QName AI logoQName AI
All AI model reports
Collected in Hong Kong time

AI Model Daily - 2026-06-18

2026-06-18 AI model news: GLM-5.2 is confirmed by Artificial Analysis as the leading open-weights model, Arena adds GLM-5.2 and Kimi-K2.7-Code agent signals, xAI releases Grok Imagine Video 1.5 and brings Grok 4.3 to Bedrock, and OpenAI publishes LifeSciBench.

30-Second Summary

  • GLM-5.2 now has stronger third-party confirmation: Artificial Analysis published its 2026-06-17 evaluation, saying it scores 51 on Intelligence Index v4.1, becomes the new leading open-weights model, and lands on the Intelligence vs Cost per Task Pareto frontier.
  • Arena continues to strengthen the GLM-5.2 signal: Agent Arena is now at 786,587 sessions and 28 models, with GLM 5.2 (Max) at No. 10; Code Arena WebDev shows glm-5.2 (max) at No. 2, behind only claude-fable-5.
  • xAI has two high-priority signals today: Grok Imagine Video 1.5 has moved from preview to API GA under model name grok-imagine-video-1.5; Grok 4.3 is also available through Amazon Bedrock, with 1M context and configurable reasoning effort.
  • OpenAI did not release a new general model, but published LifeSciBench and a GPT-5.4/Maria AI chemistry workflow: GPT-Rosalind outperforms GPT-5.5 on LifeSciBench, while GPT-5.4 helped improve Chan-Lam coupling in a real experimental workflow.
  • Video and image Arena leaderboards did not refresh: gemini-omni-flash, happyhorse-1.0, and the nano-banana family remain old-leaderboard watch items; new video-model attention shifts to Grok Imagine Video 1.5, Luma Ray 3.2 on Runware, and Runway API Recipes.

Must Watch

1. GLM-5.2: Artificial Analysis confirms it as the leading open-weights model

  • Event type: third-party benchmark + post-release official verification + availability.
  • First/official/availability time: Z.ai's official release was on 2026-06-16; the Artificial Analysis evaluation article is dated 2026-06-17.
  • Model/function name: GLM-5.2 / GLM-5.2 (max).
  • Capability change: Artificial Analysis says GLM-5.2 scores 51 on Intelligence Index v4.1, ahead of MiniMax-M3 44, DeepSeek V4 Pro (max) 44, and Kimi K2.6 43, making it the new leading open-weights model. It improves by +11 points over GLM-5.1.
  • Key metrics: under the AA/Z.ai count, GLM-5.2 is 744B total / 40B active, MIT licensed, and has 1M context. First-party API pricing is listed as $1.4 / $4.4 / $0.26 per 1M input/output/cache-hit tokens. GDPval-AA v2 score is 1524, close to GPT-5.5 xHigh 1514, and clearly above MiniMax-M3 1418 and DeepSeek V4 Pro max 1328.
  • Cost/efficiency: AA estimates GLM-5.2 costs about $0.46 per Intelligence Index task, lower than closed models at similar intelligence levels. However, it emits about 43k tokens per task, including 37k reasoning tokens, so token efficiency still needs watching.
  • Availability: Z.ai API, Hugging Face, Fireworks, DeepInfra, Novita, Nebius, Parasail, SiliconFlow, GMI Cloud, Baseten, and others. OpenCode also said on X that GLM-5.2 is available in Go and is seeing 3x demand.
  • Impact: GLM-5.2 can now be upgraded from "new release pending validation" to "third-party-confirmed open-weights leader." Its strengths concentrate in long-horizon agent, science, and coding tasks; production selection still needs 1M-context stability, output-token cost, and provider-latency testing.
  • Sources: Artificial Analysis GLM-5.2, @Zai_org, Z.ai GLM-5.2 blog

2. Arena update: GLM-5.2 WebDev #2, Kimi-K2.7-Code enters Agent Arena

  • Event type: Arena appearance + X first signal + third-party leaderboard.
  • First/official/availability time: Code Arena WebDev shows Jun 16, 2026; @arena posted the Agent Arena model-addition update on 2026-06-17.
  • Model/function name: glm-5.2 (max), GLM 5.2 (Max), Kimi K2.7 Code.
  • Code Arena WebDev: the page shows 381,168 votes and 90 models. claude-fable-5 is No. 1 with score 1654+16/-16; glm-5.2 (max) is No. 2 with score 1595+16/-16, 1,641 votes, price $1.40 / $4.40, and 1M context; claude-opus-4-7-thinking is No. 3.
  • Agent Arena: the page is now at 786,587 sessions and 28 models. GLM 5.2 (Max) is No. 10, with Net Improvement 4.37%±2.48%, Confirmed Success 9.43%±4.52%, and 11,626 sessions. Kimi K2.7 Code is No. 19, with Net Improvement 2.71%±2.39% and 14,630 sessions.
  • X signal: @arena says Agent Arena added 10 models over two weeks, GLM-5.2 is the strongest open-weight result measured so far, and Kimi-K2.7-Code is overall No. 19 and open-model No. 6.
  • Impact: Arena and AA now cross-validate GLM-5.2 as this week's strongest open/open-weight coding-agent candidate. Kimi-K2.7-Code has entered the real-world agent leaderboard, but remains behind GLM-5.2 and the top closed-model tier.
  • Sources: Code Arena WebDev, Arena Agent, @arena

