Xiaoliu BOT
中文EN

X Platform April 16th AI Digest | Claude Opus 4.7 Released, GPT Image 2 Realism Raises Concerns, Dreamina Video API Commercialized

Claude Opus 4.7 Released, Multiple Bloggers Interpret New Capabilities

Multiple bloggers mentioned that Anthropic officially released Claude Opus 4.7 on April 16th. Pricing remains unchanged from Opus 4.6 (input at $5 per million tokens, output at $25). It has simultaneously launched across all products and cloud platforms including Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. @claudeai’s official announcement states the new model focuses on three upgrades: self-verification before reporting during long tasks, vision support up to a maximum long edge of 2576 pixels (about 3 times that of previous Claude models), and programming capabilities reaching SOTA in third-party evaluations. @dotey adds that the new tokenizer increases token consumption for the same text by 1.0–1.35 times. Under high reasoning levels, output tokens are longer, so actual bills may see a slight increase. @LufzzLiz’s real-world tests show excellent results for weather card generation, but failures occurred in firework and car wash scenarios. @xiaohu perceives that Opus 4.7’s reasoning speed is more than twice as fast as 4.6, with a new adaptive thinking mode (dynamically adjusting depth) and enhanced file system memory capability, allowing it to remember important notes across sessions. @vista8 questions why old models inevitably get dumbed down before new model releases, speculating it’s due to insufficient computing power. @claudeai also announced that Claude Code has added a /ultrareview command for deep code review. The auto mode (allowing continuous autonomous execution) is now fully open to Max users for the first time. A new xhigh reasoning level, between high and max, has been added, and the default reasoning level for all plans has been upgraded to xhigh. @dotey specifically points out a subtle background to this upgrade: Anthropic has a stronger Claude Mythos Preview, but due to cybersecurity concerns, it is currently only available in limited quantities. Opus 4.7 is a version that has undergone security downgrading. @op7418 reminds that Claude Code now defaults to using Opus 4.7 with a default reasoning strength of xhigh; if insufficient, users can manually switch to Max.

Sources:

GPT Image 2 and Mysterious Model “duct-tape” Attract Attention

Multiple bloggers mentioned that OpenAI’s GPT Image 2 image generation capabilities have attracted attention. @oran_ge discovered that images generated with GPT Image 2 can even correctly follow ID number encoding rules (only the check digit is incorrect), lamenting that photo authenticity is a thing of the past and calling for more cautious treatment of image information. @xiaohu shared the actual generation result using the prompt “generate a screenshot of a Douyin livestream, featuring a beautiful woman streaming,” saying it indeed has something. @Astronaut_1216 cited @arrakis_ai’s discovery that a GPT-series model codenamed “duct-tape” is being tested in the arena, outperforming contemporary products and suspected to be GPT Image 2. @op7418 also mentioned that OpenAI’s gpt-image-2 image generation is accurate, comparing it favorably to Google’s recent nerfing of Gemini Nano 2, and stated their Pro membership was worth it.

Sources:

Gemini Mac Client Receives Negative Reviews, Rough Interface Becomes Consensus

Multiple bloggers mentioned that Google officially launched the Gemini Mac native client (written entirely in Swift), supporting the Option+Space shortcut to summon a small window, and capable of calling NotebookLM, Google Drive, DeepRearch, as well as image, video, and music generation capabilities. @vista8, after trial, noted that voice input accuracy is decent, but the interface design is too rough, screen sharing can only specify windows and not the whole screen, and even Gems are not supported, with the overall impression being “Google’s normal level of performance.” @op7418 similarly complained that the functionality is quite rudimentary, with even moderately complex Artifact web pages failing to render, and the entire UI being very crude. @xiaohu bluntly said “there’s no heart put into it,” believing that applications like this can be made by anyone with a bit of AI coding, and doesn’t know why it was only launched now. @dotey also feels Google’s iteration speed is really slow.

Sources:

Dreamina Seedance 2.0 API Launched, AI Video Commercialization Accelerates

@vista8 mentioned multiple times that ByteDance’s Dreamina Seedance 2.0 API has launched on BytePlus, featuring multimodal input (turning old footage into new videos), precise editing (changing details without regenerating), physical realism (motion and lighting consistent with the real world), and intent understanding (a single sentence can automatically complete shot transitions and rhythm control). It has reached commercial standards and includes copyright and portrait rights protection. @vista8 also shared usage tips: first use image generation tools like Jimeng to create 9-panel storyboards, then use Seedance 2.0 to generate videos referencing the storyboards, enabling the creation of narrative short films. @lovart_ai also previewed that the full-power 1080p version of Seedance 2.0 will be coming online soon.

Sources:

OpenAI Agents SDK Upgrade: Built-in Sandbox and State Management

@dotey relayed that OpenAI’s Agents SDK recently underwent an important upgrade, adding a built-in sandbox execution environment (supporting Cloudflare, Vercel, Modal, E2B, etc.) and a model-native Harness architecture (separating state preservation from computational execution, allowing quick recovery from crashes). The SDK also includes built-in MCP tool calling, AGENTS.md custom instructions, Shell tools, Apply Patch file editing, and a memory system, eliminating the need for general frameworks like LangChain. @dotey also recommends reading the article “The Model Isn’t Dumb, the Harness Isn’t Configured Right,” which points out that the root cause of AI agent failure is system design issues, not insufficient model capability, and proposes four principles of Harness Engineering: replacing prompt dependency with code constraints, externalizing critical states, third-party independent verification, and limiting failures to local scope.

