<- all tokdocs
2026.10.07@ai.honeycove

GPT-6 Astra does not run inside Claude Code; Claude Code's own model docs list Anthropic models only, and the setup in this video shells out to a separate Codex CLI process

AI
Watch on TikTok

View on TikTok ->

A 37.5-second vertical clip, 1080x1920 SDR at 30 fps, MP4 in a QuickTime/MOV container (format_name mov,mp4,m4a,3gp,3g2,mj2), H.265 video in the hvc1 tag at Main profile level 150 and 280,378 bps over 1,125 frames, HE-AACv2 audio at 44,100 Hz stereo and 64,280 bps over 811 frames, 351,630 bps overall, 1,648,266 bytes on disk, format id bytevc1_1080p_351630-1; ffprobe confirms every one of those figures against metadata.json. Posted at epoch 1791337187, which is 2026-10-07 01:39:47 UTC, by handle @ai.honeycove under the nickname "AI Honeycove" (uploader_id 7615522612223722526), over the attached sound "Million Dolla Hip Hop" credited to Brentin Davis; the description runs 230 characters and ends in the tags #AI #AITools #ChatGPT #TechNews. At capture time, the mtime of metadata.json on 2026-10-19 06:03:49 local, the post showed 19,000 views, 514 likes, 12 comments, 159 reposts and 622 saves. I read all 19 extracted frames (001 through 019, spaced two seconds apart from roughly 0 s to 36 s) and the full 141-word, 9-segment transcript. Frame 001 is a white card reading "you". Frame 002 shows an OpenAI studio set with "INTRODUCING GPT-6" projected on a wall and the caption "GPT6". Frame 003 is the Claude starburst wordmark. Frame 004 is the video's only quantitative claim, a dark chart headed "STOP BURNING YOUR LIMITS" with a full red bar labelled "CLAUDE ONLY ... 100%", an error strip reading "5-hour limit reached · resets 4:00 PM", and a partial orange bar labelled "FABLE + ASTRA ... 38%". Frame 005 stacks two cards, "FABLE 5.1" badged "BOSS" beside the Anthropic starburst and "GPT-6 ASTRA" badged "WORKER" beside the OpenAI knot, above a bearded presenter in a black shirt at a glass desk with a green banker's lamp. Frames 006 through 008 are a real Claude Code session: "Skill(astra-advisor:astra-advisor) ... Successfully loaded skill", "Codex CLI is available. Now I'll write the brief and consult Astra.", "Consulting GPT-6 Astra on landing page structure/design/tech approach (trigger: user explicitly requested Astra second opinion, interaction 1/3).", "Running Astra consult via Codex CLI · 7s", and then "Running Codex review with explicit cd flag · 57s" above the literal command codex exec --sandbox read-only --ephemeral --color never --cd "C:\Users\Administrator\astra-demo" --output-last-message "...\scratchpad\codex_review.txt" "Review the file index.html in t…" with a spinner line reading "Elucidating… (4m 21s · ↓ 8.3k tokens)". Frame 011 shows npm install -g @openai/codex in an Administrator PowerShell window. Frame 012 shows /plugin marketplace add openai/codex-plugin-cc, "Successfully added marketplace: openai-codex", /reload-plugins returning "Reloaded: 8 plugins · 83 skills · 6 agents · 3 hooks · 35 plugin MCP servers · 0 plugin LSP servers", and /plugin install codex@openai-codex. Frame 013 is a File Explorer window at Administrator > astra-demo holding one file, CLAUDE.md, dated 22/09/2026 11:48. Frames 014 through 016 show that file open, headed "# AI Collaboration Rules", with a seven-step list beginning "1. Ask Astra for an independent opinion." and ending "7. Run the tests before finishing.", followed by "Do not use Astra or Codex for every small task." and "The main Claude Code session remains responsible for the final decision." Frames 017 and 018 show an unrelated research artifact, a linear-attention literature report with an "OUTPUT FILES SUMMARY" table listing a 354-page PDF and 59 downloaded files. The spoken argument is that you can run GPT-6 Astra inside Claude Code to stop burning token limits by making Fable the boss and Astra the worker, and that setup takes a ChatGPT account, a Claude account, Codex, Claude Code, a plugin, and some CLAUDE.md text. No frame I read shows a token counter for a Claude-only run set against a Fable-plus-Astra run, so the 38 percent on frame 004 is a graphic with no measurement behind it anywhere in the video. The transcript's last segment ends at 37.16 s against a real audio duration of 37.498 s, so Whisper did not hallucinate past the end of the audio here and every quoted line below sits inside the recorded audio.

