AI Model Daily - 2026-06-23 | QName AI

30-Second Summary

The only signal in the past 24 hours that qualifies as an official release with a model-capability change is OpenAI Daybreak: OpenAI announced a security tooling suite and continues to make the full version of GPT-5.5-Cyber available on a limited basis to trusted defenders. This is an important security-specialized model and Codex Security update, not a general API rollout.
XiaoHu AI Daily 2026-06-22 put SkillOpt in the headline slot. After cross-checking, this is a Microsoft Research agent-skill optimization framework first published in May 2026. Today it should be classified as media/daily amplification, not a new model launch, but it is worth tracking for long-running agent engineering.
Google/Gemini's new signal is Gemini Spark beta availability: @GeminiApp says it is starting beta for Google AI Ultra subscribers in the U.S.; the official page says it runs on Gemini 3.5 Flash and Antigravity. This is product/agent availability progress, not a new foundation model.
The Arena item to watch today is seed-2.1-pro-preview, which entered the WebDev top 10 at No. 8 and was highlighted by Arena's official X account. The video boards are still led by gemini-omni-flash, dreamina-seedance-2.0-720p, and happyhorse-1.0.
Price Per Token's model releases page shows Last 24h 2: OpenAI Daybreak is the official model-related update, while the other item is media coverage of Anthropic's government dispute. There was no new general-purpose LLM/API model ID launch today.

Must Watch

1. OpenAI Daybreak: full GPT-5.5-Cyber remains in limited release

Event type: official release + vertical model capability update + security product update.
First/official/availability time: OpenAI's official article is dated 2026-06-22; @OpenAI posted the Daybreak thread at 2026-06-22T17:04Z.
Model/function name: GPT-5.5-Cyber, Codex Security, Daybreak Cyber Partner Program, Patch the Planet.
Capability change: OpenAI says the full version of GPT-5.5-Cyber continues advancing authorized defensive work, moving from vulnerability discovery toward patch automation. OpenAI reports a CyberGym score of 85.6%, above GPT-5.5 at 81.8%. The page also gives ExploitGym 39.5% vs 25.95% and SEC-bench Pro 69.8% vs 63.1%.
Availability: still limited to trusted defenders. OpenAI says most defenders should use GPT-5.5 Trusted Access and Codex Security. Patch the Planet works with Trail of Bits, HackerOne, Calif, and open-source maintainers; initial participating projects include cURL, Go, Python, Sigstore, and pyca/cryptography.
Impact: this is today's clearest model-capability news, but it is not a consumer or general developer API update. Track three follow-ups: whether GPT-5.5-Cyber access expands, whether Codex Security becomes part of normal Codex workflows, and whether OpenAI standardizes patch generation plus human review as an enterprise security product.
Sources: OpenAI Daybreak, OpenAI Patch the Planet, @OpenAI Daybreak, @OpenAI Patch the Planet, Price Per Token model releases

2. SkillOpt: XiaoHu headline, but the official first release was in May 2026

Event type: media coverage + official research project verification.
First/official/availability time: the Microsoft Research page is dated 2026-05; XiaoHu AI Daily put it in the headline slot on 2026-06-22.
Model/function name: SkillOpt.
Capability change: SkillOpt is not a new model. It is a text-space optimizer for agent skill documents. The loop runs a task, scores the result, analyzes failures, applies bounded add/delete/replace edits to the skill document, and uses a validation gate to filter regressions. Microsoft Research says it covered 52 combinations across 6 benchmarks, 7 target models, and 3 harnesses, and reached best/tied on all combinations. For GPT-5.5, the official publication page reports +23.5 in direct chat, +24.8 in the Codex agentic loop, and +19.1 in Claude Code.
Availability: Microsoft Research publication page and SkillOpt project page. At this stage it is more research/engineering methodology than a user-facing product entry point.
Impact: it should not be written as today's new model release, but it belongs in the agent-engineering tracking pool. It turns prompt/skill tuning from manual craft into an evaluated, rollback-friendly, cross-model optimization loop, directly relevant to Codex, Claude Code, Gemini Antigravity, and similar long-running agent workflows.
Sources: XiaoHu AI Daily 2026-06-22, Microsoft Research SkillOpt, SkillOpt project page

