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

AI Model Daily - 2026-07-01

2026-07-01 AI model news: Anthropic released Claude Sonnet 5, Google released Nano Banana 2 Lite and Gemini Omni Flash, Arena added Sonnet 5 and ranked Omni Flash #2 in Video Edit, Artificial Analysis published Sonnet 5 early results, and OpenAI released GeneBench-Pro as a biology benchmark.

30-Second Summary

  • Anthropic officially released Claude Sonnet 5, positioned as its "most agentic Sonnet" yet. The model ID is claude-sonnet-5; it is now the default for Free / Pro and is available to Max, Team, Enterprise, Claude Code, and Claude Platform. Introductory API pricing through August 31 is $2/MTok input and $10/MTok output.
  • Google released two creative-model updates: Nano Banana 2 Lite, also known as gemini-3.1-flash-lite-image, is generally available as the fastest and lowest-cost Gemini Image model; Gemini Omni Flash, also known as gemini-omni-flash-preview, is in public preview for 720p, 3-10 second video generation and conversational video editing.
  • Arena added the new models quickly: Claude Sonnet 5 is now in Agent Arena; Gemini Omni Flash ranks #2 in Video Edit Arena at 1347±16, behind Dreamina Seedance 2.0 720p at 1377±12.
  • Artificial Analysis published early Claude Sonnet 5 results: max effort scores 53 on Intelligence Index v4.1, up 6 points from Sonnet 4.6, but uses roughly 40% more output tokens per task and would cost more per task than Opus 4.8 without promotional pricing.
  • OpenAI did not release a new model today. It released GeneBench-Pro, a research-grade computational-biology benchmark. GPT-5.6 Sol reached 28.7% pass rate at highest reasoning level and 31.5% in Pro mode, measuring agent judgment on messy real-world biological data analysis.

Must Watch

1. Claude Sonnet 5: Anthropic releases the next Sonnet and makes it the Claude Free / Pro default

  • Event type: official release + X-first signal + availability change.
  • First/official/availability time: Anthropic's official page is dated 2026-06-30; the first @claudeai post was at 2026-06-30T18:00Z; @AnthropicAI reposted it.
  • Model/function names: Claude Sonnet 5, claude-sonnet-5, adaptive reasoning / effort levels.
  • Capability change: Anthropic describes Sonnet 5 as its "most agentic Sonnet" yet, emphasizing planning, browser and terminal tool use, and long autonomous task execution. Anthropic says it improves over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work, and approaches Opus 4.8 on some high-effort tasks.
  • Availability: default model on Claude Free / Pro; available to Max, Team, and Enterprise; also available in Claude Code and Claude Platform. API model ID is claude-sonnet-5.
  • Pricing: Claude Platform introductory pricing through 2026-08-31 is $2/M input tokens and $10/M output tokens; standard pricing after that is $3/M input and $15/M output.
  • Safety/limits: Anthropic says Sonnet 5 shows lower undesirable behavior than Sonnet 4.6 and is better suited for agentic scenarios. Cyber capability is below current Opus models, but cyber safeguards are still enabled by default.
  • Impact: this is today's most important general-purpose model release. Near-term testing should focus on Claude Code, browser/terminal tool use, long-running follow-through, code maintenance, and enterprise automation stability, while also tracking high-effort token expansion.
  • Sources: Anthropic: Introducing Claude Sonnet 5, @claudeai, @claudeai availability, Artificial Analysis Claude Sonnet 5

