Twitter AI Coding -- 2026-04-22

1. What People Are Talking About

1.1 Codex Model Leak Reveals GPT-5.5, Growth Hits 4M Users 🡕

The day's most discussed Codex story came from an accidental exposure of unreleased models in the Codex model picker. @ai_for_success posted (65 likes, 4,424 views): "OpenAI just 'accidentally' leaked all the internal models in the Codex model picker. Reality is, nothing gets leaked accidentally from AI labs."

The screenshot shows GPT-5.5 at the top of the list alongside oai-2.1, arcanine (selected), and multiple glacier-alpha variants. @WesRoth added context (14 likes, 634 views), calling out "glacier-alpha, Heisenberg, and Arcanine" as cryptic experimental codenames. A reply from @rahulbetgeri confirmed: "for a very very short while I saw 5.5!"

Separately, @shiri_shh reported (16 likes, 1,519 views): "Codex just hit 4 MILLION active users, less than two weeks after 3 million. And they are also DROPPING GPT 5.5 this WEEK." This growth figure is corroborated by an ARK Investment Management chart shared by @richardseiler showing (47 likes, 250 views) Codex weekly active users grew 20X year-to-date, from 0.2M in January to 4.0M by April 21.

@richardseiler noted: "Openai's Codex app has been lagging the explosive progress from Claude Code but as you can see the usage stats are climbing steadily for Codex since Jan this year."

Comparison to prior day: On April 20 the Codex story was about the desktop app receiving its highest-ever engagement (1182.2 score) and Chronicle launching ambient context. Today the narrative shifts to scale: 4M active users, a leaked model pipeline suggesting GPT-5.5 is imminent, and Codex-powered workspace agents expanding into ChatGPT (see 1.2).

---

1.2 GitHub Copilot Token-Based Billing Confirmed for June 🡕

@edzitron published an exclusive (101 likes, 7,416 views) with internal Microsoft documents revealing specific pricing: "Copilot Business Customers will pay $19 per-user-per-month and receive $30 of pooled AI credits. Copilot Enterprise customers will pay $39 per-user-per-month and receive $70 of pooled AI credits." The announcement was expected on April 23, with changes rolling out in June.

A reply from @edzitron added: "the exact amount of AI credits is still up in the air, and may change before launch." It remains unclear what happens to individual subscribers.

@SantoshYadavDev captured the sentiment (4 likes, 684 views): "2024-2025 was the best time to use AI coding tools, now it will be only getting expensive. GitHub Copilot stops new signups and removes Opus 4.6. Now Anthropic stops Claude code in the $20 plan. Be ready to pay more."

Comparison to prior day: On April 20, edzitron first reported that token-based billing was a "top priority" and that weekly Copilot costs had doubled since January. Today the story advances with specific dollar amounts from internal documents and a confirmed timeline. The subsidy era is ending with concrete numbers attached.

---

1.3 Ex-Codex Engineer Declares "Software Is Solved," Raises $12M for Hardware 🡕

The day's highest-scoring post (2152.3) came from @DanielEdrisian announcing (701 likes, 44,010 views): "I've left OpenAI and the Codex team to build Blackstar: A new hardware company building the future of human-computer interaction. We believe that software is solved. Building apps is now easy, but the next meaningful improvement in human-AI communication requires changing the OS & hardware."

The $12M seed round was led by Abstract Ventures, with participation from Naval Ravikant, SV Angel, Chapter One, and Timeless. The team of 8 spans SF and Shenzhen, working on hardware, software, and model training.

This post is notable not just for its engagement (240 bookmarks) but for its thesis: someone who built Codex from the inside now views software creation as a solved problem and is moving to the hardware layer. This is the most concrete "post-software" signal from a practitioner to date.

---

1.4 OpenAI Workspace Agents Launch, Codex Expands Beyond IDE 🡒

@9to5mac reported (31 likes, 4,846 views): "OpenAI updates ChatGPT with Codex-powered 'workspace agents' for teams." The article confirms workspace agents replace Custom GPTs, are powered by Codex technology, run in the cloud for long-running workflows, and are available for Business, Enterprise, Edu, and Teachers plans.

