AI Model Daily - 2026-06-10 | QName AI

30-Second Summary

Anthropic launched Claude Fable 5 and the restricted-access Claude Mythos 5. Fable 5 is Anthropic's strongest broadly available model today; Anthropic says it sets a new internal high bar for software engineering, knowledge work, vision, and scientific research. For sensitive cybersecurity, biology, chemistry, and similar requests, it can fall back to Claude Opus 4.8, with average fallback triggers below 5%.
Artificial Analysis now ranks Claude Fable 5 (with fallback) first on the LLM Intelligence Index, at roughly 65, ahead of Claude Opus 4.8 (max) at 61 and GPT-5.5 xhigh at 60. LMArena Text is still on its June 5 update, so it does not yet reflect Fable 5.
Google released Gemini 3.5 Live Translate, an audio model for real-time speech-to-speech translation. It supports 70+ languages, continuous low-latency translation, public preview in Gemini Live API / AI Studio, rollout in Google Translate, and private preview in Google Meet.
Cohere released its first open-source coding model, North Mini Code: a 30B total / 3B active MoE model under Apache 2.0, with 256K context and 64K output, aimed at agentic coding, terminal agents, and local deployment.
Arena Agent updated to 2026-06-10, with 535,121 sessions and 22 models. GPT 5.5 (High) remains No. 1 at 9.11%±1.26% Net Improvement, while Claude Opus 4.8 (Thinking) is No. 2 at 9.06%±1.57%.
On the video and creative-tool side, Luma API, Wan Fisheye Lens, xAI/Gopuff, and Kling all produced useful signals, but they are better treated as product/API updates or teasers, not independently confirmed new foundation model launches.

Must Watch

1. Claude Fable 5 / Claude Mythos 5: Anthropic's new broadly available and restricted model lines

Event type: official release + X first signal + leaderboard update.
First/official/availability time: @claudeai posted the official thread at 2026-06-10 01:08 HKT; Anthropic's official article is dated 2026-06-09. Anthropic says Claude Fable 5 is "available everywhere today", while Claude Mythos 5 initially goes to Project Glasswing partners in cyber defense and critical infrastructure.
Model/function name: Claude Fable 5, Claude Mythos 5; Anthropic docs list claude-fable-5 and invitation-only claude-mythos-5.
Capability change: Anthropic says Fable 5 exceeds all of its previous broadly released models and reaches state of the art on nearly all tested benchmarks, especially software engineering, knowledge work, vision, scientific research, and longer, more complex tasks. Fable 5 and Mythos 5 use the same underlying model. Fable 5 applies safeguards to sensitive cybersecurity, biology, chemistry, distillation, and related requests, falling back to Claude Opus 4.8 when needed; Anthropic says this happens in fewer than 5% of sessions on average. Mythos 5 lifts some safeguards in selected areas, but only for trusted programs.
Availability: Claude, Anthropic API, and Anthropic docs; Claude Mythos 5 has no normal public API entry yet and remains limited to Project Glasswing and later trusted access.
Price/specs: Anthropic lists Fable 5 pricing at $10/M input and $50/M output, less than half the prior Mythos Preview price. Anthropic's model page presents Fable 5 as its most capable broadly released model, with a 1M context tier.
Impact: this is the most important model event today. Fable 5 pushes Anthropic's general public model line back to the top of the current benchmark stack, while fallback turns high-risk capabilities into a "available by default, controlled when needed" product shape. The key follow-ups are whether fallback behavior is predictable for developers, whether long-task latency is acceptable, what the Mythos trusted-access standard is, and whether LMArena / real workflows reproduce the Artificial Analysis advantage.
Sources: Anthropic official article, Anthropic model docs, Claude Fable page, Claude Mythos page, Claude official X thread, Artificial Analysis analysis