2. Nano Banana 2 Lite: Google's low-latency, low-cost image model is GA

  • Event type: official release + API availability change.
  • First/official/availability time: Google AI / Google DeepMind first posted around 2026-06-30T16:03Z; the Google Cloud blog is dated 2026-07-01; the Google Cloud model page lists release date 2026-06-30.
  • Model/function names: Nano Banana 2 Lite, Gemini 3.1 Flash-Lite Image, gemini-3.1-flash-lite-image.
  • Capability change: Google positions it as the fastest and most cost-efficient image generation / editing model in the Nano Banana family. The official blog says it supports fast image generation and iteration for low-latency scenarios, with world knowledge, character consistency, object fidelity, text, and localized-draft capabilities. Google Cloud docs list image generation, image editing, multi-turn image editing, video input to image generation, and C2PA support.
  • Availability: Gemini Enterprise Agent Platform / Agent Studio; Google Cloud model ID gemini-3.1-flash-lite-image. It supports Provisioned Throughput, Standard PayGo, Flex PayGo, and Batch inference.
  • Technical information: maximum input tokens 65,536, maximum output tokens 4,096; release date 2026-06-30; knowledge cutoff January 2025; docs say 1K, about 1MP output consumes 1120 output image tokens.
  • Impact: Nano Banana 2 Lite is a cost/latency-optimized model for high-throughput visual products, not the highest-quality flagship. It is immediately relevant for ad variants, social apps, bulk asset drafts, and embedded visual generation inside agents, and should be speed/cost tested quickly.
  • Sources: Google Cloud blog, Google Cloud model card, Gemini API release notes, @GoogleDeepMind, @GoogleAI

3. Gemini Omni Flash: Google's video generation/editing model reaches public preview and API

  • Event type: official release + API availability change + leaderboard change.
  • First/official/availability time: Google DeepMind first posted around 2026-06-30T16:03Z; Gemini API release notes are dated 2026-06-30; the Google Cloud blog is dated 2026-07-01.
  • Model/function names: Gemini Omni Flash, gemini-omni-flash-preview.
  • Capability change: Google describes Omni Flash as a high-performance multimodal model for fast video generation and conversational video editing. Through the Interactions API, it can generate 3-10 second 720p videos from text or animate still images, then refine outputs through conversational editing. The Google Cloud blog highlights character/product replacement, dynamic style transfer, adding objects, relighting, and preserving original audio/video tracks.
  • Availability: Gemini API, Gemini Enterprise Agent Platform, Google AI Studio / Agent Studio, and Model Garden. Runway also announced in this window that Gemini Omni Flash is now in Runway.
  • Technical information: the Google Cloud model page lists model ID gemini-omni-flash-preview, launch stage Preview, release date 2026-06-30, discontinuation date 2026-06-30, maximum input tokens 131,072, and maximum output tokens 57,920; video specs are 720p, up to 10 seconds per clip, up to 3 videos per prompt, and up to 10 images.
  • Arena signal: Arena.ai says Gemini Omni Flash ranks #2 in Video Edit Arena at 1347±16, behind dreamina-seedance-2.0-720p at 1377±12, and ahead of HappyHorse 1.0 at 1308±12.
  • Impact: this is the key step where Google pushes Omni from consumer/demo visibility into developer/API availability. Near-term tests should focus on video-editing chains, reference image/video consistency, the 10-second cap, Runway integration, and Google Cloud fixed-quota / preview terms.
  • Sources: Gemini API release notes, Google Cloud model card, Google Cloud blog, Arena Video Edit, @GoogleDeepMind, @runwayml

4. Claude Science: Anthropic releases a scientific-research workbench beta

  • Event type: official release + vertical workflow product.
  • First/official/availability time: Anthropic's official page is dated 2026-06-30; @claudeai posted at 2026-06-30T17:02Z.
  • Model/function name: Claude Science.
  • Capability change: Claude Science is an AI workbench for scientists, not a separate new model. It integrates common research tools and packages, creates auditable artifacts, manages compute resources, and offers 60+ optional scientific databases. It supports local macOS / Linux, SSH, and HPC login nodes; outputs include code, environment, and message history for reproducibility.
  • Availability: beta for Claude Pro, Max, Team, and Enterprise; Team / Enterprise requires admin enablement.
  • Impact: Claude Science and OpenAI GeneBench-Pro landing on the same day shows frontier model companies are making scientific workflows a next competitive focus. For research institutions and biotech teams, the key tests are local/HPC data boundaries, reviewer-agent reliability, and auditable computation/citation trails.
  • Sources: Anthropic: Claude Science, @claudeai Claude Science, Claude Science app

