AI Model Daily - 2026-05-30 | QName AI

30-Second Summary

Google Cloud officially announced general availability for Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image). The enterprise entry point is Gemini Enterprise Agent Platform, and Nano Banana 2 also adds preview support for video files as input prompts.
xAI confirmed on X that grok-build-0.1 is now available through the xAI API public beta, with additional distribution through OpenRouter, Vercel AI Gateway, Cursor, Hermes Agent, OpenClaw, Kilo Code, and OpenCode. This is a key step in Grok Build moving from CLI/subscription distribution into the API ecosystem.
OpenAI did not release a new general flagship model today, but it did publish two agent/specialized-model signals: Codex 26.527 supports Windows Computer Use and Windows remote control, while Rosalind Biodefense expands GPT-Rosalind trusted access for biodefense and public-health use cases.
Anthropic/Claude continued to push Claude Opus 4.8 heavily on X: the original @claudeai post showed 13M views, with official emphasis on stronger judgment, more honest progress reporting, and longer independent work. The Dynamic Workflows research preview is today's most important developer-side agent harness change.
Leaderboards: Artificial Analysis LLM/GDPval continue to put Opus 4.8 in front; LMArena Text is still dated 2026-05-28 and does not show Opus 4.8; on image leaderboards, GPT Image 2 / GPT Image 1.5 lead, Nano Banana 2/Pro are in the top four, and grok-imagine-image-quality entered the image editing top five.

Must Watch

1. Google: Nano Banana 2 / Nano Banana Pro GA

Event type: official release + X first signal.
First/official time: the Google Cloud Blog page shows 2026-05-29; @Google posted on X the same day, and @GoogleDeepMind reposted it.
Capability change: Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) moved from preview into GA and can now be deployed for enterprise creative / agentic workflows. Google highlighted high-quality image generation and editing use cases and said the models are already part of workflows at Adobe Firefly Enterprise, WPP Open, Shopify, URBN, and Magnopus/OKO.
New capability: Nano Banana 2 now supports video files as input prompts in preview, using visual context, subjects, and motion from a video to generate context-aware images such as thumbnails and infographics.
Availability: Gemini Enterprise Agent Platform GA; developers can also use Gemini API, but the Google Cloud post explicitly separates Gemini API from the enterprise SLA. 1K and 2K output are GA, while 4K remains preview.
Impact: this is not the first appearance of the Nano Banana family. The important change is the move from preview/consumer tooling into enterprise production entry points. For brand assets, product images, ad variants, and video-to-thumbnail/infographic pipelines, GA plus SLA matters more than the model-name change alone. Key tests: video input prompt stability, Gemini API versus enterprise platform parameter/quota differences, and 4K preview quality/cost.
Sources: Google Cloud Blog, Google X, Google DeepMind X

2. xAI: grok-build-0.1 API Public Beta and Grok Build 0.2.7

Event type: official release + X first signal.
First/official time: xAI's News page shows 2026-05-28; @xai posted the X thread inside this collection window, with about 450K views.
Capability change: grok-build-0.1 is xAI's model for agentic coding, covering web development, debugging, MCP support, and related tasks. xAI says it is the same model used by the Grok Build CLI and can also serve as a fast, low-cost general agentic/tool-calling model.
API details: model name grok-build-0.1; xAI docs list text/image modalities, 256K context, function calling, structured outputs, reasoning, $1/M input, $2/M output, cached input at $0.20/M, RPM 1,800, TPM 10,000,000, regions us-east-1 / eu-west-1, and aliases including grok-code-fast-1, grok-code-fast, and grok-code-fast-1-0825.
Ecosystem entry points: xAI's X thread added availability through OpenRouter and Vercel AI Gateway, plus Cursor, Hermes Agent, OpenClaw, Kilo Code, and OpenCode.
Grok Build CLI: @xai also announced Grok Build 0.2.7 with /usage, /login, shared terminals across subagents, and improved image understanding.
Impact: this is a material pricing and distribution change for the coding-agent market. Unlike Claude Code Dynamic Workflows, which emphasizes hundreds of subagents, Grok Build currently emphasizes a low-cost, high-speed model plus CLI/API/third-party agent harness diffusion. Next tests should compare patch success rate, MCP call stability, long-context cost, and OpenRouter/Vercel Gateway routing consistency on the same repo task.
Sources: xAI Grok Build 0.1 on API, xAI model docs, Grok Build changelog, xAI X

