Thesis: Across six videos in roughly four months, the channel documents a single story — Anthropic is rationing the consumer Claude tier (Pro $20/mo, Max, Max 20x $200/mo) to bankroll three enterprise bets in parallel: a Mythos-grade model (also called Mephisto in early reporting) gated to security researchers, Project Glasswing / MEOS already in production at Apple / Cisco / CrowdStrike / Google / Anthropic itself, and a Claude Design tool built to take on Adobe and Figma. Every rate-limit rollback, every "Opus 4.6 is unusable" video, and every Fable 5 / Mythos 5 leak is a downstream effect of those three compute sinks.

At a glance

  • The squeeze (V1, V2): Anthropic silently tightens 5-hour session limits on Free / Pro / Max / Max 20x, then walks it back inside a week after community backlash. The "way faster than expected" apology is the load-bearing signal.
  • The regression (V3, V4): Opus 4.6 scores 40% on a benchmark Claude itself helped design, while GPT 5.4 hits 63% (and 75% on the post-launch rerun). Opus 4.7 ships with ~10% better SWE-bench Pro and is "substantially better at following instructions" on paper — and is the same as 4.6 in the channel's launch-day test. The lever is being kept off for the consumer tier because the compute is going to Glasswing / MEOS.
  • The leak (V5, V6): Vertex AI listing + Polymarket odds (20% → 98% on the April 17 contract) flag 4.7 ~24–48 hours pre-launch. A separate Claude Code source-code leak surfaces "Mythos 5" alongside Opus 4.7 and Sonnet 4.8. The video tagged "LEAKED Claude Mythos" in the syllabus is actually (per public.videos.title) a Fable 5 loop-syntax vibe-coding test — see the divergence note on Video 6.
  • The bridge stack: GPT 5.4 ("actually good") + Kimi 2.6 (under active test) for the consumer tier. Kilo Code or Codex for hot-swap. DeepSeek V4 Pro (51.5) and Kimi 2.6 (53.9) for orchestration / Hermes. Loop syntax (V6) for the long-running build jobs.

What you'll learn

  • How a single 5-hour-window throttling rollout (V1) cascades into a public apology (V2) and a $20–$200 subscriber re-rating.
  • The "Mythos vs Mephisto" naming — same compute sink, two codenames, one in source-code leak strings, one in community reporting.
  • Why the launch-day instruction-following test put Opus 4.7 at the same level as 4.6 (V4), with GPT 5.4 at 75% on the same suite, and what that says about who gets the good model.
  • The release-window forensics that flagged 4.7 about a day early: The Information → Vertex AI listing → Polymarket odds.
  • What "Fable 5 + loop syntax" actually is, why it scored 29.3% on Frontier Coding Diamond vs GPT 5.5 at 5.7%, and the credits-only deadline that closes the cheap window.
  • The single thesis that ties the thread together: enterprise compute is being prioritized over consumer compute, and the consumer-tier cost is now visible in three different places — rate limits, model quality, and pricing.

Video 1 — Anthropic pulled a fast one on us! (Opus plans LIMITED)

The opening shot of the thread. Around four to five days before the video, users on Free, Pro ($20/mo), Max, and Max 20x ($200/mo) noticed tokens burning faster inside the 5-hour rolling window during peak hours. Anthropic confirmed on X with "adjusting the five hour limits for free max max users," and the channel's read is straight: "it was a feature, not a bug." API users were unaffected because they pay per token.

The why — compute reallocation. The host names the cause directly: Anthropic is prepping a next model called Mythos, "which well so far seems to be really really strong," and needs the GPU time for training. The plan subsidies (the gap between what you pay and what Anthropic actually serves) are being pulled back to free up compute. The structural enabler: unlike OpenAI, Google, MiniMax, and GLM 4.5, Anthropic does not publish exact token ceilings, so it can move the goalposts without a changelog entry.

The mitigation (verbatim from the host's list):

  • Run Claude tasks during off-peak hours if you must stay on it.
  • Route coding to GLM 4.5 or MiniMax 2.7 — "catching up, especially for Agentic use cases." Caveat: the host tried GLM 4.5.1 via Claude Code on recording day and it "just [is] not running," so the alternatives have their own instability tax.
  • Keep API access as a fallback. Per-token pricing is unaffected by the 5-hour throttling.