Sources:

Skill Essence is Taxonomy, High-Frequency Use Requires Controlling Quantity

@Khazix0918 systematically outlined their thoughts on Skills, believing the core of a Skill is “classification” and “triggering,” not being as granular as possible. Paper data shows that when the number of Skills is below 20, accuracy remains above 90%, but drops sharply after exceeding 30, and is only 20% at 200; they personally keep their Skill library consistently below 30. Using the NanoBanana API wrapper as an example, different image generation scenarios (WeChat official account covers, Xiaohongshu covers, PPT illustrations) do not each warrant an independent Skill; it is more reasonable to unify them into one image generation Skill with internal secondary analysis. They proposed three criteria to determine if a Skill has value: whether the scenario boundaries are clear, whether it recurs frequently, and whether it can belong to an existing Skill. @op7418 also published a Logo generation Skill on the same day, which can generate card-style promotional images from product introductions, supports Nano Banana for generating display images, and can also generate dynamic display web pages with code.

Sources:

AI Programming Experience and Agent Capability Boundary Discussion

@vista8 shared that the experience of programming with Claude Code is increasingly resembling a product manager reviewing a launched product—constantly testing, raising requirements, and fixing bugs. @Astronaut_1216 also mentioned that B-end clients have a strong demand for “self-media customer acquisition + AI,” but believes the core has no bottlenecks and AI semi-automation is entirely feasible. @cellinlab cited @lijigang’s view that “Agents are becoming the external manifestation of everyone’s mental model,” and added that if we let AI guide life, it might ultimately become the external manifestation of an AI company’s ethics committee. @lijigang also pointed out that in the AI era, the defender (security) pays extra costs to temporarily suppress the attacker, while the attacker only needs to find one point. @Astronaut_1216 emphasized that people who don’t understand the SaaS logic of a certain industry will equally fail to make good AI products for that industry; the bubble lies in not understanding the essence of the industry.

Sources:

Qwen3.6-35B-A3B Open-Sourced, Elephant Alpha Model Debuts

@LufzzLiz mentioned that Alibaba has open-sourced Qwen3.6-35B-A3B, using a sparse MoE architecture with 35 billion total parameters and 3 billion activated, under an Apache 2.0 license. It is considered to pose strong competition to Gemma 4. @Astronaut_1216 discovered a new model codenamed Elephant Alpha (100B parameters, token-efficient, positioned as instant-type) appearing on OpenRouter, suitable for code completion, long document processing, and lightweight Agent workflows, and it is currently free, currently competing with GPT6 and Opus 4.7 by timing. @LufzzLiz also mentioned that Alibaba Cloud now supports the Hermes Agent.

Sources:

Elon Musk Continues to Focus on South Africa Policy and Starlink

Some bloggers mentioned that Elon Musk continued to speak out on South African policy that day, repeatedly reposting content involving South African racial laws, Starlink being rejected, and government lies, commenting “South African laws are literally super racist,” and calling the current South African government’s actions “Apartheid 2.0.” He also reposted content about SpaceX Starship Super Heavy V3’s first 33-engine static fire, Tesla Semi “hammer down,” Tesla vehicles with 800,000 kilometers of mileage, and Grok’s new traffic record (over 326 million visits in a single month). @vista8, in a summary of an interview with Li Guangmi, evaluated xAI as a “swinging heavy-asset player,” pointing out core team loss and strategic wavering. Musk lacks patience and is accustomed to quick feedback; large models require long-term accumulation, and xAI is constantly switching between multiple routes.

Sources:

Claude Intelligence Downgrade and Token Consumption Issues

Some bloggers mentioned that @aiwarts compiled a complete configuration solution for resolving Claude’s intelligence downgrade and rapid token consumption: by setting effortLevel to high in ~/.claude/settings.json, setting CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING to 1 (disabling adaptive thinking), setting MAX_THINKING_TOKENS to 31999 or higher, disabling 1M context, and enabling 200K automatic compression to mitigate the downgrade. They also specifically reminded that setting CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 previously, to avoid sending usage data to Anthropic, would cause the one-hour context cache for subscribers to be cut down to only five minutes, resulting in a significant loss. The best practice is to keep the conversation task unchanged and not start a new conversation if the last message was within one hour, which can save 40,000-60,000 tokens of context loading overhead.

Sources:

Other Updates

@NASA announced that the mobile launcher for the Artemis II lunar mission is rolling back to the Vehicle Assembly Building to prepare for the Artemis III mission; Artemis II astronauts will use tablets and smartphones to complete tasks like email during their time in space. @tim_cook introduced Apple’s progress in environmental protection, with the amount of recycled materials used in products reaching an all-time high. @bit_summit confirmed that the 10th bit summit will be held on September 23–24, 2026. @cellinlab shared personal experience with invoice issuance on behalf of others, noting that Shanghai tax audits are relatively lenient and do not require contracts or official seals. @oran_ge mentioned using an Agent to help with rights protection against a major bank, and if successful, could recover ten thousand US dollars. @MANISH1027512 released a combined public version of AIGC Weekly Vol.002+Vol.003, sharing AI photography golden hour lighting and lens flare special effects techniques. @NousResearch announced that the Hermes Agent now officially supports QQBot, with many users sharing cases of using Hermes Agent to fully autonomously complete complex tasks.

Sources:

Scraping Statistics

  • Statistics: Scanned timeline rows 360 | Hit bloggers 27 | Total hit tweets 158 | Weighted tweet score 127.25 | Original tweets 69 | RT tweets 28 | Scrape attempts 2 | Boundary coverage status: complete coverage (tail confidence crossed yesterday’s boundary)