AI Model Daily - 2026-06-05 | QName AI

30-Second Summary

Today's most important model release is NVIDIA Nemotron 3 Ultra: NVIDIA's official X post describes it as a 550B MoE open model for long-running agents, with up to 5x faster inference and up to 30% lower cost on complex agentic tasks. Artificial Analysis gives it an Intelligence Index of 47.7 and calls it the strongest US open-weights model.
Arena officially launched Agent Mode / Agent Arena, moving evaluation from chat battles into real agent tasks. Web search, filesystem access, terminal tools, sandboxed bash, and image generation/analysis are now part of the Arena evaluation surface. This is not a single model release, but it changes how future model rankings should be watched.
OpenAI announced and began rolling out a stronger ChatGPT memory system. It is starting with US Plus / Pro users, doubles memory capacity to 2x, and adds more visibility and control through a memory summary. This is a product capability update, not a new base model.
Anthropic published a recursive self-improvement research article and X thread, arguing that Claude is already accelerating internal AI R&D. It disclosed that Mythos Preview reached about 52x speedup on some evaluations. This is an important research signal, but Mythos Preview is not publicly available.
Other signals worth tracking: StepFun Step 3.7 Flash reached the Artificial Analysis speed/intelligence Pareto frontier with Index 43 and 400+ output tokens/s; @xiaohu mentioned an Ideogram 4.0 open image model, but the official source chain still needs verification; Gemini Omni and Runway Aleph 2.0 are product entry/workflow updates.

Must Watch

1. NVIDIA: Nemotron 3 Ultra Released, Bringing Open-Weights Agentic LLMs Into Focus

Event type: official release + X first signal + third-party leaderboard evaluation.
First/official time: @NVIDIAAI posted at 2026-06-04 21:06 HKT; @ArtificialAnlys posted the leaderboard readout at 2026-06-04 21:32 HKT and added a Terminal-Bench v2.1 low-latency agentic performance signal at 2026-06-05 02:12 HKT.
Capability change: NVIDIA says Nemotron 3 Ultra is a 550B MoE frontier-intelligence open model for long-running agents, delivering up to 5x faster inference and up to 30% lower complex agentic task cost versus other open frontier models. Artificial Analysis says it scores 47.7 on the Intelligence Index, ahead of the next US open-weights tier, and highlights its faster task pace on restricted-turn Terminal-Bench v2.1.
Availability: NVIDIA Research, Hugging Face nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16, and the Artificial Analysis model page. X circulation also mentions hosted availability through GMI Cloud, Ollama Cloud, and similar providers. Exact license terms, context window, active parameters, and deployment requirements should still be checked against the model card.
Impact: this is today's only general LLM release that should immediately enter evaluation pools. Its importance is not just raw intelligence, but the combination of open weights, long-running agent work, low latency, and cost curve. Priority tests: Terminal-Bench, SWE-like agents, long-context tool use, code repair pipelines, and enterprise internal agent tasks.
Sources: NVIDIA AI X, NVIDIA Research, Hugging Face, Artificial Analysis X, Artificial Analysis model page, Terminal-Bench signal

2. Arena: Agent Mode / Agent Arena Launched, Moving Agent Evaluation Into the Main Lane

Event type: Arena product launch + evaluation-system update + X first signal.
First/official time: @arena announced Agent Mode at 2026-06-05 00:00 HKT and Agent Arena at 00:06 HKT; at 05:30 HKT, it also announced that Battle Mode had passed 50 million votes.
Capability change: Arena says Agent Mode can perform deep research, create reports, generate images, build websites, debug code, and handle more complex work using tools such as web search, filesystem, terminal, sandboxed bash, and image generation/analysis. Agent Arena aims to evaluate agents at scale using live sessions where real users complete real tasks.
Availability: Arena Agent Mode / Agent Arena pages and the original X posts. After 2026-06-05 07:30 HKT, Arena also posted that Nemotron 3 Ultra had joined Agent Mode, but that post is outside this report's strict window and should be tracked in the next issue.
Impact: this is more structurally important than a single leaderboard item. Future "best model" claims should not be read only through chat Elo, code Elo, or static benchmarks; tool access, task length, recovery from failure, browser/terminal behavior, and real user task completion rates now matter.
Sources: Agent Mode X, Agent Arena X, Arena Agent Leaderboard