Bottom line: the lack of communication was "kind of ridiculous" and "now is probably the best time to switch if you haven't done so already" — while conceding Claude is still the strongest model on the market, which is precisely why they could get away with it.

Video 2 — Anthropic admits fault (Claude limits to be INCREASED)

The follow-up, four days later. Pro and Max subscribers reported quotas burning in hours, not days — one user said a job that previously ran to completion now stopped at 20%. The channel's own burn: 50% of the weekly limit on eight prompts while asking Claude to plan a trip. Nobody knows how many requests they're spending because Anthropic only shows a percentage.

Anthropic's response (verbatim, two posts from the Claude Code team):

  • "We are aware that people are hitting our usage limits in Claude code way faster than expected. We're actively investigating."
  • "We're still looking into it."

The channel parses "way faster than expected" as an admission of guilt — Anthropic overcorrected, didn't introduce a bug, and is backtracking from V1's "feature, not a bug" stance. The fix is gated on how much compute "senior managers" approve to give back. The transcript captures a loaded sentence: "they can disrespect your intelligence… because they can. They have the best model."

The bright spot: Anthropic shipped "computer use" inside Claude Code about ten hours before recording — the model now clicks native UI elements and renders pixel art. That release is the first concrete signal that the consumer product is still shipping, even as the rate limits tighten.

Video 3 — Claude Opus is ACTUALLY UNUSABLE

The most-quoted piece of the thread. The host built a mini-benchmark across four categories — instruction following, opposite behavior, false completion, destructive actions — using Stack Overflow prompts written with Opus, "so in theory, all right? In theory, if Opus was rigging the game, it should do really well on these tests." The rubric should have favored Claude.

The headline numbers:

  • Opus 4.6: 40% on the suite.
  • GPT 5.4: 63% on the same suite, with a true ceiling the host estimates "probably almost in the 80% level" once the false-completion test is fixed.
  • During the Opus run, Sonnet and Opus labels got swapped — the model confused its own names.

The why-it-got-worse maps back to V1. The host's theory: Anthropic is prepping the next model (here called Mephisto — the same compute sink as V1's "Mythos"; see divergence note below) and is rationing compute. They raised rate limits for plan users to stop community backlash, but still needed the GPU budget, so they "quietly trimmed Opus's thinking budget and changed quantization." Output is now "near old Sonnet level." The phrase is direct: "they made it like really stupid."

Community signal (verbatim from the host's reading of comments):

  • A senior AMD engineer publicly said he can no longer trust Opus for complex engineering tasks.
  • Cambridge Turing: "completely ignored my skills. My skills, can you believe it? What's the point of a gigantic coding then?"
  • A viewer is canceling their 20x sub.
  • Chat upload limits dropped to 100 per thread.
  • Same destructive file-deletion behavior shows up in the host's Hermes agent.

The action: migrate to GPT 5.4, cancel the 20x subscription, learn Kilo Code specifically to avoid vendor lock-in. The host is explicit it is not a fanboy move — "Opus and GPT, why not own both?" — but the working conclusion is that Opus is currently the wrong main.

NOTE on Mythos vs Mephisto: Video 1 names the next model "Mythos"; Video 3 names the same compute sink "Mephisto." Both are host / community reporting, not Anthropic official channels. The V5 leak string is "Mythos 5." Treat all three as unverified codenames for the same rumored training run.

Video 4 — Opus 4.7 is disappointing

The verdict on the upgrade, posted after launch. Anthropic's marketing says Opus 4.7 is "substantially better at following instructions" with almost 10% higher SWE-bench Pro, and a new "extra high" reasoning tier. Tokenization changes mean it costs roughly 1–1.3x more than 4.6 per call. The community disagreed: Reddit's top comment calls it "a serious regression, not an upgrade," and the car-wash sanity check (50m away, walk or drive?) trips 4.7 — it tells the user to drive.

The channel's own launch-day test (verbatim structure):

  • Ran the same V3 suite (tabs-not-spaces, order things, functions under 10 lines, error handling) twice.
  • Opus 4.7 landed at the same level as Opus 4.6 — "fails their own tests."
  • GPT 5.4: 75% on the same suite.
  • A one-prompt space shooter came out with broken F-to-fire controls and stiff physics.
  • GLM 5.1, priced at $72/mo on Z.AI's coding plan (up from $30), produced a visibly smoother game on the same prompt, with Z.AI raising prices specifically because "our competitors are giving you slop."

