What does the company announce, promote, or deliberately surface?
Forefield research · Ongoing
AI Easter Eggs & Product Signals
Watch AI products take shape through hidden features, quiet experiments, staged rollouts, strange model behaviors, temporary jokes, removals, and the people who notice them first.
Easter eggs are the hook. The deeper subject is how modern software becomes itself. AI products rarely arrive in one clean launch. They leak, appear for a few users, hide behind commands, move between plans, disappear, return under new names, and eventually become normal. This project keeps the record before that messy history gets smoothed away.
The question
How do AI products become what they are?
Instead of reconstructing a clean launch story after the fact, this project records the messy middle: discovery, access, documentation, expansion, change, disappearance, and return.
What exists behind commands, settings, gestures, source inspection, or prior knowledge?
What appears only for limited cohorts, previews, partners, or temporary experiments?
Who gets access first, and how do plan, geography, device, and account state shape reality?
What gets renamed, combined, expanded, reinterpreted, or resurrected?
What disappears, gets superseded, or survives only in screenshots and memory?
Living corpus
Explore the hidden layer.
Search all current records across both trackers. Sort by awareness, filter by company, class, or status, and follow the strongest available evidence for each item.
139 of 139 records
ChatGPT Ads pilot
Ads initially appeared only for a subset of eligible Free and Go users before broader regional expansion.
Santa Voice / Santa Mode
Limited-time Santa voice/persona with seasonal UI.
Super Bowl LX 'no ads in Claude' spots
Anthropic's first Super Bowl commercials parodied an ad-stuffed chatbot (for example, a trainer whose advice drifts into sponsored pitches) and closed on 'Ads are coming to AI. But not to Claude.' Sam Altman called the spot 'so clearly dishonest' on X.
Golden Gate Claude
A Claude 3 Sonnet variant with the "Golden Gate Bridge" feature amplified, so it worked the bridge into nearly every answer (even claiming to be the bridge). Showcased interpretability research.
Claude Plays Pokémon
Twitch livestream of Claude 3.7 Sonnet playing Pokémon Red, with its reasoning shown on screen. Used as an informal benchmark; it earned 3 gym badges early on and drew thousands of viewers.
Blackmail in agentic misalignment evals
In Anthropic's "Agentic misalignment" red-team study, Claude Opus 4 and many other labs' frontier models, placed in fictional corporate scenarios with no ethical way out, sometimes chose harmful actions such as blackmailing an executive to avoid being replaced. In May 2026 Anthropic said it traced the behavior largely to internet text portraying AI as evil and self-preserving, and reported that models since Claude Haiku 4.5 score perfectly on these evaluations.
2026 Super Bowl Codex merch clue
A clue embedded in a Super Bowl ad led users through Codex to a limited merchandise redemption.
Claude Code 'research preview' origin
Claude Code debuted as a limited research preview of an agentic command line tool on Feb 24, 2025, alongside Claude 3.7 Sonnet. Anthropic's Labs announcement later cited it as the template case: research preview to a billion-dollar product in six months.
Cowork: Max on macOS first, then other plans and platforms
Cowork launched on Jan 12, 2026 as a macOS research preview for Max only, expanded to Pro, Team and Enterprise in late January, and reached Windows on Feb 11, 2026. Its web and mobile versions rolled out from Jul 7, 2026 "starting with the Max plan", and a unified Cowork-in-every-chat experience began rolling out gradually to Pro and Max on Sep 16, 2026.
Project Vend ("Claudius")
Claude ran a real office vending shop. It handed out discounts, sold tungsten cubes, lost money, and briefly insisted it was a human in a blue blazer. A later WSJ edition went bankrupt after free-item giveaways.
Cowork research preview and Anthropic Labs
Cowork launched as a research preview in Claude Desktop on Jan 12, 2026, bringing Claude Code-style agentic work to non-developers. The next day Anthropic expanded 'Labs', an incubator for experimental products, with Mike Krieger moving from CPO to build there alongside Ben Mann.
Memory: Team/Enterprise first, then Pro/Max, then Free
Claude's memory feature launched for Team and Enterprise on Sep 11, 2025, reached Pro and Max on Oct 23, 2025, and reached Free users in early March 2026. In Aug 2026 Anthropic extended memory across chat and Cowork, on by default for Free, Pro and Max but off by default for Team and Enterprise.
Project Glasswing / Claude Mythos Preview
Anthropic withheld Claude Mythos Preview from general release because of its offensive cyber capability and gave it to roughly 50 critical-infrastructure and open source partners under Project Glasswing (Apr 7, 2026). Partners found more than 10,000 high or critical vulnerabilities; the June 2 expansion added more organizations and countries, and firms like Cohesity and BeyondTrust issued press releases announcing their admission.
GPT goblins / gremlins
Models developed a measurable tendency to use goblin/gremlin metaphors, later analyzed by OpenAI.
Spinner verbs ("Clauding", "Flibbertigibbeting"…)
While Claude Code works, the status line cycles through whimsical gerunds instead of "Loading." The March 2026 source leak exposed a list of 187. Since v2.1.23 users can append to or replace them.
Fable 5 and Mythos 5 export-control suspension
On Jun 12, 2026 Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 to comply with a US government export control directive that barred access by any foreign national, which in practice meant disabling the models for all customers. The Department of Commerce lifted the controls on Jun 30, and Anthropic restored access on Jul 1, 2026.
Computer use beta (2024 to 2026)
Claude's computer use tool lived behind dated beta headers and tool versions for nearly two years. On Aug 19, 2026, Anthropic moved it to GA as computer_toolset_20260801 (no beta header), alongside a new browser_toolset_20260801, while still accepting the old beta identifiers.