5. GeneBench-Pro: OpenAI releases a computational-biology benchmark, not a new model

  • Event type: official release + benchmark / evaluation.
  • First/official/availability time: OpenAI's official page is dated 2026-06-30; @OpenAI posted at 2026-06-30T17:10Z.
  • Model/function names: GeneBench-Pro, GPT-5.6 Sol.
  • Capability change: GeneBench-Pro uses 129 questions across 10 domains including genomics, quantitative biology, and translational medicine to evaluate whether agents can choose analysis paths, handle ambiguity, and make judgment calls on messy real biological data. OpenAI will release 10 representative questions and plans to provide a 50-question subset to Artificial Analysis for a third-party benchmark.
  • Result signal: OpenAI says GPT-5.6 Sol reached 28.7% pass rate at highest reasoning level and 31.5% in Pro mode. At the highest reasoning level, it solved nearly 6x as many questions as GPT-5.2 while using about two-thirds of GPT-5.2's tokens.
  • Availability: benchmark paper and partial data; not a new model entry.
  • Impact: GeneBench-Pro is evidence for GPT-5.6 Sol's scientific reasoning claims and part of the same AI-for-science narrative as Anthropic's Claude Science. Do not classify it as a model launch; track the later 50-question Artificial Analysis result.
  • Sources: OpenAI: Introducing GeneBench-Pro, @OpenAI GeneBench-Pro

Arena Watch

  • Agent Arena: Claude Sonnet 5 is now in Arena.ai Agent Arena. @arena says Agent Arena uses real long-horizon agentic tasks where models can use web search, filesystem, and terminal tools. The X post did not provide a score yet, so classify this as "arena appearance."
  • Video Edit Arena: dreamina-seedance-2.0-720p is #1 at 1377±12 with 3,010 votes; gemini-omni-flash is #2 at 1347±16, Preliminary, with 998 votes; happyhorse-1.0 is #3 at 1308±12 with 2,681 votes; grok-imagine-video is #4 at 1264±8; kling-o3-pro is #5 at 1251±10.
  • Artificial Analysis: Claude Sonnet 5 max effort scores 53 on Intelligence Index v4.1. AA says this is 6 points above Sonnet 4.6, but output tokens per task are roughly 40% higher, averaging about 69k output tokens / task; without promotional pricing, cost per task is higher than Opus 4.8.
  • Google Gemini: gemini-omni-flash-preview and gemini-3.1-flash-lite-image now have official API / model-card backing, so they are no longer merely anonymous Arena signals. Continue watching whether T2V, I2V, and Video Edit scores stabilize.
  • HappyHorse: no new HappyHorse 1.1 launch today. The Video Edit page currently shows HappyHorse 1.0 at #3; this window still did not show a new HappyHorse 1.1 API ID or official model card.
  • Sources: Arena Video Edit, @arena Claude Sonnet 5, @arena Gemini Omni Flash, @ArtificialAnlys Claude Sonnet 5, Artificial Analysis Claude Sonnet 5