3. Gemini Spark: beta starts for U.S. Google AI Ultra users

Event type: official X first signal / product availability + official page verification.
First/official/availability time: @GeminiApp posted at 2026-06-22T19:38Z that Gemini Spark is starting beta for Google AI Ultra subscribers in the U.S.; the official page says it is rolling out to trusted testers and will expand over the coming weeks to U.S. Google AI Ultra users over 18 and select business users.
Model/function name: Gemini Spark, Gemini 3.5 Flash, Antigravity.
Capability change: Gemini Spark is described as a 24/7 personal AI agent that can connect to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, Maps, and other Google services. Connections are off by default. Official examples include project tracker auditing and newsletter aggregation as persistent tasks.
Availability: beta for U.S. Google AI Ultra subscribers, trusted testers, and select business users; the official page does not show broad global availability.
Impact: Google is moving Gemini from a chat assistant toward an always-on personal agent. It is not Gemini Omni and not a new Gemini foundation model release, so it should be recorded separately from model launches.
Sources: @GeminiApp Spark Beta, @GeminiApp Spark overview, Gemini Spark official page

4. Seed 2.1 Pro Preview: enters Arena WebDev top 10

Event type: Arena appearance + official Arena X signal.
First/official/availability time: the Arena WebDev board is dated 2026-06-20; @arena highlighted the model entering the top 10 at 2026-06-22T01:11:54Z.
Model/function name: seed-2.1-pro-preview.
Capability change: Arena WebDev shows seed-2.1-pro-preview at No. 8 with a score of 1539+15/-15, marked preliminary, with price and context length shown as N/A. Arena's official X post says it is strong on React apps and placed top 10 in five of seven subcategories.
Availability: visible on Arena WebDev; public API, official model card, pricing, and context window remain unverified.
Impact: this is today's most important Arena signal. It shows the ByteDance/Seed model line entering a visible competitive range for frontend generation, but until official product information is available it should be classified as "Arena appearance," not a complete launch.
Sources: Arena WebDev, @arena Seed 2.1 Pro Preview

5. Grok Build `/goal`: xAI adds a long-running task entry point

Event type: official X first signal / product capability update.
First/official/availability time: @grok introduced /goal at 2026-06-22T16:28Z.
Model/function name: Grok Build /goal.
Capability change: the official description frames it as a long-running task mode that can use multiple rounds of subagents to implement and verify a goal.
Availability: Grok Build; no corresponding model ID, API documentation, or benchmark was found.
Impact: this is not a model launch, but it fits the same product trend as Codex, Claude Code, Antigravity, and Gemini Spark: model companies are turning multi-round verification, subagents, and long-running tasks into default interaction surfaces. Watch for reproducible coding/webdev benchmarks or formal documentation.
Source: @grok /goal

