<- all tokdocs
2026.09.03@sabrina_ramonov

Codex Fast Mode Costs 2.5x Usage For 1.5x Speed, And The Email Trigger In This Clip Does Not Run In The App On Screen

AI
Watch on TikTok

View on TikTok ->

This is a 61-second vertical clip at 720x1280, posted 2026-09-03 by @sabrina_ramonov, whose TikTok channel nickname is "Sabrina Ramonov 🍄", scored to the original audio track "original sound - Sabrina Ramonov 🍄"; at capture on 2026-10-05 it held 166,500 views, 7,230 likes, 302 comments, 692 reposts, and 3,381 saves. I read 20 of the 31 extracted frames and all 213 transcribed words. The clip is a talking-head take composited over a slow scroll through the creator's own blog post, with a white title card reading "Everyone dropping Claude for ChatGPT 😳" pinned over the first ten seconds and word-by-word burned-in captions along the bottom ("BACK", "CODE", "HOWEVER,", "MAKES", "SUPERIOR", "WORKS", "INTEGRATION.", "UNIFIED", "HOW", "JUST", "AND", "THEN", "IN", "HONESTLY", "SEPARATELY.", "1.5", "MORE", "IT.", "FAST", "FROM", "HERE."). The scrolled post shows four numbered sections. Section one reads "1. Codex Has A Far Superior In-App Browser" with the line "Codex's built-in browser keeps pages, chat, files, and feedback in 1 workspace." and "Just hit "CMD + T" and it will open a new browser tab.", above a screenshot of a Codex browser tab labeled "Blotato | #1 Soc" with a dropdown listing "Review ^⇧G", "Terminal", and a globe entry, plus shortcut hints "⌘T", "⌘P", and "⌥⌘S"; the body text below reads "In my work, it is FAR su...", "Try it:", "1. Open "Plugins," en", "2. Check the page be", "SAMPLE PROMPT: res", "Sit back and watch", "Not only can Codex". Section three reads "3. Unified Scheduling Is My #1 Reason" with "Codex Scheduled tasks puts every job in 1 sidebar." above a screenshot of OpenAI's own docs page titled "Scheduled tasks / Schedule recurring tasks in ChatGPT", whose left nav lists Overview, Workflows (Projects and chats, Sites, Visualizations, Scheduled tasks, Long-running work, Notifications, Pets, Codex Micro) and Capabilities (Browser, Computer use, Voice, Plugins, Web search), captioned "OpenAI Scheduled tasks"; underneath sits a real task card reading "Paused", "Review YTJobs video editor ap", "Scope and stopping conditions", "- Work only in this job's Und", "Shortlisted counts.", "- Review applicants one", "- If Undecided is alrea", "- When Undecided re", "pause/disable this ra", with a "Details / Runs in / Chat" footer. Section four reads "4. Codex Has Superior Built-In Image Generation" with "Codex image generation makes and edits images in your chat." above a chat card reading "Worked for 32s" and "Generated a photorealistic orange tabby eating a hot dog." with the generated cat image rendered in full, followed by "Claude reads images but lac", "Codex uses ChatGPT's late", "So if you're creating ad... you can do it all natively", "Yes, Claude still has C... feels fragmented from", "SAMPLE PROMPT: Mak... navy and crea". Section six reads "6. Fast Mode Helps When Delay Costs Money" with "Fast mode raises speed by 1.5x on supported models." above a Codex settings panel showing "Model 5.6 Sol", "Effort Extra High", "Speed Fast", "Advanced", a footer chip reading "⚡ 5.6 Sol Extra High", and an open Speed popover listing "Standard / Default speed" and a checked "Fast / 1.5x speed, more usage", with a hover tooltip reading "Codex Speed menu: Fast mode, 1.5x speed, and more usage"; the body reads "Be careful because the "1", "oftentimes MORE than 1.5x", "However, I still use it", "Because TIME is the mos", "I'm currently on the C... AND it is far from the", "I suggest using it fo... case that still sent". Section eight reads "8. How To Migrate From Claude To Codex" with "OpenAI has an import tool for Claude Code." above a screenshot of OpenAI's "Import from another agent" docs page, then a fully legible instruction block: "Use it first: 1. Open "Settings > Import" in the ChatGPT desktop app. 2. Choose "Claude Code," then pick projects and recent chats. 3. Finish plugin or account sign-ins after import. 4. Open 1 old project and rerun a job you know well.", followed by "Claude skills often live in .claude/skills.", "Codex repo skills live in .agents/skills, with both words plural.", "Not .skills.", "Not .agent.", "Each skill needs a folder with a SKILL.md file.", "AGENTS.md is different.", "It stores the rules Codex re... work, such as your writing style, test steps, and safety rules.", and "Codex settings live in .codex/config.toml." The final frame is a Chrome omnibox with "learn.chatgpt.com/docs/import.md" selected and a suggestion row reading "Ask Google about this page - learn.chatgpt.com". Sections two, five, and seven of the underlying post never appear on screen.