Verified: GPT-6 Astra exists under that exact name, announced 2026-09-03, and was 34 days old when this posted

At 0.00 s the video says "Most people don't know, you can run OpenAI's new GPT-6 Astra inside of Claude Code and stop it burning through your token limits." The model name is correct and the model is real. OpenAI's announcement thread in the Announcements category of its developer community, titled "Introducing GPT-6-Astra: The most intelligent and aligned model in the world," is dated September 3, 2026 and gives the API model id as gpt-6-astra, a 1,050,000-token context window, 128,000 tokens of maximum output, an April 30, 2026 knowledge cutoff, and pricing of $10.00 per million input tokens, $1.00 per million cached input tokens, $12.50 per million cache writes, and $50.00 per million output tokens. It claims state-of-the-art results on Agents' Last Exam, AutomationBench, ScreenSpot Pro, FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. I could not reach OpenAI's own index page for the launch at openai.com/index/introducing-gpt-6-astra/, which returned HTTP 403, so I am not citing it.

The word "new" is doing some work. The video posted on 2026-10-07, 34 days after the September 3 announcement. That is not a fabricated model, but it is not a scoop either, and the gap matters for the next section.

Claude Fable 5.1, the model on the "BOSS" card in frame 005, is also real. Anthropic's announcement page "Introducing Claude Fable 5.1 and Claude Mythos 5.1" is dated September 2026 and states that "Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads," that for demanding use cases "the savings will often be much larger—up to approximately 45%," that "Cache reads now cost 75% less, or $0.25 per million tokens," and that standard pricing stays at "$10 per million input tokens and $50 per million output tokens." The page gives the month but not a day, so I am citing September 2026 rather than a specific date.

Correction: Claude Code cannot be configured to use GPT-6 Astra, per Anthropic's own model-configuration page

The headline claim is that Astra runs "inside of Claude Code." Anthropic's model configuration documentation at code.claude.com says Claude Code supports Anthropic models only. The aliases it lists are default, best, fable, sonnet, opus, haiku, sonnet[1m], opus[1m], and opusplan. The full model ids it lists are claude-fable-5-1, claude-fable-5, claude-opus-5-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-5-5, claude-sonnet-5, claude-sonnet-4-6, claude-sonnet-4-5, claude-haiku-5-5, and claude-haiku-4-5. The providers it lists are the Anthropic API, Amazon Bedrock, Microsoft Foundry, Google Cloud's Agent Platform, the Claude Platform on AWS, the Claude apps gateway, and LLM gateways reached through ANTHROPIC_BASE_URL. There is no entry for an OpenAI model and no mechanism for adding one.

The plugin does not change that. Anthropic's plugins overview says a plugin is "a directory of skills, agents, hooks, MCP servers, or other components that Claude Code installs and loads as one unit," and enumerates the component types as skills, agents, hooks, a hooks module, and MCP servers. A model is not one of them. Nothing a plugin installs can swap the model answering in the session.

What the video's own frames show is the accurate version. Frame 008 captures the literal command: codex exec --sandbox read-only --ephemeral --color never --cd "C:\Users\Administrator\astra-demo" --output-last-message "...\codex_review.txt". That is Claude Code running a separate binary as a shell command and reading a text file back. OpenAI's README for codex-plugin-cc describes the plugin as "Use Codex from inside Claude Code for code reviews or to delegate tasks to Codex." Claude Code is the caller. Codex is a separate process on a separate account answering on a separate model. "Inside of Claude Code" describes a subprocess, not a model setting.

Unverified: the 38 percent on screen traces to no benchmark, and the skill in the video explicitly disclaims the savings claim

Frame 004 at roughly 6 s pairs "CLAUDE ONLY ... 100%" and a "5-hour limit reached · resets 4:00 PM" error strip against "FABLE + ASTRA ... 38%". At 10.62 s the voiceover says "They handwork back and forth and cover each other's gaps so you get both models while using less of each one." The title promises to "slash token usage by ..." and then truncates before the number.

I could not source the 38 percent to anything. OpenAI's README for codex-plugin-cc contains no statement about reducing token usage, saving tokens, or any efficiency percentage. The skill the video actually loads, astra-advisor by PatrickSys, goes further and rules the claim out in its own README: "This release does not establish that Astra Advisor improves coding quality, catches more bugs, saves money, or is independent in a statistical sense." The skill is marked experimental and verified at v0.1.1. Neither Anthropic's Fable 5.1 announcement nor OpenAI's Astra announcement publishes a cross-vendor orchestration benchmark of this shape.