3. OpenAI: ChatGPT Memory System Upgrade Rolls Out, Affecting Personal Agent Experience

Event type: official product update.
First/official time: @OpenAI posted at 2026-06-05 00:10 HKT, linking to the official chatgpt-memory-dreaming article.
Capability change: OpenAI says the new ChatGPT memory system should make cross-conversation context more useful and durable, and adds a memory summary so users can see what ChatGPT remembers. X and follow-up circulation indicate the rollout starts with US Plus / Pro users and increases memory capacity to 2x.
Availability: ChatGPT memory controls in settings and the official article. No corresponding new model ID was found in API models, Responses API, or Assistants/Agents SDK materials.
Impact: this is not a model launch, but it matters for personal assistants, long-running project context, Codex-like workflows, and how memory can be audited or edited. Watch whether Team, Enterprise, EDU, or API-side memory controls follow.
Sources: OpenAI X, OpenAI official article

4. Anthropic: Recursive Self-Improvement Research Published, Mythos Preview Remains an Internal Signal

Event type: official research article + X thread.
First/official time: @AnthropicAI posted at 2026-06-05 00:15 HKT; the official destination is Anthropic Institute's recursive-self-improvement page.
Capability change: Anthropic says internal data shows Claude is accelerating AI development, potentially pointing toward AI autonomously building more capable successors. The thread highlights two evaluations: in a same-test task where models train a small AI model, Claude Opus 4 averaged about 3x speedup in 2025-05, while Mythos Preview reached about 52x in 2026-04; in a wrong-turn research-session evaluation, Mythos Preview improved human research direction 64% of the time, versus 22% for 2024 models.
Availability: official article and X thread. No public API, Claude app, Bedrock, or Vertex access was found for Mythos Preview.
Impact: this is not a callable model release, but it belongs in the research-warning lane. Anthropic is quantifying how models accelerate AI research itself, which may shape agentic research, AI lab productivity, automated experiment platforms, and safety governance narratives.
Sources: Anthropic X, Anthropic Institute, correction post

5. StepFun: Step 3.7 Flash Becomes Competitive on the Speed/Intelligence Curve

Event type: third-party leaderboard update + open model card verification.
First/official time: @ArtificialAnlys posted at 2026-06-04 11:48 HKT; the Hugging Face model card is the official StepFun open-weights entry.
Capability change: Artificial Analysis says Step 3.7 Flash sits on the Intelligence vs Output Speed Pareto frontier, scoring 43 on the Intelligence Index and serving over 400 output tokens/s. It is a significant upgrade over Step 3.5 Flash. The model is open weights and marked Apache 2.0.
Availability: Artificial Analysis model page and Hugging Face stepfun-ai/Step-3.7-Flash.
Impact: it is not today's highest-discussion model, but it deserves testing for low-latency agents, Chinese/English mixed business workloads, batch automation, and high-throughput low-cost inference. It should enter the low-latency agent and API replacement pool.
Sources: Artificial Analysis X, Artificial Analysis model page, Hugging Face

Arena Watch

LMArena Text Overall: the public page still shows a 2026-05-28 update, 6,532,755 votes, and 360 models. The top entries remain Claude Opus 4 series plus Gemini/Qwen/GPT-5.5: claude-opus-4-6-thinking 1502+/-4, claude-opus-4-7-thinking 1500+/-5, claude-opus-4-6 1498+/-4, claude-opus-4-7 1494+/-5, muse-spark 1489+/-6, gemini-3.1-pro-preview 1487+/-4, and gpt-5.5-high 1482+/-6. No new top-board entry changed the pattern today.
Arena Agent: launched today. The key signal is not stable Elo yet, but the shift from chat to tool-based tasks involving web search, filesystem, terminal, sandboxed bash, and image generation/analysis. The next issues should watch for public ranking, model list, and methodology details.
Arena Battle Mode: Arena says it has passed 50 million votes, showing continued accumulation of real preference data. Votes, task type, and tool access should still be separated by mode.
Artificial Analysis LLM: the new focus is NVIDIA Nemotron 3 Ultra, with Index 47.7, labeled the strongest US open-weights model. Step 3.7 Flash has Index 43 and 400+ output tokens/s, putting it on the speed/intelligence Pareto frontier. The closed-model top tier remains the previous Claude Opus 4.8 / GPT-5.5 / Gemini 3.1 / Qwen3.7 pattern.
Artificial Analysis Terminal-Bench v2.1: today's signal says Nemotron 3 Ultra completes tasks faster under restricted turn usage, but the exact table was not captured in this pass. The next report should add exact score, completion rate, latency, and cost.
XiaoHu AI Daily auxiliary source: the latest public web issue is 2026-06-03, led by NVIDIA Cosmos 3, and its highlights include NVIDIA Nemotron 3 Ultra. Because it is earlier than the later part of this report's window, it is used only as an auxiliary source rather than overriding later NVIDIA / Artificial Analysis / Arena posts.
Sources: LMArena Text, Arena Agent, Arena Agent Mode X, Artificial Analysis Nemotron, Artificial Analysis Step, XiaoHu AI Daily