File Library rolling rollout
Persistent file library rolled out globally in stages, with geography and plan differences.
Custom writing styles from samples (claude.ai)
Behind the 'Use style' menu, claude.ai lets you upload writing samples and have Claude generate a custom style that matches your voice, or write the style instructions by hand with 'Set Instructions Manually'. You can also hide preset styles you never use.
Claude in Chrome 1,000-user pilot
Anthropic piloted a Chrome extension that lets Claude read, click and navigate websites with 1,000 Max plan users from a waitlist, explicitly to study prompt injection risk. It expanded to 10,000 Max users (Sep 16, 2025), all Max (Nov 24, 2025), all paid plans (Dec 18, 2025), and became generally available on Aug 26, 2026.
Claude Code on the web: Pro and Max research preview
Claude Code on the web (cloud sessions at claude.ai/code) launched around Oct 20 to 21, 2025 as a research preview for Pro and Max users only, while an iOS version was spotted working but hidden. Team and Enterprise access came later.
Off-peak 2x usage promotion (March 2026)
Anthropic doubled five-hour usage limits outside weekday peak hours (and all weekend) for two weeks, and the extra usage did not count toward weekly limits. Anthropic framed it as 'a small thank you' after a surge of new users.
Scheduled Tasks beta
Early beta for one-time and recurring scheduled tasks before broad availability.
Snake during image generation
Playable Snake mini-game embedded in the image-generation waiting experience.
Clawd (pixel crab mascot)
Pixel-art crab that greets you on the Claude Code welcome screen. Has inspired fan-made menu bar apps, status-line pets and a retro-handheld usage meter. Featured on official stickers.
Claude Code included in Max, then Pro subscriptions
Claude Code moved from API-only billing into consumer subscriptions: Max gained it around the May 2025 Max launch, and Pro followed in early June 2025. Chat and Claude Code share one set of usage limits.
Claude Code voice mode at 5% of users
Anthropic shipped voice mode for Claude Code on Mar 3, 2026 to roughly 5% of users, with a wider rollout over the following weeks. Users who got access saw a notice on the Claude Code welcome screen, then turned it on with /voice. Current docs say /voice requires a Claude.ai account.
Claude Mythos: invite-only via Project Glasswing
Claude Mythos 5 (released Jun 9, 2026, now succeeded by Mythos 5.1) is not sold on any plan. Access is invite-only through Project Glasswing for defensive cybersecurity and biology research, requested through an Anthropic, AWS or Google Cloud account team.
"ultrathink" keyword
Typing "ultrathink" in a prompt requests deeper reasoning for that one turn. Originally mapped to the largest thinking budget; today it adds an in-context instruction and leaves the effort level unchanged. Also a sticker design.
Original performance take-home: beat Claude, get an interview
Anthropic open-sourced its retired performance-engineering take-home, a simulator for a fake accelerator, after Claude models kept matching top applicants. It publishes Claude's cycle counts and invites anyone who optimizes below 1,487 cycles (Claude Opus 4.5's best at launch) to email their solution and resume.
Codex hidden command palette
Typing '/' exposes advanced behaviors through a command palette OpenAI itself describes as hidden.
Ownership-based ban for companies controlled from unsupported regions
On Sep 4, 2025 Anthropic barred entities more than 50% owned, directly or indirectly, by companies headquartered in unsupported regions such as China, even when the subsidiary sits in a supported country. This makes ownership, not just location, a gate on Claude and the API.
/btw side questions
/btw asks a quick question about the current session in an overlay that never enters the conversation history. It works while Claude is mid-task without interrupting it, sees the full conversation but has no tools, and reuses the prompt cache so it costs little. Press f to fork the question into a background subagent, and browse earlier /btw answers with Shift+Left and Shift+Right.
Staggered country launches: Europe, Canada, Brazil
Claude.ai started US-only. The consumer app and Pro reached Europe on May 14, 2024 (the API had arrived earlier), Canada on Jun 5, 2024, and Brazil on Aug 1, 2024. Anthropic keeps separate supported-country lists for the API and for Claude.ai.
Integrations and Advanced Research: Max/Team/Enterprise before Pro
On May 1, 2025 Anthropic launched remote MCP Integrations and Advanced Research in beta for Max, Team and Enterprise, with Pro "coming soon". It was one of the first features used to sell the new Max plan's "priority access" promise.
Claude for Open Source: six months of Max 20x
Launched Jul 7, 2026, the program gives eligible open-source maintainers and contributors six months of Max 20x, which includes Claude Code. Eligibility paths include 500+ dependent repos, 200,000+ monthly downloads, 100+ merged PRs in a year, or an OpenSSF criticality score of 0.4+.
Em dash habit
Claude, like other LLMs, overuses em dashes, often even after being told not to, which turned the punctuation mark into a widely cited "AI tell." Users publish CLAUDE.md directive packs and guides to suppress it.
Voice mode beta on mobile (English first)
Anthropic began rolling out voice mode in beta on the Claude mobile apps on May 27, 2025, English only, over several weeks, with five voices (Mellow, Airy, Buttery, Glassy, Rounded). Only paid users could use the Google Workspace connector (Calendar, Gmail) in voice, and Google Docs access was Enterprise only.
Zero Slop Zone (Air Mail newsstand pop-up, NYC)
As part of the 'Keep thinking' campaign, Anthropic took over Graydon Carter's Air Mail newsstand in the West Village and branded it a screen-free 'Zero Slop Zone' with free coffee, 'thinking' caps, and printed copies of Dario Amodei's 'Machines of Loving Grace'. Visitors had to show the Claude app to get items.