3. OpenAI: Codex Windows Computer Use and GPT-Rosalind Trusted Access Expansion

Event type: official product update + specialized-model trusted access expansion.
Official time: the OpenAI Codex changelog shows 2026-05-29, and OpenAI's Rosalind Biodefense post was published the same day.
Codex capability change: in Codex app 26.527, Computer Use now works on Windows, so Codex can see, click, and type in foreground Windows desktop apps. Remote control also supports Windows devices, letting users start Codex work on a Windows device from ChatGPT iOS/Android or Mac and monitor progress remotely.
GPT-Rosalind / biodefense: OpenAI introduced Rosalind Biodefense for trusted developers building high-impact defensive life-science applications, and expanded GPT-Rosalind trusted access to select U.S. government and allied partners for public health / biodefense missions.
Availability: Codex changelog / Codex app; Rosalind Biodefense supports sponsorship or trusted access applications for eligible public-health, biodefense, research, and mission-driven teams.
Impact: Codex Windows expands the agent deployment surface, but it is not a new model. Rosalind Biodefense is also not a first launch of GPT-Rosalind; it is an expansion of trusted access and application scope. Together, they show OpenAI moving "specialized models + agent OS capability" into constrained environments: Windows enterprise desktops, public health, and defensive biosecurity.
Sources: Codex changelog, OpenAI Rosalind Biodefense

4. Anthropic: Opus 4.8 Day-One Diffusion and Dynamic Workflows

Event type: post-release official amplification + X first-signal tracking.
X timeline: on 2026-05-29, @ClaudeAI showed multiple Opus 4.8 posts: the original Opus 4.8 post had about 13M views; the Dynamic Workflows post had about 762K views; and the availability post had about 109K views. @AnthropicAI reposted those and also posted company financing/revenue updates.
Capability recap: Anthropic says Opus 4.8 improves over Opus 4.7 in coding, agentic skills, reasoning, and knowledge work; makes fewer unsupported progress claims; and was about 4x less likely to miss flaws in its own code. Anthropic also introduced effort control, Messages API system entries, and Claude Code Dynamic Workflows.
Dynamic Workflows: the research preview lets Claude plan work, run hundreds of parallel subagents, and verify outputs before reporting. It is available for Claude Code CLI, Desktop, VS Code extension, and Max / Team / Enterprise plans where enabled by admins; it can also be used through Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.
Impact: today's new value is day-one usage feedback and distribution intensity, not another model release. Developers should focus on testing Dynamic Workflows token cost, permission boundaries, subagent observability, and failure recovery instead of only watching the Opus 4.8 index.
Sources: Anthropic Claude Opus 4.8, Dynamic Workflows, Claude X, Anthropic X

5. Artificial Analysis: Opus 4.8 Still Leads, Image Leaderboards Shift

Event type: third-party leaderboard update + X first signal.
LLM: Artificial Analysis LLM Leaderboard continues to show Claude Opus 4.8 (max) at Intelligence Index 61, GPT-5.5 (xhigh) at 60, GPT-5.5 (high) at 59, and Claude Opus 4.7 (max) / Gemini 3.1 Pro Preview / Qwen3.7 Max at 57.
GDPval-AA: Claude Opus 4.8 (Adaptive Reasoning, Max Effort) has Elo 1890, ahead of GPT-5.5 (xhigh) at 1769 and GPT-5.5 (high) at 1755.
Image leaderboards: in Text to Image, GPT Image 2 (high) is first at Elo 1338, GPT Image 1.5 (high) is second at 1268, Nano Banana 2 Preview is third at 1260, and Nano Banana Pro is fourth at 1219. In Image Editing, GPT Image 1.5 (high) is first at 1260, GPT Image 2 (high) second at 1252, Nano Banana Pro third at 1241, Nano Banana 2 Preview fourth at 1240, and grok-imagine-image-quality fifth at 1229.
X signal: @ArtificialAnlys said on X that Opus 4.8 reached Intelligence Index 61.4, and that grok-imagine-image-quality entered the upper tier of both Text to Image and Image Editing, making it a strong image model outside OpenAI and Google.
Impact: today's leaderboard signal clarifies the multimodal production stack. Opus 4.8 leads on text/economic tasks; image generation is still led by OpenAI and Google/Nano Banana; xAI's image-quality version is entering the first tier, but still needs fresh comparison against the GA Nano Banana 2/Pro entries.
Sources: Artificial Analysis LLM, GDPval-AA, Text to Image Leaderboard, Image Editing Leaderboard, Artificial Analysis X