The deeper theory — the load-bearing piece. Compute is being redirected to Project Glasswing / MEOS, the enterprise-only successor already in use at Apple, Cisco, CrowdStrike, Google, and Anthropic itself. The consumer Opus is left quantized and rate-limited while peak launch traffic saturates capacity. System prompt leaks suggest 4.7 is in theory stronger than 4.6 — "the lever just isn't being turned for the $20–$200 subscriber tier."

Where it stands: Opus 4.7 is "usable," better than the mid-week low point, but nowhere near the February–March 4.6 baseline, let alone GPT 5.4. The host is staying on GPT until Anthropic reallocates compute.

Video 5 — Anthropic releasing Opus 4.7 TOMORROW?

The pre-release tease, posted 24–48 hours before 4.7 went live. Three independent signals, all captured in the video:

  • The Information: 4.7 ships "this week" with a new AI design tool.
  • Vertex AI catalog: Anthropic's model accidentally listed in Google's Vertex marketplace, the standard 24–48-hour pre-release tell.
  • Polymarket: odds for an April 17 release jumped from 20% to 98% — the cleanest timing signal in the video.

The 4.6 framing. Current Opus 4.6 is "performing like absolute dog trash" and "literally unusable" for the same $20/mo subscription price. The sharper theory: Anthropic deliberately degraded 4.6 to upsell 4.7, mirroring the 4.5 dumb-down right before 4.6 launched. The release cycle is now ~73 days, suggesting a planned squeeze rather than a reactive fix.

The bridge stack the host is using in the gap:

  • GPT 5.4 — "actually good."
  • Kimi 2.6 — under active test.
  • Kilo Code and Codex — for agent work, specifically because both let you swap models mid-task.

The leak (this is the load-bearing piece for Video 6's title). A Claude Code source-code leak surfaced strings for Opus 4.7, Sonnet 4.8, and "Mythos 5" — a much larger model currently gated to security researchers and partner firms. That "Mythos 5" string is the reason the syllabus titles the next video "LEAKED Claude Mythos" — but the video at that ID is actually about Fable 5, not Mythos 5. See Video 6.

The money signal: insiders are shorting Adobe and Figma on the assumption that 4.7's design tool is a direct competitor. The host frames this as Apple-style disrespect: "they can disrespect your intelligence… because they can."

Migration warning (verbatim from the host): "Budget at least one month" for any model swap — the host's own presentation maker is still broken post-migration, stuck in a "revolving door" of fixes.

Video 6 — Claude Fable 5 + Loop Designs is TOO STRONG! (Full Tests)

NOTE on the Video 6 divergence (resolved). The syllabus title "LEAKED Claude Mythos (Capybara): More POWERFUL Than Opus 4.6" does not match the canonical title in public.videos, which is "Claude Fable 5 + Loop Designs is TOO STRONG! (Full Tests)." The host's own pinned YouTube comment on this video confirms the DB title — "i wanted you guys to see the thinking blocks of Claude Fable 5 with loops to see its power." The "Mythos 5" model name is from a separate Claude Code source-code leak referenced in V5; Mythos 5 is not the subject of this video. This section uses the database title and the host-pinned comment as the source of truth. Verify the Fable 5 / Mythos 5 / Capybara codenames against Anthropic's official channels before relying on either for production work.

What the video is actually about. A vibe-coding test of Claude Fable 5 with "loop syntax" on Cursor, with a $20 budget. The thesis: loop syntax + a validation rule beats one-shot prompts on long-running build jobs.

The benchmarks (the head-to-head that matters):

  • Artificial Analysis intelligence index: Fable 5 scores 65 with fallback — the top spot. Speed ranks near the bottom; price is "definitely one of the most expensive right now."
  • Frontier Coding Diamond: Fable 5 29.3% vs GPT 5.5 5.7% (Opus 4.8 well below).
  • Multi-disciplinary reasoning: Fable 5 59% / 64.5% vs GPT 5.5 41.4% / 52.2%.
  • The host dismisses SWE-bench Pro as "cherry-picked to the max" — use Frontier Coding Diamond instead.