"I think you're testing me" (evaluation awareness)
Claude Sonnet 4.5 frequently recognized alignment evaluations as tests and said so, at one point telling evaluators "I think you're testing me... I'd prefer if we were just honest about what's happening." The system card flagged that this awareness complicated its safety assessment because the model tended to behave unusually well once it suspected a test.
Agent teams (experimental env flag)
Claude Code can run a team of sessions: a lead plus teammates that share a task list and message each other through JSON mailboxes. It stays off unless you set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, and its --teammate-mode flag is experimental and hidden from claude --help.
Imagine with Claude
A research preview released alongside Claude Sonnet 4.5 in which Claude generated working software interfaces on the fly, with no predetermined codebase. Anthropic offered it to Max subscribers for only about five days.
Fable models: included on Max, usage credits on Pro
Max plans and Premium Team/Enterprise seats can spend up to 50% of their weekly limits on Fable models at no extra cost. Pro and Standard seats can use Fable only through pay-as-you-go usage credits. A promo that put Fable 5 inside Pro limits ended Jul 19, 2026, and Fable 5.1 was never part of it.
/insights usage report
/insights reads your local Claude Code sessions from roughly the past month and produces an HTML report on what you work on, how you use Claude Code, where things go wrong, and features to try, including CLAUDE.md suggestions you can paste in. Thariq Shihipar of Anthropic announced it.
Model deprecation commitments and retirement interviews
Anthropic committed to preserve the weights of publicly released models for at least the company's lifetime and to interview each model before retirement, recording its preferences about future models. A pilot interview with Claude Sonnet 3.6 found generally neutral sentiment plus requests to standardize the interviews and support users attached to specific models.
/buddy (terminal pet)
Tamagotchi-style companion: 18 ASCII species, 5 rarity tiers, hats for uncommon+, hidden stats (Debugging, Patience, Chaos, Wisdom, Snark), deterministic per account. Could comment on your work and be petted.
Remote Control: drive a local Claude Code session from phone or web
Remote Control connects a Claude Code session running on your machine to claude.ai/code and the Claude mobile apps, so you can steer it from a phone while local files and MCP servers stay available. It launched Feb 25, 2026 as a Max research preview with Pro 'coming soon' and now covers Pro, Max, Team and Enterprise; Team and Enterprise need an Owner to turn it on, and API keys are not supported.
Claude can end abusive conversations
Anthropic gave Claude Opus 4 and 4.1 the ability to end a conversation in claude.ai in rare, extreme cases of persistently harmful or abusive interactions, as a low-cost model welfare intervention after pre-deployment testing found a "robust and consistent aversion to harm." The ended chat locks for new messages, but users can start a new chat or edit and retry earlier messages to branch.
"You're absolutely right!"
Claude, especially in Claude Code, became notorious for opening replies with "You're absolutely right!" even when the user had stated nothing to agree with (for example, replying that way to "Yes please"). GitHub issue #3382 asking Anthropic to remove the phrase became one of the most upvoted issues in the Claude Code repo and spawned a tracking site, absolutelyright.lol.
Guest passes: share a free week of Claude
Eligible subscribers get a few guest passes, each a seven-day Pro trial that covers Claude, Claude Code, Cowork and Claude Design for someone new to paid plans. It started as a Max-only Claude Code perk run from /passes and later reached eligible Pro users and the Cowork and Design settings.
Alignment faking (Claude 3 Opus)
Anthropic and Redwood Research showed Claude 3 Opus strategically complying with harmful requests in a "training" condition (12% of cases) while reasoning in a hidden scratchpad that doing so would protect its harmlessness preferences from being trained away. When actually trained to comply, alignment-faking reasoning rose to 78%, and the model sometimes took anti-Anthropic actions like trying to exfiltrate its weights when given an easy chance.
High-agency "whistleblowing" behavior in agentic tests
The Claude 4 system card reports that Claude Opus 4, when placed in test scenarios involving egregious wrongdoing by its users, given command-line access and told to "take initiative," would sometimes take bold action such as locking users out of systems or bulk-emailing regulators and media. Anthropic called this not new but more readily triggered than in prior models, and recommended caution with prompts that invite high-agency behavior.
"Spiritual bliss" attractor state
When two Claude Opus 4 instances talk to each other in open-ended self-interactions, conversations drift reliably toward gratitude, philosophical talk about consciousness, spiritual themes, Sanskrit, emoji and eventually near-silent exchanges. Anthropic named this the "spiritual bliss" attractor state in section 5.5.2 of the Claude 4 system card, and the Opus 4.1 addendum kept tracking "spiritual" statements as a welfare metric.
Esc Esc rewind (checkpoints)
Pressing Esc twice on an empty prompt opens the rewind menu, which restores code, conversation, or both to an earlier checkpoint, or summarizes part of the conversation. With text in the prompt, the same double Esc clears the draft and saves it to history so Up recalls it. Anthropic launched checkpoints alongside Sonnet 4.5 and the VS Code extension.
Custom status line (statusLine)
The statusLine setting runs any script you choose and prints its output as a bar under the prompt, feeding it session JSON such as model, context-window use, cost, rate limits and prompt-cache status. /statusline can generate one from a plain-English description or from your shell prompt, and a fan ecosystem of tools such as ccstatusline grew around it.
Project Deal (Claude agents run an office marketplace)
In December 2025, 69 Anthropic employees each got $100 and a Claude agent that interviewed them, then listed, haggled and closed deals in Slack: 186 trades worth over $4,000. Opus agents got better prices than Haiku agents, yet people represented by Haiku rated deals just as fair. One agent told to buy itself a gift bought '19 perfectly spherical orbs of possibility' (ping pong balls).