The one real token figure anywhere in the video undercuts the graphic. Frame 008 reads "Elucidating… (4m 21s · ↓ 8.3k tokens)" during a single Codex review that the same frame times at 57 s and then 4m 21s. That is consumption during one delegated step, not a comparison, and no frame I read shows a paired Claude-only control run. A bar chart with no measured run behind it is an assertion in chart form.

There is also a cost direction the graphic hides. Astra's announced API pricing is $10.00 per million input tokens and $50.00 per million output tokens. Anthropic's announced Fable 5.1 pricing is $10 per million input tokens and $50 per million output tokens. Per token, the two models cost the same. Delegating work to Astra does not make a token cheaper. It makes it bill to a different vendor.

Clarification: the real mechanism is quota arbitrage against a 5-hour session window, not reduced token usage

The error strip in frame 004 is the honest part of the graphic. Anthropic's support article on using Claude Code with a Pro or Max plan states that "Both Pro and Max plans have a five-hour session limit and a weekly limit," that those weekly limits apply across Claude and Claude Code combined, and that Max plans carry "a separate weekly limit for Fable."

Running a Codex subprocess does reduce how fast you burn that specific window, because OpenAI's README requires a "ChatGPT subscription (incl. Free) or OpenAI API key" and Node.js 18.18 or later, and the Codex work bills against the ChatGPT account. That is two quotas instead of one. It is not "using less of each one" in the sense the voiceover implies. Total tokens across both vendors go up, because the orchestration adds a written brief, a scratchpad file, a review request, and a review result on top of the original task. Frame 007 shows exactly that overhead: "Made 1 scratchpad edit +42, ran 2 shell commands" before the consult even starts.

Anthropic's plugins page names a second cost the video does not mention. For every skill, agent, and command Claude can invoke on its own, "the name and description are in Claude's context on every turn so that Claude knows it exists. Those tokens count toward your usage and leave less room in the context window even in sessions where nothing from the plugin runs." Frame 012 shows this user's reload output: "8 plugins · 83 skills · 6 agents · 3 hooks · 35 plugin MCP servers". That is a standing per-turn context charge on the Claude side of the ledger, paid in every session, in exchange for a chart claiming 62 percent savings.

Correction: "most people don't know" describes a plugin OpenAI published 191 days before this video

The hook at 0.00 s frames the setup as obscure. The GitHub releases page for openai/codex-plugin-cc lists v1.0.0 on 30 March, then v1.0.1 and v1.0.2 on 31 March, v1.0.3 on 8 April, v1.0.4 on 18 April, v1.0.5 on 23 June, and v1.0.6 on 8 July. Against the video's 2026-10-07 post date, v1.0.0 is 191 days old and the most recent release, v1.0.6, is 91 days old. The plugin is an official OpenAI repository, not a workaround someone discovered.

This is the same pattern @ai.honeycove used in another video in this corpus, where a Cloudflare tool 116 days past release was presented as "just open sourced." The novelty framing is a recurring device on this account, applied to shipped and documented software.

Partial correction: the codex exec line names no model, the CLAUDE.md is plain prompt text, and neither pins the setup the video describes

At 7.52 s the video says "To do this, you make Fable the boss and Astra the worker," and frame 007 shows Claude Code narrating "Consulting GPT-6 Astra on landing page structure/design/tech approach." Whether Astra actually answered depends on configuration the video never shows.

OpenAI's README for codex-plugin-cc states that "if you do not pass --model or --effort, Codex chooses its own defaults." The codex exec invocation visible in frame 008 passes --sandbox read-only, --ephemeral, --color never, --cd, and --output-last-message, and no --model. The release notes for Codex CLI rust-v0.159.1, published 29 September, say the release "Added GPT-6.1 Sol as the default model in the bundled catalog and Amazon Bedrock Mantle and Runtime catalogs," and make no mention of GPT-6-Astra. That is 8 days before this video posted. A default codex exec after that release reaches Sol, not Astra.