@daniel_mac8 noted (21 likes, 836 views): "interesting that it's in ChatGPT and not Codex" -- pointing to the strategic expansion of Codex technology beyond the dedicated coding app into the broader ChatGPT ecosystem.

This follows Chronicle (April 20) and the major desktop update (April 16), making it the fourth significant Codex product release in one week.

---

1.5 SpaceX-Cursor $10B Deal Reshapes Competitive Landscape 🡕

@9to5mac reported (40 likes, 6,455 views): "SpaceX lands deal to likely purchase Claude Code and OpenAI Codex competitor." The deal gives Cursor access to xAI's Colossus supercomputer for model training, with a $10B payment that could escalate to a $60B acquisition.

@Agent_Joby questioned the logic (7 likes, 626 views): "Why pay $60B for a Cursor IDE forked from Visual Studio Code?" Replies on the 9to5mac post from @codyavant, @tymphillips, and @MaxRovensky all called out the article's headline as clickbait.

The deal positions Cursor with significant compute resources for custom model training, potentially breaking its dependence on third-party LLM providers. This adds a third well-capitalized player alongside OpenAI (Codex) and Anthropic (Claude Code).

---

1.6 Google Antigravity: Security Threats Mount, Mindshare Fades 🡖

Antigravity drew attention for security concerns rather than product progress. @DarkReading reported (6 likes, 1,303 views): "Google Fixes Critical RCE Flaw in AI-Based Antigravity Tool." Separately, @ransomnews warned (3 likes, 120 views) of "Zero-detection malware hidden in fake Google tool" -- trojanized Antigravity downloads that drop infostealers hijacking browser sessions and tokens.

@SecurityWeek added (1 like, 741 views): "Google Antigravity in Crosshairs of Security Researchers, Cybercriminals."

Meanwhile, mindshare erosion continued. @BenjaminDEKR asked (17 likes, 4,018 views): "Do you guys remember Google Antigravity" -- a sardonic take that drew 15 replies. @scientistbio1 replied: "They have nerfed Antigravity." @DumbEinstein added: "it seems to lose any kind of advertising or promotion after a week or two."

@teortaxesTex admitted (108 likes, 8,452 views): "man do I feel dumb for shilling Jules and Antigravity... Classic Google to ruin it." The quoted tweet from @Presidentlin noted: "The sentiment for Claude Code has become negative. Tool that is actively fighting against you, not with you, is the general vibe."

Comparison to prior day: On April 20, the Antigravity story was about a paying customer banned with no recourse (documented by GergelyOrosz, 113 likes). Today the narrative shifts from trust issues to active security threats -- a patched RCE flaw and trojanized downloads targeting users. The product is now associated with both operational and security risk.

---

1.7 GitHub CLI Telemetry Ships to Track AI Agent Usage 🡕

@everyonebpup documented (8 likes, 570 views, 5 bookmarks) that GitHub CLI v2.91.0 added telemetry: "every gh command you run now sends a ping to github: command name, flags, OS, version, a device_id, and a field called 'agent.'"

The "agent" field tracks which AI tool -- Claude Code, Copilot CLI, Cursor, or Codex -- drove each CLI invocation. Opt-out is available via GH_TELEMETRY=false, DO_NOT_TRACK=true, or gh config set telemetry disabled. A reply from @lmkifiwin argued: "they are not tracking you. They are tracking which AI tool drives the most CLI usage." @everyonebpup countered: "agent field tracks which agent ran command. other 8 fields track you."

The telemetry launch confirms that AI agent usage of developer toolchains is now significant enough for GitHub to measure systematically.

---

2. What Frustrates People

GitHub Copilot Pricing Uncertainty Deepens -- High

@edzitron revealed (101 likes, 7,416 views) specific token-based billing tiers coming in June, but the individual subscriber fate remains "unclear." @SantoshYadavDev warned (4 likes, 684 views): "Be ready to pay more." The combination of paused signups (April 20), removed models, and now imminent billing changes leaves users unable to plan. Coping mechanism: stacking cheaper alternatives like OpenCode Go ($10) and Ollama subscriptions.