2. Gemini 3.5 Live Translate: real-time speech translation enters the Gemini 3.5 line

Event type: official release + priority X-account first propagation.
First/official/availability time: Google's official article is dated 2026-06-10; @xiaohu posted a Chinese tracking thread at 2026-06-10 16:33 HKT.
Model/function name: Gemini 3.5 Live Translate; the developer docs list the model as gemini-3.5-live-translate-preview.
Capability change: Google describes the model as live speech-to-speech translation. It supports 70+ languages, automatically detects the input language, continuously generates speech in the target language, and aims to preserve the speaker's intonation, pacing, and pitch. Unlike turn-by-turn translation, it can keep translating while the speaker continues talking, staying only a few seconds behind, with stronger filtering for background noise and music.
Availability: developers can use it through Gemini Live API and Google AI Studio public preview; Google Translate for Android/iOS is rolling out Live translate with headphones, including a new Android earpiece listening mode; Google Meet has a private preview for selected Google Workspace enterprise customers, expanding from 5 languages to 70+ languages and 2000+ language pairs.
Impact: this is both an audio-model release and a consumer-product rollout. It matters directly for meetings, cross-border service, customer support, and travel. Google also says Grab is testing it for multilingual driver/passenger voice calls. The risk areas are latency, accent fidelity, noisy environments, privacy, and enterprise compliance. Developers should also note that the current model is audio-input-only, not a general multimodal conversation model.
Sources: Google official article, Gemini Live Translate docs, Gemini 3.5 Live Translate model page, @xiaohu X thread

3. Cohere North Mini Code: Cohere's first open-source agentic coding model

Event type: official release + open-source model availability + leaderboard/media interpretation.
First/official/availability time: @cohere posted the official thread at 2026-06-10 00:04 HKT; Cohere's official blog and docs went live in parallel.
Model/function name: North Mini Code / North-Mini-Code-1.0.
Capability change: Cohere positions it as its first open-source coding model and the first model in the North family of agentic coding models. It is a 30B total / 3B active parameter MoE model under Apache 2.0, with 256K context and 64K max output tokens, built for agentic software engineering, terminal-based agents, local/private deployment, and code generation. Cohere's X thread cites a 33.4 score on the Artificial Analysis Coding Index, competitive among similarly sized models.
Availability: Hugging Face CohereLabs/North-Mini-Code-1.0, Cohere docs, Cohere Model Vault, and free availability on OpenCode.
Impact: this is important for open, locally deployable coding agents. Cohere is not trying to beat frontier general models here; it is targeting coding-agent economics with a small-active-parameter MoE. If real terminal tasks hold up, it becomes a candidate for private enterprise coding agents, low-cost CI agents, and local developer assistants. Follow-up tests should cover SWE-bench, terminal tool use, long-repo context, license/compliance, and inference cost.
Sources: Cohere official blog, Cohere docs, Hugging Face model card, Hugging Face launch post, Cohere official X, Artificial Analysis analysis

4. Luma API video update: multi-keyframe, 20-second V2V, HDR/EXR

Event type: official X first signal + official API page update.
First/official/availability time: @LumaLabsAI started posting API update details at 2026-06-10 01:18 HKT.
Model/function name: Luma API / video generation workflow update.
Capability change: Luma says Multi-keyframe supports up to 16 keyframes inside a single clip, with sequence-level control exposed through the API. V2V supports up to 20 seconds, alongside 1080p output, native HDR generation, and 16-bit EXR export.
Availability: Luma API.
Impact: this is a video workflow upgrade, not a new named base video model. Its value is control and production integration: multi-keyframe, longer V2V, HDR, and EXR matter more for advertising, previsualization, and brand-template automation than one-off generation. Next checks: pricing, throughput, failure rate, commercial licensing, and asset-consistency constraints.
Sources: Luma API, Luma X multi-keyframe, Luma X V2V/HDR

5. xAI / Gopuff personalized shopping assistant: Grok multimodality enters instant retail

Event type: official product partnership.
First/official/availability time: @xai posted the Grok/Gopuff personalized shopping assistant partnership at 2026-06-10 03:14 HKT.
Model/function name: Grok shopping assistant for Gopuff.
Capability change: xAI describes the assistant as combining chat, voice, and image models for personalized shopping recommendations, discovery, and instant retail workflows.
Availability: Gopuff / xAI official news page.
Impact: this is not a new Grok model release, but it shows xAI embedding multimodal models into a transaction workflow. The follow-up question is whether it only handles recommendation/search, or whether it also supports image recognition, voice ordering, inventory constraints, substitute products, pre-payment confirmation, and safety boundaries.
Sources: xAI official X, xAI news page