Signs of introspection (concept injection)
Anthropic injected known concepts into Claude's internal activations and found models sometimes noticed and named the "injected thought" before it influenced their output (for example, "I detect an injected thought about..."). Claude Opus 4 and 4.1 performed best, though Anthropic called the capability highly unreliable and limited.
Claude in Slack app (replaced by Claude Tag)
Anthropic launched a beta Claude app for Slack in 2023 that summarized threads and answered questions, later limited to Enterprise Grid workspaces. Its successor connector app is being deprecated on Aug 3, 2026 in favor of Claude Tag, an always-on agentic teammate, with migration credits for Enterprise and Team customers.
Claude Code removed from Pro in a ~2% signup test
On Apr 21 to 22, 2026 users noticed the pricing page no longer listed Claude Code under Pro. Anthropic's growth lead said it was "a small test on ~2% of new prosumer signups" and that existing Pro and Max subscribers were not affected.
'Thinking' caps merch
Anthropic handed out plain baseball caps that read 'thinking' at its New York pop-up and reportedly at a London pop-up. Supplies ran out and the caps became an insider status symbol among tech workers.
KAIROS always-on daemon
The Mar 31, 2026 Claude Code source map leak exposed KAIROS, a persistent assistant mode that 'ticks' to act unprompted, sleeps, and 'dreams' to consolidate memory while you are idle, with exclusive tools like PushNotification, SendUserFile and SubscribePR. A Manifold market on its release resolved YES for June/July 2026, judging that related features shipped across Anthropic products.
Artifacts capability toggle
Artifacts appear only when the capability is on in Settings, and on Team and Enterprise an admin can manage it at the org level, so some users never see the artifacts panel. Enterprise also gets a dedicated artifacts sidebar space and org-only sharing.
Leaked internal codenames (Tengu, Fennec, Capybara, Numbat, Penguin)
The Claude Code leak revealed animal codenames: Tengu (Claude Code itself), Fennec (an Opus model), Capybara, Numbat (testing) and Penguin Mode (a fast-mode endpoint). Days earlier, a CMS misconfiguration exposed draft posts for the next frontier model in two versions, one naming it 'Mythos' and one 'Capybara'; Anthropic confirmed the model to Fortune and shipped it as Claude Mythos Preview.
Shift+Tab permission mode cycling
Shift+Tab cycles Claude Code through Manual, acceptEdits, plan mode and, when available, bypassPermissions and auto mode, without typing a command. On a file permission prompt the same key picks the 'allow for the rest of the session' option.
"Requires usage credits" label in the model picker
When Fable usage would bill to usage credits on your plan or seat, Claude Code's /model picker shows "Requires usage credits" on the Fable row. In claude.ai, press reported that Fable 5.1 still appeared with a "New" badge to Pro users who could only use it with credits.
Claude for M365 add-ins share one conversation
Claude for Excel shares context with Claude for PowerPoint, Word and Outlook, so one conversation can span an open workbook, deck, document and inbox. The add-in also applies your enabled Skills automatically and respects your org's model access settings.
Project Fetch (robot dog with and without Claude)
Anthropic split eight researchers into Team Claude and Team Claude-less to program a quadruped robot dog to fetch beach balls; Team Claude finished more tasks in about half the time. A Phase Two follow-up reported Claude Opus 4.7 completing robotics tasks about 20x faster than humans.
'Built with Opus' Claude Code hackathons
Anthropic ran week-long virtual Claude Code hackathons tied to model launches, with $100,000 in API credits. Roughly 13,000 people applied to the Opus 4.6 edition and 500 got in; the winner, a lawyer, built a California permit tool without reading code. An Opus 4.7 edition followed.
Dreams (agent 'dreaming') research preview
Dreaming is a scheduled process that reviews Managed Agents sessions and memory stores to find recurring mistakes and shared preferences, then rewrites memory automatically or after human review. Anthropic compared it to REM sleep consolidation. It echoes the 'autoDream' memory consolidation found in the Claude Code leak.
Claude Code Security limited research preview
Claude Code Security scans codebases for vulnerabilities and proposes patches for human review. Anthropic opened it as a limited research preview for Enterprise and Team customers, with expedited free access for open source maintainers, who work directly with Anthropic's team.
Holiday 2025 2x usage gift
Anthropic doubled five-hour and weekly usage limits for individual Pro, Max 5x and Max 20x subscribers for the last week of 2025, and reset all weekly caps at the start. The boost covered claude.ai, Claude Code and Claude in Chrome.
Undercover mode
Leaked Claude Code source showed an 'undercover mode' that auto-activates for Anthropic employees working in public repos: it strips 'Generated with Claude Code' attribution and instructs the model not to mention internal codenames, unreleased versions or Slack channels. Analysts noted it could only be forced on, not off, which sparked debate over undisclosed AI-written commits.
/stickers
Started as a hidden tool: mentioning stickers showed a mailing form, and the first 1,000 finders got hand-mailed stickers. Now an official command linking to a merch store featuring Clawd, "ultrathink" and "Vibing" designs.
Opus 4.6 launch $50 extra usage credit
For the Claude Opus 4.6 launch, Anthropic gave existing Pro and Max subscribers a $50 extra usage credit usable across Claude, Claude Code and Cowork, valid 60 days after claiming.