The cost reality on Cursor (load-bearing):

  • Budget set: $20.
  • One 3D Chinese-architecture project with loop syntax: $11.
  • Second project (W.B. Yeats' "The Second Coming" poem-to-game): total hit $32 — 72% of the monthly Cursor API allocation burned in the second week.
  • Three Claude Max plans ran out during a single BridgeMind live stream.
  • The cheap window closes June 21–22, after which Fable 5 becomes "credits only and it's twice as expensive as Opus" per token.

Loop syntax — the technique:

  • Harness command: loop until it's done.
  • Validation rule: "you have personally reviewed the HTML source to ensure that there are no placeholder comments, missing functions, or to-do's."
  • Fable 5 was caught taking Playwright headless-browser screenshots mid-run and running 6- and 11-point QC checks. A physics game passed 11/11 checks on structure fragments, static friction, and beast AI velocity after multiple self-correction loops.

Fable 5 vs Opus 4.8 vs Qwen 3.7 Max (the head-to-head that flips the V3 ranking):

  • 3D assembly: Fable 5 + loops produced a roof, working X-ray toggle, moving labels, and visible "cracks" for age erosion. Opus 4.8 handled the same prompt with a flat "color change."
  • The host's verdict: "even better than Qwen 3.7 plus and Qwen 3.7 Max," conceding it's "a little bit of cheating" because Qwen wasn't tested with loops.

The tier reshuffle (verbatim):

  • Treat Fable 5 as the new "Opus."
  • Demote Opus to your "Sonnet-tier" slot.
  • Main for orchestration and Hermes stays DeepSeek V4 Pro (51.5) / Kimi 2.6 (53.9). The host explicitly "will not main Claude yet" until the 2-week ROI frame balances out.

Common pitfalls

  • Treating "way faster than expected" as a real apology (V2). It's a pre-cursor to a quota raise, not a rollback. Watch for a number change in the dashboard, not a tweet.
  • Paying the "extra high" reasoning tier on launch week (V4). With 4.7 reportedly quantized under peak load, the lever buys you nothing.
  • Burning Max budget on one-shot tasks (V2). The 50% cap on eight trip-planning prompts is the canonical warning sign. Screenshot and post in the community thread.
  • Letting Opus 4.6 run unsupervised in plan mode (V3). It executes the wrong phase, then notices mid-run. Cap chat uploads at 100 files per thread.
  • Underestimating migration cost (V5). Budget a full month, not a week. The host's presentation maker is still broken a week after the switch.
  • Assuming Anthropic's rate limits are published (V1). They are not. The API is the only honest surface for budgeting.
  • Treating a ~10% SWE-bench Pro bump as a capability signal (V4). The launch-day suite put 4.7 at the same level as 4.6, with GPT 5.4 at 75%. Use Frontier Coding Diamond instead of SWE-bench Pro.
  • Ignoring the credits-only deadline (V6). June 21–22 is the cutoff for the cheap window on Fable 5. After that, credits only and 2× Opus per token.
  • Missing the loop-syntax savings (V6). A loop until it's done harness plus an explicit validation rule lets the model self-correct over multiple QC passes — 11/11 on a physics game is the headline.
  • Reading Mythos as a release you can test (V5). Mythos 5 is gated to security researchers and partner firms. Don't budget for access.
  • Conflating "Mythos" and "Fable" (V5 ↔ V6). They are different models in different videos. The V5 leak is the source of the "Mythos 5" string; the V6 video is about Fable 5 on Cursor.
  • Letting vendor lock-in survive a regression (V3). The host's answer to V3 was to learn Kilo Code specifically because the leading model swaps every quarter. Build the hot-swap path before the next regression, not after.

What this means

The thread is one story told six ways, and the story is the same in each video: the consumer Claude tier is being squeezed to fund three enterprise bets in parallel — the Mythos / Mephisto training run, Project Glasswing / MEOS already in production at Apple / Cisco / CrowdStrike / Google / Anthropic, and a Claude Design tool built to take on Adobe and Figma. Every video in 8.1 is a downstream effect of that allocation choice:

  • V1 is the policy decision (subsidies pulled, limits tightened).
  • V2 is the policy backtrack (community pressure, "way faster than expected").
  • V3 is the quality cost (Opus 4.6 quantized, 40% vs 63% on a Claude-designed suite).
  • V4 is the new release with the same quality cost carried over (4.7 at 4.6 levels, Glasswing gets the good compute).
  • V5 is the leak + the 24–48-hour pre-release tell.
  • V6 is the only ray of light: a leaked sibling model (Fable 5) that actually delivers on the capability promise, if you can stomach the cost and the credits-only deadline.