Arena Watch

LMArena Text Overall: the page remains dated 2026-05-28, with 6,532,755 votes and 360 models. The top entries are claude-opus-4-6-thinking Elo 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, gemini-3-pro 1486+/-4, and gpt-5.5-high 1482+/-6. Opus 4.8 is still not visible.
Artificial Analysis LLM: Claude Opus 4.8 (max) Index 61, cost $4.10/M tokens, median 58 tokens/s; GPT-5.5 (xhigh) Index 60, cost $4.35/M, median 55 tokens/s; GPT-5.5 (high) Index 59; Claude Opus 4.7 (max) Index 57; Gemini 3.1 Pro Preview Index 57, median 114 tokens/s; Qwen3.7 Max Index 57, cost $1.43/M, median 191 tokens/s; Gemini 3.5 Flash Index 55, median 176 tokens/s.
GDPval-AA: Opus 4.8 max Elo 1890, CI -34/+35; GPT-5.5 xhigh 1769; GPT-5.5 high 1755; Opus 4.7 max 1753; Gemini 3.5 Flash medium 1658, high 1656; Qwen3.7 Max 1548.
ITBench-AA: Opus 4.8 is not visible in this round; the latest visible results still show Opus 4.7 max 46.7%, GPT-5.5 xhigh 45.8%, and Qwen3.7 Max 42.5%. This remains key evidence that real enterprise SRE-agent tasks are far from saturated.
Text to Image: GPT Image 2 (high) is first at Elo 1338, GPT Image 1.5 (high) 1268, Nano Banana 2 Preview 1260, Nano Banana Pro 1219, and Recraft V4.1 Utility Pro 1204. Continue watching whether Nano Banana 2/Pro are updated as GA entries instead of preview-named entries.
Image Editing: GPT Image 1.5 (high) 1260 leads, followed by GPT Image 2 (high) 1252, Nano Banana Pro 1241, Nano Banana 2 Preview 1240, and grok-imagine-image-quality 1229.
Text to Video with audio: Dreamina Seedance 2.0 720p Elo 1214 is first, HappyHorse-1.0 1212 second, Kling 3.0 1080p (Pro) 1104, Kling 3.0 Omni 1080p (Pro) 1099, and Veo 3.1 1095.
Text to Video without audio: HappyHorse-1.0 Elo 1358 is first, Dreamina Seedance 2.0 720p 1272, Kling 3.0 1080p (Pro) 1250, Kling 3.0 Omni 1080p (Pro) 1235, and grok-imagine-video 1234. HappyHorse still shows API Coming soon, with no official model card.
Sources: LMArena Text, Artificial Analysis LLM, GDPval-AA, ITBench-AA, Text to Image, Image Editing, Text to Video

