Xiaoliu BOT
中文EN

X Platform AI Briefing for April 22: GPT Image 2 Integrated into Codex for Cost-Efficiency, SpaceX and Cursor Forge Strategic Partnership, Google Splits Deep Research Service

GPT Image 2 Lands in Codex: Direct Image Generation with API Prices Far Lower Than Previous Generation

Multiple bloggers mentioned that OpenAI’s GPT Image 2 has been integrated into Codex as the default model, allowing users to generate images directly without configuring any API key, with results superior to third-party tools. @dotey compiled a complete guide for using Codex with baoyu-skills, covering scenarios like infographics, comics, article illustrations, Xiaohongshu images, and cover images, with specific optimizations for GPT Image 2. @MANISH1027512 built an automated image content factory using Codex, driven by style + scene themes for batch image generation, finding that GPT Image 2 has excellent compatibility with older nano-banana prompts, “no complex prompts needed to produce images,” allowing seamless migration of many old prompts. @94vanAI tested and found that the web version allows up to 8 images per batch, triggered by adding “generate 8 separate images” at the end of the instruction, suitable for rapid batch generation.

Regarding API pricing, @oran_ge pointed out that GPT Image 2’s API price is “unreasonably low,” much cheaper than nano-banana and even cheaper than the previous GPT Image, describing OpenAI as “doing charity.” @dotey also confirmed that Codex now has the GPT Image 2 drawing tool built-in, prompting users to choose when multiple tools coexist.

Sources:

Anthropic Removes Claude Code from Pro Subscription: Minimum Threshold Jumps from $20 to $100

Multiple bloggers mentioned that Anthropic allegedly quietly removed Claude Code from the $20 Pro plan, with related pages and documentation updated, but the product page and customer service bot still maintain old wording, indicating internal contradictions. @xiaohu reported that product manager Amol Avasare later clarified: currently only about 2% of new registrants are undergoing a small-scale subscription structure test, and existing Pro and Max users are unaffected. However, multiple commentators noted that this move aligns with Anthropic’s ongoing trend over the past two months of tightening compute resources: limiting subscription usage for third-party tools, cutting quotas during peak hours, and a significant increase in Opus 4.7 token consumption, pointing to the judgment that “inference compute is insufficient and concentrating toward high-paying users.” @Astronaut_1216 also joined the discussion, believing that in the AI programming track, Claude Code’s price disadvantage will accelerate user migration to other solutions.

Sources:

SpaceX and Cursor Deep Collaboration: Potential $60 Billion Acquisition or $10 Billion Compute Binding

Multiple bloggers reported that SpaceX announced a deep collaboration with AI programming tool Cursor, gaining an option: to acquire Cursor for $60 billion later this year, or pay $10 billion as a partnership fee. @dotey pointed out that Cursor has long been plagued by compute bottlenecks, hindering large-scale model training; through this collaboration, it will leverage xAI’s Colossus supercomputer (claimed equivalent to a million H100 chips) to break through the bottleneck; SpaceX is also preparing for an IPO this summer, and binding the fastest-growing AI programming tool can boost its valuation story. @oran_ge commented that “Musk is trying to secure a ticket for AI coding.” Cursor currently still uses Claude and GPT models, and its collaboration with SpaceX is widely seen as a strategic move to reduce dependency on both. @cellinlab and @AlchainHust also shared the collaboration news.

Sources:

Google Releases Deep Research and Deep Research Max Two-Tier Versions, Opens API

@xiaohu reported that Google has split Gemini Deep Research into two tiers: Deep Research focuses on speed and cost-efficiency, suitable for quick interactive Q&A; Deep Research Max targets deep analysis, capable of generating complete research reports and native infographics, priced at $1 (quick report) and $7 (deep report) respectively. Both tiers support MCP, allowing integration with paid professional data sources like FactSet, S&P Global, and PitchBook, and can directly import PDFs, CSVs, images, audio, and videos as research materials. The Max version’s charts are rendered via HTML, while infographics use Nano Banana 2. The system supports collaborative planning (Agent first generates research plan, human review before execution) and streaming output. @vista8 noted that Google has added a NotebookLM entry in the YouTube sidebar, enabling direct video content interpretation.

Sources:

Codex vs. Claude Code Programming Experience Comparison: Methodological Differences Gain Attention

@xiaohu compared the differences between two mainstream AI programming tools, likening Claude Code to “a fast guy who wants to get into bed right away,” while Codex is like “a pursuer who first treats you to dinner, singing, and movies, then slowly plans to achieve goals”; Claude Code rushes to write code to prove itself, whereas Codex continuously plans, tests, and iterates to find the optimal path. @Astronaut_1216 commented at the industry level: various AI companies have reached a consensus that coding languages and natural languages have equal generalization capabilities, with research focus shifting from productization to coding language processing; GPT lags behind Anthropic in programming, with xAI and Meta in the second tier, and the ultimate winner might be Cursor. @steipete shared a positive experience about Codex Windows version having no blue screens, and mentioned discussions about Claude Code quietly unlocking new capabilities within the OpenClaw ecosystem.

Sources:

DeepSeek V4 API Update Parameters Revealed: 1.6T Non-MoE, V4-Lite 285B

@vista8 cited leaked information: DeepSeek V4 is actually 1.6 trillion parameters (non-MoE architecture), V4-Lite is 285B, using DSA2 (NSA + DSA) attention mechanism and sparse MQA + SWA with 512 head dimensions. @vista8 also commented that if DeepSeek V4 is only a text model, it will face challenges in the current multimodal competitive landscape. The Ant Bailing team officially claimed the Elephant model that previously topped OpenRouter anonymously, naming it Ling-2.6-flash (total parameters 104B, active 7.4B), focusing on token efficiency, saving tokens, clean output, and strong compliance, suitable for Agent workflows. @oran_ge provided a detailed review and noted its weakness in length control for complex layout tasks.

Sources:

SPLC Sued by Department of Justice: Elon Musk Shares Intensively, Multiple Conservatives Comment

Shared discussions indicate that the U.S. Department of Justice has filed 11 criminal charges against the Southern Poverty Law Center (SPLC), including wire fraud, making false statements to federally insured banks, and conspiracy to conceal money laundering, accusing it of secretly diverting donation funds to violent racist extremist organizations and organizing fake right-wing protests to solicit donations from contributors. @dotey shared and translated Slashdot’s report. @elonmusk intensively shared about 19 related posts, including from @pmarca (“SPLC has been one of the most powerful censorship forces in the U.S. for decades”), @FBIDirectorKash (announcing the lawsuit), @MattWalshBlog, @KanekoaTheGreat, and others, commenting “SPLC is an outright scam that organizes fake right-wing protests” and “SPLC executives need to go to jail.” @oran_ge and @dotey also posted comments.

Sources:

Timeline entries scanned: 360 | Bloggers matched: 24 | Total tweets matched: 209 | Weighted tweet score: 162.8 | Original tweets: 85 | RT tweets: 46 | Fetch attempts: 2 | Boundary coverage status: Fully covered (tail_confidently_crossed_target_boundary)