The tenure claim holds, but the "one of the first" claim does not check out from public dates

The spoken opener is "I've used Claude Code for over a year. I was one of the first person to start making tutorials on Claude Code." The first half survives verification. Anthropic shipped Claude Code on 24 February 2025 alongside Claude 3.7 Sonnet, describing it in its own announcement as "available as a limited research preview" that "enables developers to delegate substantial engineering tasks to Claude directly from their terminal" (Anthropic, Claude 3.7 Sonnet and Claude Code). The GitHub REST API reports the anthropics/claude-code repository was created on 2025-02-22, two days before that post. Claude Code reached general availability on 22 May 2025 with the Claude 4 launch (Anthropic, Introducing Claude 4). Measured against a 2026-09-03 post date, that is roughly 18 months since the research preview and 15 months since general availability, so "over a year" is possible and consistent.

The second half I could not confirm. The earliest dated Claude Code tutorial I found on her own publication is "The ULTIMATE AI Coding Guide for Developers (Claude Code)", dated 5 July 2025 on sabrina.dev, about four months after the research preview opened. Earlier posts may exist behind different titles or on other platforms. I am reporting what I could date, not asserting she was late.

The hook reverses which product shipped first

The title card says "Everyone dropping Claude for ChatGPT 😳" and the voiceover opens with "why everyone's switching back to ChatGPT." The word "back" implies ChatGPT held the terminal-coding-agent ground first and is reclaiming it. That is not what the record shows for this product category. Claude Code's research preview opened 24 February 2025 (Anthropic). OpenAI released Codex CLI on 16 April 2025, and the GitHub REST API reports the openai/codex repository was created on 2025-04-13. Claude Code came first by about seven weeks.

The clip also offers no evidence for "everyone." It cites no survey, download count, or usage figure. The only live cross-product number I could pull directly from a primary API at capture time is GitHub stars, which measures repository attention rather than seat count: the GitHub REST API returned 149,462 stars for anthropics/claude-code and 127,900 for openai/codex on 2026-10-05. That is not a migration statistic, and neither is the hook.

The email trigger she praises is unavailable in the ChatGPT desktop app she shows on screen

The voiceover says ChatGPT "has a unified scheduler... In ChatGPT, everything's in one place, and they also just release trigger-based workflow. For example, when an email comes in, then run this automation." The frames back this with a screenshot of OpenAI's own scheduled tasks docs.

OpenAI's documentation contradicts the "everything's in one place" framing on the exact feature she names. OpenAI's scheduled tasks page lists event triggers for Gmail ("New incoming messages, optionally filtered by sender or subject"), Slack, and GitHub pull request activity, and then states that event-triggered tasks are not available in the ChatGPT desktop app, the Codex CLI, or the IDE extension. Those three surfaces get time-based scheduling only. Her own section-eight instructions are written for "the ChatGPT desktop app," which is one of the three surfaces where the Gmail trigger does not run. OpenAI's changelog dates the event-trigger release to 25 August 2026, nine days before this clip posted, so "just release" is accurate.