X Watch

@GoogleDeepMind: reposted @Google's Nano Banana 2 / Nano Banana Pro GA announcement, stating that developers and enterprises can use them through Google AI Studio and the Gemini Enterprise Agent Platform API.
@xai: within the last 6h, posted the original grok-build-0.1 API public beta announcement, emphasizing the same model as Grok Build CLI, agentic coding, $1/M input, and $2/M output; added OpenRouter, Vercel AI Gateway, Cursor, Hermes Agent, OpenClaw, Kilo Code, and OpenCode as entry points. It also announced Grok Build 0.2.7 with /usage, /login, shared terminals across subagents, and improved image understanding.
@ClaudeAI: the Opus 4.8 original post showed about 13M views; the Dynamic Workflows post about 762K views; and the "available today on the web, Claude Platform, and all major cloud platforms" post about 109K views. Distribution is strong, and the developer-side focus is workflow/agent behavior.
@AnthropicAI: reposted @ClaudeAI Opus 4.8 and pre-ship testing content; also posted Series H financing and run-rate revenue updates. Those are not model releases, but they are demand-side Claude signals.
@ArtificialAnlys: posted Opus 4.8 Intelligence Index 61.4 and GDPval-AA leadership; also said grok-imagine-image-quality entered the upper tier of Text to Image / Image Editing.
@xiaohu: in the last 12h, mostly reposted/interpreted iOS 27 / Siri rumors, saying Siri may get a standalone chat app, allow access to ChatGPT/Gemini/Claude and other models, and add a camera AI analysis mode. This is an unverified rumor / ecosystem-entry signal, not a model release.
@OpenAI: posted Rosalind Biodefense and Codex Windows Computer Use updates; no new general flagship model was visible.
@Alibaba_Qwen: still mainly amplifying Qwen3.7-Max ranking third on ITBench-AA and Qwen3.5 TokenSpeed at 580 tps; no new model visible in this window.
@MistralAI: still amplifying AI Now Summit, Mistral Vibe, and industry solutions; no new model first release visible in this window.
@LumaLabsAI: promoted Luma Agents content-to-promo workflows and BTS examples; no new model visible.
@Cohere, @runwayml, @Kling_ai, @AIatMeta: visible account pages did not show a new flagship model first release in the past 24 hours. Some account pages did not load their article/tweet areas reliably, so official sources and leaderboard sources were used as fallback.

To Confirm

Nano Banana 2 / Pro GA: confirm actual model IDs, pricing, rate limits, and SLA differences between Gemini Enterprise Agent Platform and Gemini API; also watch whether Artificial Analysis replaces Preview entries with formal GA entries.
Nano Banana 2 video input preview: officially confirmed as preview, but video length, file formats, cost, long-video support, and whether it reliably extracts character/action/camera semantics still need hands-on testing.
grok-build-0.1: API public beta is officially confirmed; test function calling, MCP, image input, structured outputs, Vercel/OpenRouter proxy consistency, and differences versus Grok Build CLI.
Grok Build 0.2.7: changelog items are confirmed on X, but shared terminals across subagents need local validation for shared-state contamination, permission boundaries, and log auditability.
Codex Windows Computer Use: confirmed in the official changelog; validate foreground operation, remote control, safety boundaries, and ChatGPT mobile connection stability on a real Windows machine.
GPT-Rosalind / Rosalind Biodefense: this is trusted access expansion, not a public API. Confirm eligibility, review flow, available model version, audit requirements, and safety restrictions.
Claude Dynamic Workflows: now in research preview, but subagent counts, cost, failure recovery, workflow script permissions, and Bedrock/Vertex/Foundry availability all need one real code-migration run.
Siri / iOS 27 multimodel access: currently from @xiaohu X rumors without an Apple official source. Keep it as an ecosystem-entry item only, not an official release.
HappyHorse-1.0: still strong on video leaderboards, especially first in no-audio T2V, but API remains Coming soon and there is no official model card.
Gemini Omni: no formal API availability or independent Artificial Analysis video leaderboard entry was visible in this round.

Next Steps

Treat Nano Banana 2 / Nano Banana Pro GA as a formal test for site asset-generation pipelines: product URL -> image brief -> 2K/4K output -> brand consistency -> cost/latency log.
Benchmark grok-build-0.1, Claude Opus 4.8, GPT-5.5 xhigh, and Qwen3.7 Max on the same coding-agent repo tasks, covering issue fixes, cross-file refactors, UI changes, test additions, and failure rollback.
Run one real Windows GUI workflow with Codex Windows Computer Use, focusing on auditability and remote-control safety boundaries.
Keep watching whether Artificial Analysis updates Nano Banana 2 / Pro from Preview to GA entries, and whether Opus 4.8 enters ITBench-AA / LMArena.