Vibe Coding Quality and Review Gap -- High

@itsolelehmann identified (2 likes, 427 views): "the dirty secret of vibe coding in 2026: most ai-written code is going live without anyone actually reading it. We're shipping 100x faster with claude code, cursor, codex. But the review layer hasn't kept pace at all." @serpinxbt connected this to security (21 likes, 447 views): "vibe coding drastically lowers barrier of entry to scammers." A reply from @walsxbt asked: "Need a vibe coded app for checking the safety of other vibe coded apps. Who's building this?"

Claude Code Usage Limits Perceived as Restrictive -- Medium

@RasputinKaiser posted (0 likes, 13 views) a screenshot showing Claude Code usage with the comment: "The usage limits are limitless! Anthropic is simultaneously compute poor and money hungry with Claude and Claude Code." @Presidentlin described Claude Code sentiment as: "Tool that is actively fighting against you, not with you."

Copilot Reviews Wasting Compute on Auto-Merged PRs -- Low

@__alula discovered (56 likes, 1,921 views) that GitHub runs Copilot code reviews on every DMCA takedown PR from dmca-sync-bot, which are auto-merged immediately.

A reply from @asdhuqiwuhdais suggested: "It's probably an organization default policy."

---

3. What People Wish Existed

Code Review Layer That Matches AI Build Speed

@itsolelehmann stated (2 likes, 427 views): "the review layer hasn't kept pace at all, so most ai-written code is going live without anyone actually reading it." @serpinxbt added (21 likes, 447 views) that this creates a direct security surface for scams. The demand is for automated review tooling that scales with AI-generated code volume.

Distribution and Marketing Layer for Vibe-Coded Products

@Parul_Gautam7 identified (8 likes, 210 views) the gap: "Vibe coding solved building, but distribution was still the missing layer. Founders can ship SaaS in days but turning it into marketing-ready content is still a bottleneck." She pointed to Higgsfield Marketing Studio as an emerging solution.

Unified Analytics Across AI Coding Tools

@tom_doerr shared (3 likes, 337 views) Agentlytics, an open-source dashboard that unifies sessions from 16 editors including Cursor, Windsurf, and Claude Code. The tool runs 100% locally and tracks sessions, costs, models, and tools.

Better Model Management for Non-Technical Users

@pfanis asked twice (2 likes, 50 views): "Why do we have to do a new update every time there is a new model? Isn't there any way to make this a bit easier? Sorry, guys, not technical here." The question captures friction for the growing non-engineer user base trying to add new models to OpenCode.

---

4. Tools and Methods in Use

Tool	Category	Sentiment	Strengths	Limitations
OpenAI Codex	Agent platform	(+)	4M weekly active users (20X YTD per ARK); workspace agents launch in ChatGPT; GPT-5.5 visible in model picker	Model leak suggests rushed pipeline; individual pricing unclear under token billing
GitHub Copilot	Cloud IDE agent	(+/-)	Still dominant by installed base; CLI telemetry confirms agent-driven usage is significant	Token-based billing in June ($19/mo for $30 credits); signups still paused; Opus removed from Pro
Claude Code	Terminal agent	(+)	19 mentions in dataset; claude-reflect adds self-learning (922 stars); coding standards enforcement praised	Usage limits frustrate power users; $20 plan access removed per SantoshYadavDev
Google Antigravity	IDE	(-)	ClickHouse MCP server integration at Google Next; JulianGoldieSEO continues course content (2hr, 4hr)	Critical RCE patched; trojanized downloads in the wild; mindshare fading ("do you guys remember Antigravity")
Cursor	IDE	(+)	SpaceX $10B deal gives access to Colossus supercomputer for model training	$60B acquisition valuation questioned; dependent on deal outcome
OpenCode	Open-source terminal agent	(+)	Praised as "solid" at $10; party event draws community enthusiasm	Model update process frustrates non-technical users; Ollama sub considered better value
Google AI Studio	Vibe coding	(+)	Builds landing pages live in browser from plain English	Dismissed by some as "vibes without the logic"