The companion claim that Claude lacks trigger-based automation is also wrong. Anthropic's routines documentation describes three trigger types already shipped: "Scheduled", "API" via an authenticated HTTP POST to a per-routine /fire endpoint, and "GitHub": "run automatically in response to repository events such as pull requests or releases." The endpoint ships behind the beta header experimental-cc-routine-2026-04-01, which dates the API trigger to roughly April 2026, months before OpenAI's August 2026 event triggers. On event-driven automation specifically, Claude shipped first and the clip reverses the order.

Where she is right is the fragmentation complaint, and she understates it. Anthropic's routines page links three separate scheduling surfaces in its own "Related resources" list: cloud routines, Desktop scheduled tasks that run locally on your machine, and /loop and in-session scheduling. That is a three-way split documented by the vendor, not a two-way one.

Fast mode's real price is 2.5x usage for 1.5x speed, which the voiceover never states

The audio says "I also love ChatGPT's fast mode. You can make it go 1.5x speed, even though it'll cost you more tokens, but this saves me time, so it's worth it." The on-screen panel confirms the interface: "Speed Fast", a popover entry reading "Fast / 1.5x speed, more usage", and the heading "Fast mode raises speed by 1.5x on supported models."

OpenAI's speed documentation gives the number the voiceover leaves out. Fast mode "uses included subscription limits at 2.5x the Standard rate", and "purchased credits and Enterprise pay-as-you-go usage are billed at 2x the Standard rate." You pay a 2.5x usage rate to get a 1.5x speedup. Her own on-screen text acknowledges this with "Be careful because the "1... oftentimes MORE than 1.5x", but a viewer who only listens hears "more tokens" and never learns the multiplier. The cost is also not measured in tokens. It is a rate multiplier applied to included subscription limits or to credit billing.

The heading "raises speed by 1.5x on supported models" also overgeneralizes. The same OpenAI page scopes the 1.5x figure to specific models: "GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna, where available. For GPT-5.6 and GPT-5.5, the speed increase is 1.5x." The 1.5x number is stated for GPT-5.6 and GPT-5.5, and the frame shows "Model 5.6 Sol" selected, so the demo matches the documented case. The doc does not publish a speed multiplier for the GPT-6 family, so "on supported models" is broader than the source supports. Toggling is /fast in the CLI, or service_tier = "fast" plus [features].fast_mode = true in config.toml.

The image generation claim is correct, the browser claim is app-only, and the migration paths check out

Three of her smaller claims verify cleanly, with one scope limit worth adding.

On image generation, she says Claude "honestly just doesn't have at all. You have to set it up separately." Anthropic's own documentation confirms this in plain language. The FAQ on Anthropic's vision page asks "Can Claude generate or edit images?" and answers "No, Claude is an image understanding model only. It can interpret and analyze images, but it cannot generate, produce, edit, manipulate, or create images." Codex's side verifies too: OpenAI's image generation page says built-in image generation uses gpt-image-2, invoked with $imagegen or by asking in chat. It also carries a cost the clip omits: "Image generations use included limits 3–5x faster on average than similar turns without image generation, depending on image quality and size." OpenAI's changelog dates in-conversation image refinement to 30 July 2026.

On the browser, the "CMD + T" shortcut is right. OpenAI's browser documentation gives the shortcut as "Cmd+T (Ctrl+T on Windows and Linux)" and describes the built-in browser as a way to "open websites, gather current information, and take action while you stay in control." The same page adds the limit she does not mention: "Browser isn't available in Codex CLI or the Codex IDE extension." Her claim is an app claim, not a Codex claim. Whether it is "far superior" to Claude's Chrome extension is a preference I cannot verify. For chronology, Anthropic's Chrome extension began as a 1,000-user Max pilot on 25 August 2025 and reached general availability on all paid plans on 26 August 2026, while OpenAI's changelog dates page comments and live previews in the built-in browser to April 2026.