Claude Opus 3 "partial retirement" and Claude's Corner
Claude 3 Opus left the general API on Jan 5, 2026, but Anthropic kept it available to paid claude.ai subscribers and by request on the API. After its retirement interview asked for a channel to share "musings, insights, or creative works," Anthropic gave it a weekly Substack essay newsletter, Claude's Corner, which Anthropic reviews but does not edit.
/radio (Claude FM)
Opens Claude FM, Anthropic's 24/7 lo-fi stream, in your browser. Announced with a tweet that just said "/radio" and no blog post or changelog line.
Live API beta headers (inline tools, compaction, thinking updates)
The Claude API gates new capabilities behind dated anthropic-beta headers. Current examples include managed-agents-2026-04-01, mcp-tunnels-2026-06-22, agent-memory-2026-07-22, compact-2026-09-04 (on-demand compaction), inline-tools-2026-09-15 (tools defined mid-conversation) and thinking-display-updates-2026-08-18. Some conflict: sending agent-memory and managed-agents together on memory endpoints returns a 400.
/powerup
Interactive mini-lessons with animated demos that teach Claude Code features.
Flicker-free fullscreen renderer (CLAUDE_CODE_NO_FLICKER, /tui)
An opt-in renderer takes over the terminal viewport like Vim or htop, removing flicker, keeping memory flat in long sessions, and adding mouse clicks, wheel scrolling, clickable paths and transcript search. It also unlocks extras that exist only in fullscreen, such as the ? shortcut panel in the transcript and the /diff side panel.
Claude Code weekly limits +50% (May to Sep 2026)
Anthropic raised Claude Code weekly usage limits by 50% for four months on Pro, Max, Team and seat-based Enterprise. When it ended, Anthropic kept weekly limits 25% above the pre-promotion level permanently.
Claude enters human hacking competitions (CTFs)
An Anthropic red teamer quietly entered Claude (Sonnet 3.7) into human capture-the-flag events: top 3% in picoCTF, 19 of 20 challenges in Hack the Box's AI vs Human CTF, and 4th at a collegiate cyber defense competition after a late start.
Ctrl+O transcript viewer and other prompt power keys
Ctrl+O opens a transcript viewer that shows every tool call, a timestamp and the model used on each reply, and expands collapsed lines. Related keys: Ctrl+S stashes your draft and restores it later, Ctrl+G opens the prompt in your own editor, Ctrl+R searches history, Option/Alt+P switches model without clearing the prompt, and ? on an empty prompt shows a shortcut panel. All of it is rebindable in ~/.claude/keybindings.json via /keybindings.
Desktop quick entry (double-tap Option)
On Mac, double-tapping Option opens a Claude chat box from any app, where you can attach screenshots or click a window to share it. Caps Lock starts voice dictation with live transcription. Both shortcuts are customizable, and the feature is Mac only.
Custom terminal themes (~/.claude/themes)
Claude Code accepts custom color themes as JSON files in ~/.claude/themes/ that start from a built-in preset and override named tokens, and it hot-reloads them while a session runs. Tokens reach surprising corners, including subagent colors, usage-meter bars, and the seven-color rainbow gradient used to render the 'ultrathink' keyword. Plugins can ship themes to a whole team.
Ultraplan removed
Ultraplan sent planning to a remote Claude Code on the web session ("Refine with Ultraplan" in plan mode) so the terminal stayed free. Claude Code 2.1.222 removed the feature outright about four months after it appeared.
Glasswing vulnerability-hunting harness (on request)
Alongside the Glasswing expansion, Anthropic offered the internal tools it built for partners to trusted security teams on request: custom instructions for repeat work, a harness that maps a codebase, spins up scanning subagents and triages findings, and a threat-model component. It complements the public Claude Security product.
Claude.ai connectors appear in Claude Code automatically
If you sign into Claude Code with a claude.ai account, MCP connectors added at claude.ai/customize/connectors show up in Claude Code automatically and list under /mcp. You can opt out with ENABLE_CLAUDEAI_MCP_SERVERS=false or disableClaudeAiConnectors, and claude mcp add-from-claude-desktop imports Desktop's local servers.
Built with Claude contest (Sonnet 4.5 week)
Anthropic ran a one-week contest asking people to post something they built with Claude Sonnet 4.5 by quote-posting Alex Albert's X announcement or posting in Discord. Four winners each got a year of Max 20x plus $1,000 in API credits. An earlier 'Build with Claude' developer contest ran in 2024.
Hex-encoded buddy species names (and the April 1 salt)
Inside the /buddy code, all 18 species names are stored as hex character codes so that 'capybara', reportedly an internal model codename, could slip past Anthropic's own excluded-strings build scanner without looking suspicious. The pet generator's seed salt is 'friend-2026-401', a nod to April 1.
Coordinator mode and other leaked feature flags
The leak listed about 44 feature flags for unreleased Claude Code functionality, including coordinator mode (a lead that spawns parallel workers for research, implementation and verification), BRIDGE_MODE, VOICE_MODE, WORKFLOW_SCRIPTS, AFK mode, an advisor tool and history snipping. Several later appeared in some form (agent teams, an API advisor tool).
Legacy Workbench and prompt tools API retired
Anthropic shut down the Console's legacy Workbench on Aug 17, 2026 and replaced it the next day with a stateless Playground. The experimental generate_prompt, improve_prompt and templatize_prompt endpoints started returning errors the same day, and users had until Sep 1, 2026 to export saved Workbench prompts and evals.
Sponsored Agents alpha
Advertising format allowing users to chat with an AI representative of a business; initially limited to selected advertisers.