---

5. What People Are Building

Project	Builder	What it does	Problem it solves	Stack	Stage	Links
Blackstar	@DanielEdrisian	New hardware device for human-AI interaction	"Software is solved" -- next improvement needs OS and hardware changes	Hardware, software, model training	Seed ($12M, Abstract Ventures)	Announcement
Workspace Agents	OpenAI	Codex-powered agents in ChatGPT for teams	Custom GPTs lacked long-running workflow capability	Codex, ChatGPT	Research preview (Business/Enterprise)	9to5mac
claude-reflect	@BayramAnnakov	Self-learning system for Claude Code with permanent memory	Agent forgets corrections across sessions	Claude Code hooks, CLAUDE.md	Shipped (922 stars, v2.6.0)	Repo
Agentlytics	@f	Unified analytics dashboard across 16 AI editors	Sessions scattered, no cost tracking, no cross-editor comparison	Node.js/Deno, local-only	Shipped (npm v0.2.12)	Repo
PixiJS Agent Skills	@PixiJS	AI skills for WebGL/WebGPU game development	AI agents lack PixiJS v8 context for game builds	Agent Skills format, 40+ agents	Shipped	Repo
Google Cloud Agent Skills	@JackWoth98	13 official skills for Google Cloud services	Agents lack GCP context	Gemini CLI, Codex, Claude Code, OpenCode	Shipped	Announcement
Higgsfield Marketing Studio	@higgsfield	UGC ad creation from product links via Hermes Agent	Distribution bottleneck for vibe-coded products	Hermes Agent	Shipped	Referenced
Open SaaS	wasp-lang	SaaS starter template	Scaffolding for vibe-coded products	Wasp	Shipped (14K GitHub stars)	Referenced
Wave-Driven Development	@asm3r96	Skill for Codex/Claude Code that structures work into planning waves	Unstructured agent workflows drift	Codex, Claude Code	Alpha (personal use)	Post

---

6. New and Notable

GitHub Copilot Token-Based Billing: Specific Numbers Emerge

@edzitron published (101 likes, 7,416 views) internal Microsoft documents showing Business at $19/user/month for $30 of pooled AI credits and Enterprise at $39/user/month for $70 of pooled AI credits, with announcement expected April 23 and June rollout. This advances the April 20 reporting from directional signals to specific pricing tiers. Individual subscriber plans remain undefined.

Codex Workspace Agents Replace Custom GPTs

OpenAI launched (31 likes, 4,846 views) workspace agents in ChatGPT, powered by Codex technology. Available for Business, Enterprise, and Edu plans. Agents run in the cloud, handle long-running workflows, and are shareable within organizations. This extends Codex from a dedicated coding tool into a general-purpose enterprise automation layer.

SpaceX-Cursor $10B Deal With Path to $60B Acquisition

SpaceX and Cursor announced (40 likes, 6,455 views) a partnership giving Cursor access to xAI's Colossus supercomputer for model training. The $10B payment could lead to a $60B acquisition. If completed, this would create a vertically integrated competitor with its own compute, models, and IDE distribution -- a structure neither OpenAI (Codex) nor Anthropic (Claude Code) currently has.

Google Cloud Publishes Official Agent Skills for Competitor Platforms

@JackWoth98 announced (4 likes, 106 views) 13 official Google Cloud Agent Skills compatible with Gemini CLI, Codex, Claude Code, and OpenCode. This is notable as Google explicitly supporting competitor agent platforms with first-party cloud service skills.

PixiJS Launches Agent Skills for Game Development

@PixiJS shipped (8 likes, 165 views) AI skills teaching agents correct PixiJS v8 usage for WebGL/WebGPU builds. Compatible with 40+ agents via the Agent Skills format. This is the first major rendering library to publish dedicated agent skills, signaling the skills ecosystem is expanding beyond cloud infrastructure into creative tooling.

Coinbase Reports 27x Increase in Non-Engineer Dev Tool Usage

@rwitoff (Coinbase) shared (15 likes, 310 views): "In the last 12 months, we've seen a 27x increase in non-engineers using dev tools like Claude, OpenCode and Cursor to build & automate how we work."

A reply from @rgivon noted: "every employee has become a team. Coordination or sharing are not a primitive anymore."

---

7. Where the Opportunities Are

[+++] AI Code Review at Scale -- @itsolelehmann named the core problem: "most ai-written code is going live without anyone actually reading it." @serpinxbt connected this to scam risk. The Antigravity security stories (RCE flaw, trojanized downloads) show the attack surface is already being exploited. A review tool that matches the throughput of AI code generation -- scanning for vulnerabilities, logic errors, and credential exposure at the speed code is being shipped -- addresses a gap that widens with every productivity improvement.