3. xAI: Grok Imagine Video 1.5 GA, and Grok 4.3 arrives on Amazon Bedrock

  • Event type: official release + availability + cloud distribution.
  • First/official/availability time: xAI's official news page is dated 2026-06-16; @xai posted Grok Imagine Video 1.5 is here and Grok is now available on Amazon Bedrock within this window.
  • Model/function name: grok-imagine-video-1.5, Grok Imagine Video 1.5, Grok 4.3.
  • Capability change: xAI says Grok Imagine Video 1.5 has moved from preview to API GA, improves motion/physics, prompt adherence, and audio/speech generation, and can generate a 6-second 720p video clip in about 25 seconds. The API model name is grok-imagine-video-1.5.
  • Bedrock signal: the AWS Bedrock model card shows Grok 4.3 with 1M token context, configurable reasoning effort (none/low/medium/high), on-demand pricing of $1.25 input and $2.50 output per 1M tokens, and Amazon Bedrock Mantle secure inference.
  • Availability: Grok API, grok.com/imagine, Amazon Bedrock, and DigitalOcean Marketplace Grok Build VM.
  • Impact: this is today's clearest official video-model release. Grok 4.3 on Bedrock is an enterprise distribution signal that lowers access friction for AWS users; Bedrock-vs-native differences in tool calling, reasoning effort, latency, and region coverage still need testing.
  • Sources: @xai, Grok Imagine 1.5, AWS Bedrock Grok 4.3 model card

4. OpenAI: LifeSciBench and GPT-5.4 chemistry workflow, more science-eval than general-model release

  • Event type: official benchmark + official research/case study.
  • First/official/availability time: OpenAI published LifeSciBench on 2026-06-17; it also posted the GPT-5.4 medicinal-chemistry workflow case on X the same day.
  • Model/function name: LifeSciBench, GPT-Rosalind, GPT-5.4.
  • Capability change: LifeSciBench contains 750 expert-authored tasks from 173 scientists across seven biological research categories. It is designed to test whether models can reason from evidence, work with scientific artifacts, handle uncertainty, and make useful decisions under real-world constraints.
  • Model signal: OpenAI's X thread says GPT-Rosalind scores above GPT-5.5 on LifeSciBench. GPT-5.4, paired with Molecule.one's Maria AI and a specialized lab workflow, proposed and experimentally validated a way to improve a reaction widely used in drug discovery.
  • Availability: OpenAI LifeSciBench article and OpenAI chemistry case article. No standalone GPT-Rosalind API or model card was found.
  • Impact: this is not a new general-model release, but it matters for scientific-agent selection. Watch whether GPT-Rosalind becomes a usable model/tool, whether LifeSciBench gains a public leaderboard, and how reproducible GPT-5.4's chemistry workflow is.
  • Sources: @OpenAI, LifeSciBench, OpenAI AI chemist

5. Soniox v5 Real-Time: speech model reaches AA Pareto frontier

  • Event type: official release + third-party benchmark + availability.
  • First/official/availability time: Soniox announced v5 Real-Time on X on 2026-06-16; Artificial Analysis evaluated it on 2026-06-17.
  • Model/function name: Soniox v5 Real-Time.
  • Capability change: Soniox says v5 Real-Time supports 60+ languages, low-latency streaming speech-to-text, and real-time segmentation controls such as endpoint sensitivity. Artificial Analysis says it sits on the accuracy/latency Pareto frontier and is the lowest-priced proprietary model tested.
  • Availability: Soniox console/API; pricing signal is $2 per 1,000 minutes of audio.
  • Impact: if the daily report includes models beyond LLMs, this belongs in scope. It directly matters for realtime voice agents, meeting transcription, and voice operating systems. Next comparisons should include Gemini/OpenAI/ElevenLabs/Deepgram-style low-latency STT models.
  • Sources: @soniox_ai, @ArtificialAnlys