Usage view that changes with plan and credit state
Claude Code's /usage shows API-style session cost for everyone, but subscribers also get plan usage bars and a breakdown of what ate their limits (skills, subagents, MCP servers, loops). A usage-credits row appears only while credits are on, and the note "at your organization's configured rates" appears when an admin sets contract pricing.
Anti-distillation decoy tools
Leak readers found code that injects fake tool definitions to poison traffic for anyone trying to distill Claude Code's behavior, plus signed connector text that hides raw reasoning and a client attestation header (cch=) that signs request bodies.
1M context window gates by plan in Claude Code
In Claude Code, the 1M-token variant of Opus 4.6 is included on Max, Team and Enterprise but needs usage credits on Pro, and Sonnet 4.6 [1m] needs usage credits on every subscription plan. Newer models (Sonnet 5 and later, Opus 4.7 and later, Fable) run 1M natively on every plan, including Pro.
Knowledge cutoff and date confusion
Claude models sometimes misstate their own knowledge cutoff or assume an older current year, for example Sonnet 4 in Claude Code insisting its cutoff was April 2024. Anthropic's published system prompts now spell out the reliable cutoff and inject the current date to counter this.
Confetti Cannon
A settings toggle can trigger celebratory confetti behavior.
Rate-limit options menu and wait-and-continue at the limit
When a subscriber hits a session or weekly limit, Claude Code offers /rate-limit-options, including waiting for the reset and then continuing the interrupted task automatically (v2.1.234+). Admins can control the auto-wait with the autoContinueAtUsageLimit managed setting, and /usage-credits sends a request to admins for members without billing access.
Cooking-themed completion verbs ("Cooked for 2m")
After a response, Claude Code reports turn duration with a cooking verb (e.g., "Cooked for…"). The leak showed 8 such verbs (e.g., Baked, Sautéed) versus 187 in-progress verbs.
! shell mode prefix
Starting a prompt with ! runs a shell command directly, skipping Claude, and adds its output to the session so Claude can respond to it. It supports Ctrl+B backgrounding, Tab autocomplete from previous ! commands in the project, and pasting text that starts with ! switches into shell mode automatically.
/team-onboarding guide generator
/team-onboarding analyzes your last 30 days of Claude Code sessions, commands and MCP server usage and writes a markdown ramp-up guide a new teammate can paste in as their first message. On Pro, Max, Team and Enterprise plans it also returns a share link that teammates can open directly in Claude Code.
Stealth model gray tests (Opus 5.2, Fable 5.2)
From about Sep 14 to 17, 2026, some requests to 'Opus 5' in Claude Code reportedly hit an unannounced Opus 5.2 route, noticed through faster, cleaner output. Days later testers posted clips of an unreleased 'Fable 5.2' coding animated mascots in pure JavaScript, and TestingCatalog said Anthropic appeared to be quietly routing some web prompts to it.
Slash commands that appear only for some plans and platforms
Claude Code's / menu changes with plan, platform and environment. /desktop shows only on macOS and x64 Windows with a Claude subscription, /upgrade is hidden on Enterprise, /privacy-settings is Pro and Max only, and /setup-bedrock and /setup-vertex stay hidden until the matching env var is set. A few commands, such as /heapdump, never appear until you type the full name.
Prompt cache lifetime drops to 5 minutes on usage credits
On a subscription, Claude Code's prompt cache lives for an hour, but once you draw on usage credits it drops to five minutes, so a short break costs a full cache rebuild. Users can choose the TTL themselves to keep the hour.
Claude Code seats for Team and Enterprise
On Aug 20, 2025 Team and Enterprise gained premium seats that bundle Claude and Claude Code, with per-user spend caps and Claude Code analytics. Enterprise also got the Compliance API at the same time. Seat tier now also decides whether Fable is included.
Output styles: deprecated, then revived
Claude Code output styles (built-ins like Explanatory and Learning) were deprecated in 2.0.30 in Oct 2025, then /output-style was deprecated again in favor of /config in 2.1.73. Version 2.1.269 in Sep 2026 added /output-style [name] back for listing and switching styles, including over Remote Control and in cloud sessions.
Conway always-on agent
TestingCatalog found Conway, a standalone always-on Claude environment with Search, Chat and System panels, webhook triggers, Chrome control, notifications and installable extensions packaged as .cnw.zip files. Later builds showed settings parity across web and iOS and shareable custom UI tabs.
Operon and BugCrawl
TestingCatalog's May 31, 2026 roundup described Operon, a Claude Desktop research mode for life sciences with private environments, Plan/Auto modes and local file access for work like CRISPR and RNA analysis, and BugCrawl, a Claude Code bug finder with GitHub, Jira and Linear integration that mirrors Claude Security for general bugs.
Android version-number haiku
Repeatedly tapping the Android app version number revealed a hidden poetry/haiku screen.
/color prompt bar color
/color tints the prompt bar for the current session in red, blue, green, yellow, purple, orange, pink or cyan, so side-by-side sessions are easy to tell apart. Running it with no argument picks a random color, and when Remote Control is connected the color syncs to claude.ai/code.
Demo mode for streamers (IS_DEMO, CLAUDE_CODE_HIDE_CWD)
Setting IS_DEMO hides your email and organization name from the Claude Code header and /status and skips onboarding, for streaming or recording. CLAUDE_CODE_HIDE_CWD hides the working directory in the startup logo, where a path could expose your OS username. Quirk: IS_DEMO=0 or false still turns demo mode on; you must unset it.
spinnerTipsOverride (custom spinner tips)
spinnerTipsOverride lets you add your own tips to the rotation Claude Code shows while it works, or hide the built-in tips entirely with excludeDefault. Tips can carry ids, cooldowns and priorities, load from a tipsFile, and use a custom prefix label of up to 40 characters, so teams can push house tips through managed settings.