Arena Watch

LMArena / Arena Agent: Claude Fable 5 (High) remains No. 1 with Net Improvement 14.05%+/-1.53%; Claude Opus 4.8 (Thinking) is No. 2, GPT 5.5 (xHigh) No. 3, Claude Opus 4.7 Thinking No. 4, GPT 5.5 High No. 5, and GLM 5.2 Max No. 10. The page is dated 2026-06-19; no new agent-board reshuffle announcement was found in this window.
Arena Text: claude-fable-5 is No. 1 at 1508+/-9, claude-opus-4-6-thinking No. 2 at 1504+/-4, and claude-opus-4-7-thinking No. 3 at 1502+/-5; muse-spark is No. 6, gemini-3.1-pro-preview No. 7, gemini-3-pro No. 8, claude-opus-4-8-thinking No. 9, and gpt-5.5-high No. 10.
Arena WebDev: claude-fable-5 is No. 1 at 1654+16/-16, glm-5.2 (max) No. 2 at 1593+15/-15, claude-opus-4-8-thinking No. 3, and claude-opus-4-7-thinking No. 4. seed-2.1-pro-preview is No. 8 at 1539+15/-15, preliminary; qwen3.7-max-20260517 is No. 10, glm-5.1 No. 11, and gpt-5.5-xhigh (codex-harness) No. 16.
Arena Vision: claude-opus-4-7-thinking is No. 1 at 1309+/-8, claude-fable-5 No. 2 at 1307+/-15, and claude-opus-4-6-thinking No. 3 at 1301+/-8.
Arena Document: claude-opus-4-6 at 1507+/-6 and claude-opus-4-6-thinking at 1507+/-7 share the top band, followed by claude-opus-4-7-thinking at 1498+/-7.
Arena Text-to-Image: gpt-image-2 (medium) is No. 1 at 1385+/-6, 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. The nano-banana line is a clear leaderboard item, but official naming and product-line mapping still require separate verification.
Arena Text-to-Video: the page is dated 2026-06-10; gemini-omni-flash is No. 1 at 1527+/-14, dreamina-seedance-2.0-720p No. 2, happyhorse-1.0 No. 3, veo-3.1-audio-1080p No. 4, wan2.7-t2v No. 5, sora-2-pro No. 7, and grok-imagine-video-720p No. 10.
Arena Image-to-Video: the page is dated 2026-06-10; gemini-omni-flash at 1475+/-14 and dreamina-seedance-2.0-720p at 1475+/-10 sit in the top band. grok-imagine-video-1.5-preview-720p is No. 3, happyhorse-1.0 No. 4, grok-imagine-video-720p No. 5, and kling-v3-pro No. 12.
Arena Video Edit: the page is dated 2026-05-13; dreamina-seedance-2.0-720p is No. 1, happyhorse-1.0 No. 2, grok-imagine-video No. 3, kling-o3-pro No. 4, kling-o1-pro No. 5, and runway-gen4-aleph No. 6.
Artificial Analysis: the homepage still highlights Intelligence Index v4.1 and AA-Briefcase as core evaluation updates. The current top range includes Claude Fable 5 (with fallback) at 60, marked not currently available; GPT-5.5 (xhigh) at 55; GLM-5.2 (max) at 51; Gemini 3.1 Pro Preview at 46; MiniMax-M3 at 44; and DeepSeek V4 Pro (Max) at 44. On cost per task, the page shows DeepSeek V4 Pro at $0.04/task, gpt-oss-120b at $0.06/task, MiniMax-M3 at $0.12/task, GLM-5.2 at $0.52/task, GPT-5.5 xhigh at $0.86/task, and Claude Fable 5 at $2.75/task.
Sources: Arena overview, Arena Agent, Arena WebDev, Arena Text-to-Video, Arena Image-to-Video, Arena Video Edit, Artificial Analysis