X Watch

@NVIDIAAI: released Nemotron 3 Ultra, describing it as a 550B MoE open model for long-running agents and emphasizing 5x faster inference and 30% lower cost. This is today's highest-priority official release.
@ArtificialAnlys: posted two model signals. First, Nemotron 3 Ultra, with Index 47.7, strongest US open-weights positioning, and a Terminal-Bench v2.1 low-latency agentic performance signal. Second, Step 3.7 Flash, with Index 43 and 400+ output tokens/s.
@arena: launched Agent Mode / Agent Arena and said Battle Mode passed 50 million votes. Agent Mode will directly affect how we rank whether a model can actually do work.
@OpenAI: announced the ChatGPT memory system upgrade, described as more capable cross-conversation memory with a more transparent memory summary. No new model or API model ID was visible.
@AnthropicAI: published recursive self-improvement research and later corrected that Claude Opus 4's about 3x speedup refers to 2025-05, not 2024-05. Mythos Preview remains an internal/research model signal.
@GeminiApp: posted a Gemini Omni "Create videos" product-entry demo supporting text, video, or up to five images as input; it also posted a macOS Gemini app screen-aware feature using the double Command key. Neither is a base-model release.
@runwayml: posted an Aleph 2.0 Edit Studio update, emphasizing targeted edits that leave the rest of the shot untouched. This is a video-editing workflow update; keep watching for a model card or API signal.
@xiaohu: posted an Ideogram 4.0 open image model lead mentioning bounding-box control and structured JSON prompts; another post summarized how Anthropic uses Claude for automated business analysis. The Ideogram item still needs official model card, repository, and license verification.
@xai / @grok: no new main-account model release was found in this window; Grok Imagine 1.5 remains a previous-days carryover.
@Alibaba_Qwen / @Alibaba_Wan: no new model release was found in this window; Qwen3.7, Wan2.7, and WanMuse+ remain previous-days signals.
@Cohere: window activity centered on a NATO Agentic AI challenge and Canadian AI strategy content, not a model release.
@GoogleDeepMind / @GoogleAI / @AIatMeta / @MistralAI / @LumaLabsAI / @Kling_ai: no new verifiable model first release was found in this window.

To Confirm

Nemotron 3 Ultra: add model-card-level details, including license terms, context window, active parameters, recommended inference stack, quantized versions, commercial restrictions, and actual availability through GMI/Ollama/Prime-style hosted entry points.
Nemotron 3 Ultra Terminal-Bench v2.1: Artificial Analysis has posted the low-latency agentic performance signal, but exact score, average latency, completion rate, and cost curve still need a full leaderboard or article.
Agent Arena: confirm the stable public leaderboard URL, model list, tool permissions, sandbox behavior, task samples, vote methodology, and whether different agent models receive identical tool budgets.
OpenAI memory system: confirm rollout beyond US Plus/Pro, Team/Enterprise/EDU policy, any API-side equivalent memory controls, and the full UI for deleting or auditing memory.
Anthropic Mythos Preview: no public entry point was found. Treat it as a research capability signal, not a Claude model release.
Step 3.7 Flash: test Chinese/English quality, function/tool calling, long context, code ability, and true throughput cost before treating it as a low-latency replacement.
Ideogram 4.0: the @xiaohu lead is worth tracking, but this pass only verified official technical/product pages and community circulation. The formal model card, repository, license, and download entry still need confirmation.
Gemini Omni: consumer-side video creation keeps spreading, but Gemini API / AI Studio / Vertex / Enterprise developer access, pricing, and geography remain unclear.
Runway Aleph 2.0: Edit Studio is available, but whether it maps to a new model version, whether there is API access, and how it relates to Gen-4 / Act-Two still need verification.