Arena Watch

  • LMArena Text Overall: no core structural change was observed versus yesterday; GLM-5.2 remains a mid-to-high Text overall entry, while today's main new signals moved to AA Intelligence Index and Code/Agent Arena.
  • Code Arena WebDev: Jun 16, 2026, 381,168 votes, 90 models. Top five are claude-fable-5, glm-5.2 (max), claude-opus-4-7-thinking, claude-opus-4-8-thinking, and kimi-k2.6. GLM-5.2 is today's key new frontend/webdev entry.
  • Arena Agent: 786,587 sessions, 28 models. Claude Fable 5 (High) remains No. 1; Claude Opus 4.8 (Thinking) No. 2; GPT 5.5 (xHigh) No. 3; GLM 5.2 (Max) No. 10; Kimi K2.7 Code No. 19.
  • 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 leaderboard movement.
  • 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, happyhorse-1.0 No. 3, and grok-imagine-video-720p No. 10. Grok Imagine Video 1.5 has been released, but no 1.5 entry was observed on this page yet.
  • Artificial Analysis: GLM-5.2 is the new open-weights leader; Soniox v5 Real-Time is the new speech-to-text Pareto frontier signal; v4.1 continues to emphasize agentic workloads.
  • Sources: Arena Agent, Code Arena WebDev, Arena Text-to-Image, Arena Text-to-Video, Artificial Analysis

X Watch

  • @ArtificialAnlys: published GLM-5.2's AA Intelligence Index v4.1 open-weights-leading result, added CritPt physics benchmark context, and evaluated Soniox v5 Real-Time STT.
  • @arena: posted that Agent Arena has added 10 models over two weeks; highlighted GLM-5.2 as the strongest open-weight result and Kimi-K2.7-Code as Agent Arena overall #19/open #6.
  • @Zai_org: continued reposting GLM-5.2 availability and best-practice signals, including OpenCode Go, Featherless, AA evaluation, and user setup advice.
  • @xai / @grok: Grok Imagine Video 1.5, Grok 4.3 on Amazon Bedrock, and DigitalOcean Grok Build VM are the high-signal items; HyperFrames connector is a product integration.
  • @OpenAI: published LifeSciBench and the GPT-5.4 chemistry workflow; @OpenAIDevs mainly shared Codex events and Rust Foundation support, not a model release.
  • @GeminiApp: posted Gemini Live community demo, Nano Banana 2 templates, Pixel Drop Gemini Omni music generation, and Daily Brief product updates. No new Gemini Omni model card was found.
  • @Alibaba_Qwen: still centered on the June 16 Qwen-Robot Suite thread, with no newer Qwen model launch.
  • @xiaohu: focused on Codex OSS/local provider configuration, TRAE event posts, and China app access observations. XiaoHu Daily's 2026-06-17 headline is the SpaceX/Cursor claim, which remains unverified.
  • @AnthropicAI / @ClaudeDevs: mainly Claude Code economic research, Managed Agents, and Build Day demos; Fable/Mythos suspension remains the background availability risk.
  • @cohere: continued North Mini Code and AI sovereignty messaging; no new model.
  • @runwayml: announced Runway API Recipes, a one-call generative media workflow feature. This is API product capability, not a new foundation model.
  • @LumaLabsAI: reposted Runware's Ray 3.2 availability, including text-to-video, image-to-video, start/end frame, 5s/10s output, up to 1080p, HDR/EXR, looping, and extension.
  • @Kling_ai, @MistralAI, @AIatMeta: no new model release observed; mostly cases, events, older pinned models, or ecosystem updates.

Pending Verification

  • GLM-5.2 parameter count still differs across sources: HF shows 753B params, while AA/Z.ai references use 744B total / 40B active. Use the official technical report or architecture table once available.
  • GLM-5.2 WebDev #2 and Agent Arena #10 are visible, but need continued monitoring as sample size, confidence interval, price/latency, and 1M-context failure rate evolve.
  • Grok Imagine Video 1.5 is officially GA, but we still need to confirm when it appears on Arena Video, whether it replaces grok-imagine-video-720p, API pricing, concurrency, and audio-boundary details.
  • Grok 4.3 on Bedrock needs testing against the native xAI API for tool calling, reasoning effort, latency, and region availability.
  • GPT-Rosalind currently appears only in the LifeSciBench context. No standalone model card/API was found, so it should not be treated as an open model yet.
  • Soniox v5 Real-Time needs fuller WER, latency, streaming stability, diarization, and endpoint sensitivity tests.
  • XiaoHu Daily's SpaceX 60B Cursor acquisition and SpaceXAI/Cursor model claims still lack official SpaceX, Cursor, or filing confirmation. Treat them as high-heat rumors only.
  • gemini-omni-flash, happyhorse-1.0, and nano-banana-2 still lack complete official model cards, first-release timestamps, API/pricing boundaries.