X Watch

@OpenAI: posted multiple Daybreak, Patch the Planet, and Cyber Partner Program updates at 2026-06-22T17:04Z. This is today's strongest official model/product signal. Credibility is high and cross-checked with OpenAI's official articles.
@OpenAIDevs: posted a Codex-as-research-partner example for iOS/macOS development at 2026-06-22T20:21Z. This is Codex workflow amplification, not a model update. OpenAI also published the Codex-maxxing long-running work guide the same day; classify it as methodology/workflow.
@GeminiApp: posted a Gemini Spark thread at 2026-06-22T19:38Z, describing it as a 24/7 personal AI agent and saying it is starting beta for Google AI Ultra subscribers in the U.S. Credibility is high and cross-checked with the official Gemini page.
@GoogleDeepMind: posted about a research partnership with A24 at 2026-06-22T14:35Z, focused on creator-shaped tools. This is not a new model release.
@arena: posted the seed-2.1-pro-preview WebDev signal at 2026-06-22T01:11Z; at 2026-06-22T19:25Z, it reiterated that the Arena leaderboard is a continuously refreshed living benchmark. Seed 2.1 Pro Preview should be marked as an Arena appearance pending an official model card.
@xiaohu: no usable new model-launch post was extracted from the X timeline; the XiaoHu AI Daily site for 2026-06-22 was readable and listed SkillOpt, OpenAI's partner network, Anthropic/Fable 5 access disputes, and Claude dynamic workflows as high-score themes. Today XiaoHu is used as an auxiliary source, not as direct proof of official first release.
@ArtificialAnlys: no new leaderboard-launch post was extracted in this window; the Artificial Analysis homepage still provides ongoing leaderboard evidence for Intelligence Index v4.1, AA-Briefcase, GLM-5.2, and related models.
@Zai_org: GLM-5.2-related reposts appeared in this window, including prompt methods and AWS Marketplace availability discussion. The original capability comparison came from an earlier date, so it should not be upgraded into a new release today.
@grok / @xai: @grok introduced Grok Build /goal, emphasizing long-running tasks and multi-round subagents. It is not a model ID update, but it belongs in the agent-product trend line.
@MistralAI: this window mainly contained an organizational signal about reaching 1,000 people; no new model release.
@cohere: this window mainly contained sovereign AI discussion; no new model release.
@runwayml: this window mainly contained an Aleph 2.0 aspect-ratio tutorial; no new model release.
@Kling_ai: this window contained creator profile / showcase content; no new Kling model release.
@AnthropicAI / @claudeai: no new Claude model release was extracted. Price Per Token listed media coverage of Anthropic's government dispute; classify it as media/policy signal, not a model update.
@AIatMeta, @LumaLabsAI, @midjourney: no new model launch strong enough to upgrade into Must Watch was extracted in this window.

Pending Confirmation

Gemini Omni / gemini-omni-flash: Arena video-board rank is clear, and older Google blog/release-note material helps explain the Gemini Omni/Gemini Omni Flash product direction, but this window did not produce a new official model card, API pricing, or first-release timestamp. Continue tracking Google AI, Gemini, Vertex AI, Flow, and YouTube Shorts entry points.
happyhorse-1.0: high on Arena Text-to-Video, Image-to-Video, and Video Edit, but still missing official team attribution, release date, availability, API/pricing, and licensing. Treat it as a high-heat anonymous or semi-anonymous Arena model.
nano-banana-2 / nano-banana-pro: the Arena Text-to-Image entries are clear and map to gemini-3.1-flash-image-preview and gemini-3-pro-image-preview-2k, but official Gemini/Google AI pages still need to confirm product naming, first-release timing, pricing, and API availability.
seed-2.1-pro-preview: today's Arena signal is strong, but the ByteDance/Seed official model card, context window, pricing, API entry point, and whether this is Arena-only testing remain unconfirmed.
Claude Fable 5: Arena and Artificial Analysis capability signals remain strong, but AA marks it as not currently available. Anthropic access limits and fallback rules need daily verification. Do not infer callability from leaderboard rank alone.
GLM-5.2: GLM-5.2 (max) remains strong on Arena/AA, and X had AWS Marketplace and prompt-method discussion, but there was no new official Z.ai release in this window. Continue verifying provider limits, stability, and real coding-agent reproduction tasks.
GPT-5.5-Cyber: OpenAI officially confirmed the full version and security benchmarks, but access is still limited to trusted defenders. Follow up on developer API availability, partner application paths, Codex Security plugin boundaries, and audit workflow.
Grok Build /goal: official X post exists, but model ID, formal docs, reproducible benchmark, and account availability are still missing. Keep it as a product capability pending documentation.