Arena Watch

LMArena Text Overall: the page is still updated through 2026-06-05, with 6,703,075 votes and 365 models. The top remains unchanged: claude-opus-4-6-thinking at 1504±4 in first place, claude-opus-4-7-thinking at 1501±5 in second, claude-opus-4-6 at 1498±4 in third, and claude-opus-4-7 at 1493±5 in fourth. Conclusion: Claude Fable 5 is not visible there yet, so Artificial Analysis No. 1 should not be treated as LMArena No. 1.
Arena Agent: the page is updated to 2026-06-10, with 535,121 sessions and 22 models. GPT 5.5 (High) is No. 1 with Net Improvement 9.11%±1.26%; Claude Opus 4.8 (Thinking) is No. 2 at 9.06%±1.57%; Claude Opus 4.7 (Thinking) is No. 3 at 8.42%±1.24%; Claude Opus 4.6 is No. 4 at 8.15%±1.21%; GPT 5.4 (High) is No. 5 at 7.97%±1.33%. Conclusion: model count and sessions rose materially, and Opus 4.8 Thinking is now near the top of the second slot, but Fable 5 is not yet visible.
Arena Text-to-Image: the page remains at 2026-06-05, with 5,391,418 votes and 70 models. The top list remains gpt-image-2 (medium) at 1385±6 in first, reve-2.0 at 1273±9 in second, gemini-3.1-flash-image-preview (nano-banana-2) [web-search] at 1269±4 in third, mai-image-2.5 at 1253±6 in fourth, and gemini-3-pro-image-preview-2k (nano-banana-pro) at 1245±4 in fifth. Conclusion: no new image leaderboard top spot; nano-banana-2 remains in the top tier.
Artificial Analysis LLM: Claude Fable 5 (with fallback) is now first, with Intelligence Index around 65, 1M context, blended price around $8.20/M, output speed around 60 tokens/s, first-chunk latency around 108s, and total response latency around 116s. It is followed by Claude Opus 4.8 (max) at Index 61, GPT-5.5 (xhigh) at 60, GPT-5.5 high at 59, and Claude Opus 4.7 max at 57. Conclusion: Fable 5 is the new leader under the AA methodology, but latency is high enough that long-task and interactive use should be tested separately.
Artificial Analysis Coding: the standalone coding leaderboard page was not stable enough to read as a full ranking; Cohere and Artificial Analysis public analysis give North Mini Code a Coding Index of 33.4, describing it as competitive among similarly sized open models. Conclusion: North Mini Code belongs on the open coding-agent watchlist, but one index should not replace real repository-task testing.
Artificial Analysis STT Non-streaming: no independent STT/translation leaderboard score was visible for Gemini 3.5 Live Translate. Existing Non-streaming STT leaders still include Fun-Realtime-ASR-preview, ElevenLabs Scribe v2, MAI-Transcribe-1.5, MAI-Transcribe-1, Voxtral Small, and Gemini 3.1 Pro Preview. Conclusion: Live Translate is a translation model/product release, and STT WER should not be used as a direct substitute for translation quality.
XiaoHu AI Daily auxiliary source: the 2026-06-10 issue highlights Anthropic Fable 5 / Mythos 5, Cognition FrontierCode, Apple's Siri rebuild, Google Gemini 3.5 Live Translate, Cohere North Mini Code, and Responses API web-search image results. Difference from this report's scope: Apple, ChatGPT superapp, and Google GPU procurement are product, media, or industry signals; this report promotes only verified model, leaderboard, and availability items into the core section.
Sources: LMArena Text, Arena Agent, Arena Text-to-Image, Artificial Analysis LLM, Artificial Analysis Claude Fable analysis, Artificial Analysis North Mini Code analysis, XiaoHu AI Daily