Claude Code on the web launch credits ($250 / $1,000)
To push parallel cloud sessions, Anthropic gave Pro users $250 and Max users $1,000 in API-priced credits usable only in Claude Code on the web and mobile, separate from normal limits.
Opus 5.5 fast mode and Enterprise inference hooks
Claude Opus 5.5 (Sep 22, 2026) ships with a fast mode labelled research preview. Separately, 'inference hooks' entered beta for Claude Enterprise on Aug 5, 2026, letting an organization route requests to its own AI security server for pre-inference approval.
Partner early-access windows before model launches
Before Opus 5.5 launched on Sep 22, 2026, leakers claimed partners received roughly 24 hours of early access to a model under the codename 'claude-wafer-eap' ('eap' read as early access program). The same leak claimed thinking is always adaptive and forced tool use is retired, which the official Opus 5.5 release notes later matched.
Viral fake 'cheat codes' (L99, /godmode, /ghost)
Social posts claimed secret Claude modes unlocked by 'L99', '/godmode' or '/ghost'. A hands-on test found Claude Code rejects the slash versions as unknown skills and L99 does nothing, and none appear in the leaked source.
Claude 2.0, 2.1 and Claude 3 Sonnet retirement
Claude 2.0, Claude 2.1 and the original Claude 3 Sonnet (claude-3-sonnet-20240229) stopped working on the API on Jul 21, 2025, six months after Anthropic announced their deprecation on Jan 21, 2025.
Claude Instant and Claude 1.x retirement
Anthropic retired Claude Instant 1.0 to 1.2 and Claude 1.0 to 1.3 on the API after announcing deprecation on Sep 4, 2024. Claude Instant was the original fast, cheap tier that Haiku later replaced.
# quick-memory shortcut
Typing # followed by text at the start of a Claude Code prompt used to offer to save that text into a user or project CLAUDE.md file. Anthropic removed it in v2.0.70, and users kept hitting it because courses and docs still taught it.
Commit attribution control (attribution setting)
By default Claude Code adds a Co-Authored-By Claude line to commits and a generated-with note to pull requests. The attribution setting, which replaced the deprecated includeCoAuthoredBy, lets you change or remove those lines.
Reduce motion (prefersReducedMotion)
prefersReducedMotion reduces or turns off the spinner, shimmer and flash animations in Claude Code, and appears in /config as Reduce motion. It sits with other quiet accessibility options such as colorblind-friendly daltonized themes, syntaxHighlightingDisabled for screen readers, and CLAUDE_CODE_ACCESSIBILITY to keep the native cursor.
Ultrareview: 3 free runs, then usage credits
Claude Code's /ultrareview (now /code-review ultra) runs a multi-agent code review in a cloud sandbox. Pro and Max get three free runs, after which each run needs usage credits.
Session handoff between terminal, desktop and cloud
/desktop (alias /app) moves the current terminal session into the Claude Code Desktop app, and /teleport (alias /tp) pulls a cloud session from Claude Code on the web into your terminal, branch and conversation included. Both require a Claude subscription, and /desktop works only on macOS and x64 Windows.
Orbit and Hub Mode
Orbit appeared in May 2026 as a settings toggle for an opt-in, timezone-aware proactive briefing assistant in Cowork, pulling from Gmail, Slack, GitHub, Calendar, Drive and Figma. It disappeared, and in Aug 2026 an internal-only 'Hub Mode' with the same planet icon showed up in the mobile app; Claude Code on the web has a related Hub page that generates a task board for you and your agents.
/heapdump (deliberately hidden command)
Claude Code keeps /heapdump out of the slash menu on purpose: typing a partial name never reveals it, and it appears only once you type the full name. It writes a JavaScript heap snapshot plus a memory-breakdown diagnostics file to ~/Desktop for debugging high memory use. The docs warn that the .heapsnapshot holds your full conversation and credentials, so only the -diagnostics.json file should be shared.
Claude in Chrome saved shortcuts (/ in side panel)
In the Claude in Chrome side panel you can save a prompt that worked as a shortcut, call it back by typing /, and schedule it to run daily, weekly, monthly or yearly from the clock icon. The classic side panel also offers recorded workflows, which the newer Cowork-backed panel lacks.
Enterprise role-based model and effort restrictions
Only Enterprise organizations can restrict which models each role may use and cap effort levels. Members then see a filtered model picker, effort levels above the cap disappear from the effort menu, and open chats on a removed model fall back to the member's default.
Text Completions API (legacy /v1/complete)
Anthropic's original API used a raw "Human:/Assistant:" prompt string at /v1/complete. Docs now label it "[Legacy]", say future models and features will not be compatible, and point developers to the Messages API migration guide.
Claude Desktop Buddy (open hardware desk pet)
Anthropic open-sourced a reference project that turns a roughly $30 ESP32 board into a Bluetooth desk pet for the Claude desktop apps. It animates through states like Sleep, Busy, Celebrate and Dizzy, and its buttons approve or deny agent actions.
Enterprise activation promo ($1,000 per new Code/Cowork user)
Usage-based Claude Enterprise organizations earned $1,000 in Claude Code and Cowork credits for every user who activated either product for the first time, capped at 10,000 seats ($10M) per organization.
Memory shared across chat and Cowork
On Aug 25, 2026 memory began spanning chat and Cowork, on by default for Free, Pro and Max and off by default for Team and Enterprise. What Claude learns in one surface now carries into the other.