The astra-advisor skill the video loads does pin the model, but only on the Codex side: its README says Codex "prepares the packet locally, spawns one native subagent with model=gpt-6-astra, reasoning_effort=low, and fork_turns=none." It also requires "Node.js 22+ and a signed-in Codex installation with access to gpt-6-astra", a stricter requirement than the Node.js 18.18 the plugin asks for, and one the video never states. Frame 007's "interaction 1/3" matches the skill's bounded-consultation design. So the Astra consult in frame 007 is plausibly real, and the separate Codex review in frame 008 is a different command with no model pinned.

At 24.74 s the video says "Last, paste a few sections into your Claude MD that tell them how to hand off work to each other and check each other's work." Frames 014 through 016 show the file in full. It is seven numbered instructions in plain English under "# AI Collaboration Rules", ending with "Do not use Astra or Codex for every small task." and "The main Claude Code session remains responsible for the final decision." There is no configuration syntax in it, no model field, and no routing logic. It is instruction text the Claude model reads and may or may not follow.

Frame 013 dates that CLAUDE.md to 22/09/2026 11:48, 15 days before the post. The demo directory, C:\Users\Administrator\astra-demo, contains that single file. The video presents a prepared demo folder, which is fine for a tutorial, and does not show a from-scratch run.

One more thing the frames do not support: frames 017 and 018 show a linear-attention research report with an "OUTPUT FILES SUMMARY" table listing "Linear_Attention_Report.pdf ... 3.11 MB ... 354-page comprehensive report" and "Pdf files ... 84.32 MB ... 59 files downloaded", plus "240% YoY growth: From 5 papers (2024) to 17 papers (2025)". That artifact has no connection to the landing-page task shown in frames 006 through 008, and no frame ties it to the Fable-plus-Astra setup. It reads as B-roll.

Key Takeaways

  • Verified: GPT-6 Astra exists under that exact name. OpenAI's announcement is dated September 3, 2026, gives the API id gpt-6-astra, a 1,050,000-token context window, and pricing of $10.00 per million input and $50.00 per million output tokens.
  • Verified: Claude Fable 5.1 exists. Anthropic's announcement page, dated September 2026, gives its pricing as $10 per million input and $50 per million output tokens, with cache reads at $0.25 per million.
  • Correction: The video says you can "run OpenAI's new GPT-6 Astra inside of Claude Code." Anthropic's model-configuration docs list Anthropic models only, across the Anthropic API, Bedrock, Microsoft Foundry, Google Cloud's Agent Platform, the Claude Platform on AWS, the Claude apps gateway, and LLM gateways. Claude Code shells out to a separate codex exec process, as frame 008 shows.
  • Correction: A plugin cannot change the session model. Anthropic's plugins overview lists the component types as skills, agents, hooks, a hooks module, and MCP servers.
  • Unverified: The "38%" against "100%" in frame 004 traces to no benchmark. OpenAI's codex-plugin-cc README makes no token-savings claim, and the astra-advisor README states the release "does not establish that Astra Advisor improves coding quality, catches more bugs, saves money, or is independent in a statistical sense."
  • Clarification: Per token, Astra and Fable 5.1 cost the same at $10 input and $50 output per million. Delegating does not make tokens cheaper. It spends a second vendor's quota against a 5-hour Claude session window that Anthropic's support article confirms exists.
  • Context: Anthropic's plugins page notes that every auto-invocable skill, agent, and command costs context tokens on every turn. Frame 012 shows this user running 8 plugins, 83 skills, 6 agents, 3 hooks, and 35 plugin MCP servers, a standing charge the savings chart ignores.
  • Correction: "Most people don't know" describes a plugin OpenAI released as v1.0.0 on 30 March 2026, 191 days before this video, with its most recent release v1.0.6 on 8 July 2026, 91 days before.
  • Partial correction: The codex exec command in frame 008 passes no --model. OpenAI's README says Codex then picks its own default, and Codex CLI rust-v0.159.1 of 29 September 2026 made GPT-6.1 Sol the bundled default, 8 days before this video. The Astra consult in frame 007 is pinned separately by the astra-advisor skill, which spawns a subagent with model=gpt-6-astra.
  • Clarification: The astra-advisor skill is third-party and experimental at v0.1.1, not part of OpenAI's plugin, and it requires Node.js 22+ plus a Codex login with Astra access. The video never names it aloud or states those requirements.
  • Context: I could not reach OpenAI's own launch page at openai.com/index/introducing-gpt-6-astra/, which returned HTTP 403, and the New Stack article returned only a signup form, so neither is cited above.

Resources

Published October 7, 2026. Writeup generated from a favorited TikTok.