X Watch

@claudeai: posted the official Claude Fable 5 / Claude Mythos 5 thread at 2026-06-10 01:08 HKT. Credibility: official release. Needs follow-up: actual availability across Claude subscription/API accounts, fallback behavior in practice, and Mythos trusted-access application path.
@xiaohu: posted a Chinese tracking thread for Gemini 3.5 Live Translate at 2026-06-10 16:33 HKT, summarizing 70+ languages, low-latency continuous translation, Google Translate / Meet / Grab testing, and linking official Google docs. Credibility: priority-account first propagation plus official-source verification.
@cohere: posted the official North Mini Code thread at 2026-06-10 00:04 HKT, confirming 30B/3B active, Apache 2.0, AA Coding Index 33.4, Hugging Face, and OpenCode availability. Credibility: official release.
@LumaLabsAI: posted Luma API updates from 2026-06-10 01:18 HKT, including 16 keyframes, V2V up to 20s, 1080p, HDR, and 16-bit EXR. Credibility: official product/API update.
@xai: posted the xAI/Gopuff personalized shopping assistant at 2026-06-10 03:14 HKT. Credibility: official product partnership. Needs follow-up: whether this uses any new Grok model, or only existing chat/voice/image models integrated into commerce.
@Alibaba_Wan: posted Wan Fisheye Lens at 2026-06-10 15:24 HKT, saying it can turn normal images into ultra-wide circular views. Credibility: official product feature. Needs follow-up: whether it uses a new model, whether video is supported, commercial licensing, and international availability.
@Kling_ai: posted an anniversary/new-capability teaser at 2026-06-10 22:00 HKT, pointing toward more realistic character motion and "she could speak" style video-character expression. Credibility: official-account teaser; not enough to treat as a full model launch. Needs follow-up: model name, entry point, pricing, and whether this is Kling 3.x or a new audio/video feature.
@GeminiApp: posted a Gemini 3.5 Flash visual-to-3D/interactive-code demo at 2026-06-10 00:13 HKT. Credibility: official product demo; not a new model launch.
@ArtificialAnlys: the core new signals in this window came from its site article and leaderboard: Claude Fable 5 taking the LLM Intelligence Index lead, and North Mini Code entering coding-model tracking. The X account was not used as a sole factual source.
@arena / LMArena-related accounts: no new Arena first signal was visible for Fable 5 or Gemini 3.5 Live Translate; official leaderboard pages are the source of record here.
@OpenAI / @OpenAIDevs / @AnthropicAI / @GoogleDeepMind / @MistralAI / @AIatMeta / @Alibaba_Qwen / @runwayml: no new verified base-model launch was visible in this window. OpenAI-related Responses API / web-search image result notes belong to developer-feature tracking, not today's model-launch list.

Pending Confirmation

Claude Fable 5 on LMArena: AA ranks it first, but LMArena Text/Agent does not yet show Fable 5. Wait for Arena availability, vote stability, and variant separation across with-fallback / no-fallback / thinking / non-thinking modes.
Claude Fable 5 fallback developer experience: Anthropic says fewer than 5% of sessions trigger fallback on average, but developers need to test which requests fall back, whether fallback is clearly visible in API responses, and whether this changes benchmark or real agent workflows.
Claude Mythos 5 trusted access: currently limited to Project Glasswing partners. Track application standards, audit requirements, regional availability, compliance limits, and whether commercial products can embed it.
Gemini 3.5 Live Translate real latency and language quality: Google claims a few seconds of lag and 70+ language support, but Mandarin/Cantonese/Japanese/English code-switching, heavy noise, interruptions, proper nouns, long meetings, headphone routing, and enterprise privacy all need practical testing.
Gemini 3.5 Live Translate leaderboard gap: no Artificial Analysis or LMArena speech-to-speech translation score was visible yet. STT WER should not be used as a direct proxy for translation quality.
North Mini Code real agentic coding performance: Cohere and AA cite Coding Index 33.4, but it still needs SWE-bench, terminal tasks, long-repo edits, test fixing, tool-call stability, inference cost, and local deployment throughput checks.
Luma API update and possible underlying video model change: multi-keyframe, 20-second V2V, and HDR/EXR are product capabilities. Confirm whether there is a new model ID, API migration, price change, or generation-stability change.
Wan Fisheye Lens: only the official X product feature was visible. Confirm global availability, batch/API support, video support, subject consistency, and license terms.
Kling June 10 teaser: the official account posted a strong teaser, but without model name, docs, entry point, or demo detail. Upgrade it to "official release" only after launch details land.
xAI/Gopuff: currently a vertical product partnership, not a new model. Watch for public voice/image model versions and whether developers can reuse similar shopping-agent capabilities through an API.