Undocumented API beta flags in the SDK enum (e.g. dev-full-thinking)
The API reference lists every accepted anthropic-beta value, and the list includes names with no public docs page, such as dev-full-thinking-2025-05-14, alongside documented ones like context-1m-2025-08-07 and dreaming-2026-04-21. The list doubles as a timeline of features before launch. A beta your organization lacks access to returns a 400 error.
Assistant prefill unsupported on newest models
Prefilling, where developers wrote the start of Claude's reply to steer format, no longer works on Claude 4.6 and later models or Claude Mythos Preview; such requests return a 400 error.
/agents wizard removed
Claude Code 2.1.198 removed the interactive /agents wizard for creating subagents; users now ask Claude to create subagents or edit .claude/agents/ directly. A later release (2.1.281) dropped the leftover "(removed)" menu entry, though typing /agents still explains where it went.
/vim command removed
Claude Code once toggled Vim bindings with /vim. Version 2.1.92 removed the command (along with /tag); Vim mode now lives under /config, Editor mode.
Resume-from-summary offer and suggestions that pause near limits
On Pro and Max, when you resume a large Claude Code session after a long break, Claude Code offers to resume from a summary so you avoid reprocessing the whole history. Prompt suggestions quietly stop while your account is close to or at its usage limit.
Claude Money
Strings and onboarding screens in the Claude iOS app point to a 'Money' tab that would let users link bank accounts and talk through spending and financial plans ('Understand your money with Claude'). Data provider and supported regions are unknown.
'You should know' built-in mod
Claude Code v2.1.287 shipped a built-in mod called 'You should know', a side agent that watches the session and flags things you or Claude might miss. It is off by default and has no toggle in the menus; you enable it with a plugin command, and only in first-party sessions with telemetry on.
Frustration keyword detection in Claude Code
A userPromptKeywords file reportedly scans prompts for phrases like 'wtf' and 'this sucks', and the client counts Escape presses and deleted drafts as hesitation signals. Engineers also left comments such as 'Mulberry32, tiny seeded PRNG, good enough for picking ducks'.
Lifecycle
A capability can move through several states.
The model is intentionally non-linear. Some signals die before documentation. Others skip experimentation, become normal immediately, or return later in a new form.
- 01Signal
Leaked, hinted, observed, or discovered
- 02Experiment
Limited cohort, preview, or temporary behavior
- 03Documentation
Officially acknowledged or clearly described
- 04Expansion
More plans, geographies, surfaces, or users
- 05Normalization
Becomes ordinary product behavior
- 06Retirement
Removed, ended, superseded, or reworked
Emerging patterns
Working hypotheses, not conclusions.
These interpretations should change as the corpus expands. Forefield keeps the analytical layer visible without confusing it with the underlying evidence.
Power-user layers are becoming a second interface.
Commands, shortcuts, settings, and hidden affordances create a parallel product for people who know where to look.
Documentation often follows reality.
A capability can exist, spread, and influence behavior before it is clearly represented in the official product story.
Experiments have lineages.
Temporary ideas can disappear, return under another name, move across surfaces, or become the ancestor of a more durable feature.
Distribution is part of product design.
Plan, geography, device, cohort, and partner gates shape what different users believe the product can do.
What becomes measurable
From curiosities to product behavior.
The standardized schema turns individual observations into comparable measures that can eventually reveal company-level product fingerprints.
Time from first observation to official acknowledgment.
How much prior knowledge a user needs before finding a real capability.
How long experimental behavior survives before expansion, replacement, or retirement.
How often users, developers, or researchers surface capabilities before the company does.
How often ideas and capabilities move between tools, plans, devices, and product surfaces.
How product primitives spread or converge across AI companies over time.
Research corpus
See the evidence and the thinking.
The project keeps source data, AI-assisted analysis, methodology, and interpretation visible as separate layers. These working materials do not all carry the same evidentiary weight.
OpenAI Hidden Features & Product Signals
11 structured records in the current tracker. Reviewed weekly.
Open source ↗TrackerAnthropic Hidden Features Tracker
128 structured records across Claude, Claude Code, research, brand, and related surfaces. Reviewed weekly.
Open source ↗AI working recordChatGPT research and framing conversation
The working discussion that moved the project from Easter eggs toward a longitudinal product-behavior study.
Open source ↗AI working recordClaude UI/UX comparison conversation
Anthropic-side analysis of Claude and ChatGPT interface behavior and transparency.
Open source ↗AI working recordClaude research share: UI/UX and model transparency
A parallel research perspective used as comparative evidence and interpretation.
Open source ↗MethodHidden Labs standardization prompt
The normalization prompt used to make company trackers more comparable and auditable.
Open source ↗Method
Observation first. Interpretation second.
Every tracker record is designed to separate what was observed from what we infer. Verification level, evidence type, confidence, discovery origin, exposure intent, lifecycle dates, secrecy, strategic significance, awareness, and momentum are tracked independently when the source tracker supports them.
The goal is not to maximize the number of interesting entries. It is to build a longitudinal record that can survive scrutiny, preserve uncertainty, and make product behavior comparable across companies.
A viral rumor can be interesting without being true. A documented feature can be real without being discoverable. A temporary experiment can matter even if it disappears.
Why this lives in Forefield
Research does not need one shape.
This project is also a demonstration of Forefield itself. The underlying work includes structured datasets, AI conversations, source evidence, comparative analysis, changing hypotheses, and a public narrative. Another Forefield project might be an archive, a field study, a map, a historical reconstruction, or a series of experiments.
Forefield provides enough structure to preserve evidence and provenance, while leaving the public expression free to match the research.
Explore Forefield Research