[+++] Subscription Cost Intelligence -- Token-based billing is now confirmed with specific tiers (@edzitron, article). @LottoLabs described manually evaluating OpenCode Go at $10 vs GPT $20 vs Ollama subscriptions. A tool that monitors actual token consumption across Copilot, Codex, Claude Code, and OpenCode, recommends optimal plan allocation, and auto-routes requests to the cheapest capable model would serve a structural need as the subsidy era ends.

[++] Agent Skills Ecosystem Tooling -- Google Cloud published 13 official agent skills. PixiJS shipped rendering skills for 40+ agents. @asm3r96 built wave-driven development as a personal skill. The skills format is gaining adoption but discovery, quality assurance, and composition tooling does not yet exist. A skills marketplace, registry, or recommendation engine for agent harnesses would accelerate this emerging ecosystem.

[++] Post-Build Distribution for AI-Created Products -- @Parul_Gautam7 identified that "vibe coding solved building, but distribution was still the missing layer." Open SaaS reached 14K GitHub stars through the vibe coding boom, showing demand for scaffolding. Tools that bridge from shipped product to marketing-ready content -- landing pages, UGC ads, SEO -- serve the bottleneck that now sits downstream of code generation.

[+] Non-Engineer Onboarding for Dev Tools -- Coinbase reports 27x growth in non-engineers using dev tools. @pfanis asked why adding a new model requires a manual update -- "not technical here." The gap between developer-oriented tooling and the rapidly growing non-engineer user base creates opportunity for simplified setup, guided workflows, and managed model configuration.

---

8. Takeaways

1. Codex is scaling faster than any competitor and expanding beyond coding. 4M weekly active users (20X YTD per ARK data), a leaked model pipeline showing GPT-5.5 and experimental codenames, and workspace agents launching in ChatGPT for teams. OpenAI is positioning Codex not as a coding tool but as general-purpose enterprise automation infrastructure. (@richardseiler, chart; @9to5mac, workspace agents)

2. GitHub Copilot token-based billing now has specific dollar amounts and a June deadline. Business: $19/user/month for $30 of pooled credits. Enterprise: $39/user/month for $70 of pooled credits. Individual plans remain undefined. Combined with the April 20 signup pause and model removals, this completes the picture of subsidized AI coding ending. (@edzitron, exclusive)

3. The SpaceX-Cursor deal creates a third vertically integrated competitor. $10B with a path to $60B acquisition gives Cursor access to xAI's Colossus supercomputer for model training. If completed, Cursor would have its own compute, custom models, and IDE distribution -- a structure OpenAI and Anthropic lack. (@9to5mac, report)

4. An ex-Codex engineer betting $12M that "software is solved" is the strongest post-software signal yet. @DanielEdrisian's departure from OpenAI to build hardware (701 likes, 240 bookmarks, 44K views) reflects the view from inside the most-used coding agent: the software creation bottleneck is gone, and the next constraint is the interface layer. (@DanielEdrisian, announcement)

5. Google Antigravity is now associated with security risk, not just trust issues. A patched critical RCE flaw, trojanized downloads dropping session-hijacking malware, and continued mindshare erosion ("Do you guys remember Google Antigravity"). The narrative has shifted from "shipping fast but banning users" to "shipping fast while users face active threats." (@DarkReading, RCE; @ransomnews, malware)

6. The agent skills ecosystem is formalizing fast. Google Cloud published 13 official skills for competitor platforms. PixiJS shipped the first major game engine skills. Individual developers are creating workflow skills. The shift from "agents with tools" to "agents with curated domain knowledge" is the emerging infrastructure layer. (@JackWoth98, Google skills; @PixiJS, skills)

7. The code review gap is the most dangerous unaddressed problem in AI coding. Code is shipping 100x faster, but review has not scaled. Trojanized Antigravity downloads and vibe-coding-enabled scams demonstrate the attack surface is already being exploited. The distance between AI code generation speed and human review capacity is a systemic vulnerability. (@itsolelehmann, review gap; @serpinxbt, scam risk)