification.
@OpenAI: no new model release in this window; the official main account focused on Codex rate-limit reset bank and invites, while @OpenAIDevs spread Codex mobile/community and docs-agent updates. Credibility: official product updates, not a model first signal.
@GoogleDeepMind / @GeminiApp: DeepMind stayed focused on Robotics Accelerator, TacticAI, and Project Genie access expansion; Gemini App had a recent F1 interaction post and earlier small-business / Gemini in Chrome feature expansion. No official Gemini Omni model card was observed.
@Alibaba_Qwen: the latest high-signal item remains Qwen3.7-Plus from June 2; no new Qwen model first signal appeared in this window.
@MistralAI: no new model; the latest high-signal line remains Mistral Vibe / Medium 3.5.
@xai: Grok Build added terminal rendering for math, formulas, and LaTeX within the past few hours; this is a product capability, not a new Grok model.
@AIatMeta: Muse Spark remains the pinned model signal; no new model appeared.
@cohere: continued posting around North Mini Code and AI sovereignty, using the Fable 5 suspension to emphasize controllable open/private deployment; not a new model.
@runwayml, @LumaLabsAI, @Kling_ai: mostly events, partnerships, examples, and demos; no new video model card or API release was observed.

To Confirm

Kimi-K2.7-Code independent evaluation: the official/HF benchmark and Cloudflare availability are confirmed, but LMArena, Artificial Analysis, SWE, and agentic-coding third-party benchmarks are still needed, especially for fair cost/latency comparisons against GPT-5.5 xHigh, Opus 4.8, and Fable 5.
Kimi-K2.7-Code license and commercial boundaries: Hugging Face shows Modified MIT, but production deployment still needs a full LICENSE review, third-party notices, Moonshot API terms, and hosting-platform terms.
Gemini-SQL2 availability: currently only Google Research X and the BIRD leaderboard are confirmed. There is no model card, paper, API, product integration, or pricing; watch BigQuery, Cloud SQL, and Vertex AI.
SkillOpt reproducibility: inspect the GitHub repo's training scripts, benchmark splits, migration path into the team's existing Codex / Claude Code skill system, and overfitting risks.
Fable 5 / Mythos 5 suspension scope: whether U.S. enterprise, government, safety-research, or Project Glasswing exceptions exist remains unclear; leaderboards also have no unified labeling for unavailable models.
gemini-omni-flash and happyhorse-1.0: still missing official model cards, first-release times, API/product entry points, pricing, and whether they include audio or editing capabilities.
XiaoHu AI Daily's Apple Siri AI, Fable 5 FrontierMath, Google Cloud Open Knowledge Format, Azure sandbox, and related items are partly product/tool/infrastructure updates rather than model releases; keep them separately classified to avoid mixing them into foundation-model launch tracking.