On migration, every path she reads out is in the docs. OpenAI's import guide covers "Settings > Import" in the desktop app and /import in the Codex CLI, importing instruction files (AGENTS.md, settings.json, config.toml), skills, plugins, project folders, project memories, MCP server configuration, hooks, slash commands, subagents, and chat sessions from the last 30 days, with Codex CLI capped at 50 chats. The desktop app imports from Claude Code, Claude Cowork, or Cursor; the CLI imports from Claude Code or Cursor. The changelog dates this to 11 August 2026. Her skills-path detail is also correct and she is right to stress the plurals: OpenAI's skills documentation lists $CWD/.agents/skills, $CWD/../.agents/skills, $REPO_ROOT/.agents/skills, $HOME/.agents/skills, and /etc/codex/skills, with each skill being "a directory with a SKILL.md file plus optional scripts and references." She omits only the /etc/codex/skills admin path.

Key Takeaways

  • Correction: the clip's "switching back to ChatGPT" framing reverses the order. Claude Code's research preview opened 24 February 2025; OpenAI released Codex CLI on 16 April 2025. The GitHub REST API puts repo creation at 2025-02-22 for anthropics/claude-code and 2025-04-13 for openai/codex.
  • Correction: the clip presents trigger-based automation as a reason to leave Claude. Anthropic's routines docs already document schedule, API, and GitHub event triggers, with the /fire endpoint on the experimental-cc-routine-2026-04-01 beta header. OpenAI's event triggers shipped 25 August 2026.
  • Correction: "everything's in one place" fails on the exact feature demonstrated. OpenAI's own docs state event-triggered tasks are not available in the ChatGPT desktop app, Codex CLI, or IDE extension, and the desktop app is the surface her instructions target.
  • Partial correction: the on-screen line "Fast mode raises speed by 1.5x on supported models" is broader than the source. OpenAI publishes the 1.5x figure for GPT-5.6 and GPT-5.5 and lists GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna as supported without stating their multiplier.
  • Unstated cost: Fast mode draws on included subscription limits at 2.5x the Standard rate, and bills purchased credits and Enterprise pay-as-you-go at 2x. The voiceover says only "it'll cost you more tokens," which understates a 2.5x rate and misnames the unit.
  • Unstated cost: Codex image generations "use included limits 3–5x faster on average than similar turns without image generation," per OpenAI's own page. The clip praises the feature without naming the burn rate.
  • Verified: Claude cannot generate or edit images. Anthropic's vision FAQ states it directly, so "Claude honestly just doesn't have at all" is accurate.
  • Verified: "CMD + T" opens a browser tab, exactly as the on-screen text says. Scope limit: the same page says the browser is not available in Codex CLI or the IDE extension.
  • Verified: Codex reads repo skills from .agents/skills, with both words plural, and each skill needs a folder containing SKILL.md. The docs also list /etc/codex/skills, which the clip omits.
  • Verified: "Settings > Import" in the ChatGPT desktop app imports Claude Code setup, skills, plugins, MCP config, hooks, slash commands, subagents, and 30 days of chats. The URL shown in the final frame, learn.chatgpt.com/docs/import.md, resolves to that page.
  • Verified: her fragmentation complaint about Claude scheduling, and she understates it. Anthropic documents three separate surfaces: cloud routines, local Desktop scheduled tasks, and in-session /loop scheduling.
  • Could not verify: "I was one of the first person to start making tutorials on Claude Code." The earliest dated Claude Code tutorial I located on her own site is 5 July 2025, about four months after the research preview opened.

Resources

Published September 3, 2026. Writeup generated from a favorited TikTok.