X Watch

  • @claudeai: first posted Claude Sonnet 5 at 2026-06-30T18:00Z, calling it the "most agentic Sonnet yet"; the same thread confirmed Sonnet 5 is now the Free / Pro default and is available in Max, Team, Enterprise, Claude apps, and Claude Platform. Classification: official release + X-first signal + availability change.
  • @AnthropicAI: reposted Claude Sonnet 5. No other new Claude model appeared in this window beyond Sonnet 5 / Claude Science. Classification: official release.
  • @GoogleDeepMind: posted Nano Banana 2 Lite and Gemini Omni Flash at 2026-06-30T16:03Z; highlighted the Interactions API workflow: generate an image with Nano Banana 2 Lite, then animate it with Gemini Omni Flash, with up to three sequential edits. Classification: official release + API availability change.
  • @GoogleAI: posted the same creative-workflow update, emphasizing fast image generation followed by immediate animation. Classification: official release.
  • @arena: at 2026-06-30T19:12Z, announced Claude Sonnet 5 in Agent Arena; at 2026-06-30T20:06Z, announced Gemini Omni Flash at #2 in Video Edit Arena. Classification: arena appearance + leaderboard change.
  • @ArtificialAnlys: at 2026-06-30T20:59Z, published early Claude Sonnet 5 results: Intelligence Index 53, +6 vs Sonnet 4.6, higher token/cost profile, and agentic knowledge work close to or above Opus 4.8. Classification: benchmark/evaluation source.
  • @xiaohu: at 2026-06-30T14:08Z, caught the same-day Google model and Claude Sonnet 5 launch signals ahead of release; also posted a personal/rumor item about Claude blocking Zhejiang/Hangzhou IPs, which should be marked pending verification and not written as official fact. Classification: X-first/media signal + pending rumor.
  • @OpenAI: at 2026-06-30T17:10Z, released GeneBench-Pro. Not a new model. Classification: official benchmark.
  • @OpenAIDevs: this window was mostly Codex / AI Engineer conference / data infrastructure debugging content. Not a new model. Classification: official product/engineering signal.
  • @runwayml: at 2026-06-30T19:24Z, announced Gemini Omni Flash is now available in Runway to generate and edit video from prompts, images, or video, and can be used through Agent. Classification: official product integration + availability change.
  • @LumaLabsAI: at 2026-06-30T19:33Z, announced Seedance 2.0 Mini is now available in Luma. This is a third-party model product entry in Luma, not a new Luma in-house model. Classification: official product integration.
  • @Kling_ai: this window was mainly Cannes Lions / work showcase content; no new Kling model ID appeared. Classification: no new model launch.
  • @Alibaba_Qwen, @Alibaba_Wan, @MistralAI, @xai, @grok, @AIatMeta, @cohere, @midjourney, @MicrosoftAI: no official X signal in this window was strong enough to classify as a new model release. Classification: no new model launch.
  • Sources: @claudeai, @GoogleDeepMind, @GoogleAI, @arena, @ArtificialAnlys, @xiaohu, @OpenAI, @runwayml, @LumaLabsAI

Pending Confirmation

  • Claude Sonnet 5 real-world agent success rate: official and early-partner examples are strong, but Claude Code, browser/terminal tools, enterprise automation, and long-context project maintenance need hands-on testing, especially whether high effort makes token costs hard to control.
  • Claude Sonnet 5 cost on Artificial Analysis: AA says per-task cost is higher than Opus 4.8 without promotional pricing and output tokens are roughly +40%. More provider data and real task bills are needed.
  • Claude Sonnet 5 safety boundaries: Anthropic says cyber capability is below Opus / Mythos, but cyber safeguards are enabled by default; watch for false positives in legitimate security research and code auditing.
  • Gemini Omni Flash preview constraints: Google Cloud docs show preview status, a 10-second cap, 720p output, and some "soon" capabilities. Test fixed quota, generation time, audio consistency, and multi-turn conversational editing stability.
  • Nano Banana 2 Lite cost: official docs provide image-token consumption and pricing entry points, but real per-image price depends on text/video input tokens, PayGo/Flex/PT, and regional quota.
  • Runway's Gemini Omni Flash integration: confirm whether it is only an app entry or whether API/Agent calls are stable, and verify Runway-side billing, output watermarking, and commercial terms.
  • Luma's Seedance 2.0 Mini entry: X confirms it is now available in Luma, but API access, pricing, resolution, duration, and commercial terms still need verification.
  • Xiaohu's Claude Zhejiang/Hangzhou IP blocking claim: currently a personal/media X source; it needs official mail evidence, Anthropic status/source confirmation, multiple independent samples, or an official statement.
  • OpenAI GeneBench-Pro third-party evaluation: OpenAI says it will provide a 50-question subset to Artificial Analysis; wait for AA's independent result instead of relying only on OpenAI's self-evaluation.
  • Grok 4.5: yesterday's Elon private-beta signal still has no xAI official docs/API/model card; today @xai / @grok did not add confirmation.