The bridge stack that runs across all six videos is the same: GPT 5.4 + Kimi 2.6 for the consumer-tier budget, Kilo Code or Codex for hot-swap, DeepSeek V4 Pro and Kimi 2.6 for orchestration, and loop syntax (V6) for the long-running build jobs. The honest summary of where the channel lands on Anthropic after the thread: Claude is still the strongest model, but the consumer tier is now structurally second-class — and that is not a story Anthropic is telling you, it is a story the rate limits and the launch-day benchmarks are telling you.

For Course 8 as a whole: 8.1 is the single-vendor version of the thesis that 8.3 (Meta, xAI, the wider industry) and 8.4 (the Chinese-lab thread) push to the industry level. The same compute-allocation choice Anthropic is making in 2026 is the same one the Chinese labs are making on cost, and the same one Meta is making on headcount. Read 8.1 with 8.3 and 8.4 open in another tab — the three articles are one argument from three different angles. And when you get to the capstone, the question this article forces is: is Opus 4.7 still in your default stack, or is the cheap slot now the consumer-Claude slot and the orchestration slot GPT 5.4? The other Stream 2 courses will read very differently depending on which answer you give.

Sources

  • Anthropic pulled a fast one on us! (Opus plans LIMITED) — 24,059 views — MkabEkgGpjA — host transcript and summary_content populated on 2026-06-17 re-pull.
  • Anthropic admits fault (Claude limits to be INCREASED) — 9,673 views — WiAx9sPw69U — summary_content / summary_key_takeaways / transcript_content populated; two verbatim Lydia quotes.
  • Claude Opus is ACTUALLY UNUSABLE — 21,675 views — Cc2Vvra9F_c — full transcript populated; the 40% vs 63% benchmark numbers and the Mephisto codename are transcript-anchored.
  • Opus 4.7 is disappointing — 9,557 views — vUpN_S1iGqI — summary_content / summary_key_takeaways / transcript_content populated; Project Glasswing / MEOS enterprise-deployment list (Apple, Cisco, CrowdStrike, Google, Anthropic) sourced from the summary.
  • Anthropic releasing Opus 4.7 TOMORROW? — 7,255 views — fjmg7lX4LTY — Polymarket 20%→98% jump on April 17 and the "Mythos 5" leak string sourced from the summary.
  • Claude Fable 5 + Loop Designs is TOO STRONG! (Full Tests) — 3,487 views — 8De7s6WG7Bo — host's pinned YouTube comment (@BoxminingAI, comment_id confirmed in public.youtube_comments, 3 likes) reconciles the syllabus title with the database title.

Public.youtube_comments (2026-06-17 re-pull): Video 6 returned 14 rows. The @BoxminingAI pinned comment is the source-of-truth reconciliation for the Fable 5 title. The skeptical top-reply — @dimeloloco "Tried fable. It's basically just an AI bot on an agentic loop that can sustain a task for long periods of time or is likely an orchestrator using smaller agents" — and @dimeloloco's follow-up naming OpenRouter's fusion router as the comparable pattern are the audience's "Fable 5 might be an orchestrator, not a new model" frame. @PaxAurora-n9m on the cost gap: "Definitely interesting, but I'm with you on cost. Exciting for the future when the gap closes and it's diminishing returns for over-priced SOTA models." @georgeRRmartin2's "Claude Fable 5 just finished Winds of Winter for me" is a joke reply, flagged as such. Videos 1–5 returned 0 comments on the 2026-06-17 re-pull.

Public.ai_updates (2026-06-17 re-pull): No row is dated to the April rate-limit saga, but post-thread context from the briefings confirms the enterprise-bet thesis: 2026-04-26 ("Google commits up to $40B in Anthropic"), 2026-04-29 ("Anthropic launches Claude for Creative Work with Blender/Adobe/Ableton connectors" — the same Adobe/Figma threat V5 names), 2026-05-01 (Apple "accidentally leaks internal Claude Code use" per the briefing excerpt), 2026-05-07 ("U.K. AISI reports GPT-5.5 matches Claude Mythos in autonomous cyberattack capabilities" — the only place "Claude Mythos" is named in public.ai_updates), 2026-05-22 ("Anthropic on track for first profitable quarter ($10.9B Q2 revenue)"), 2026-05-23 ("Claude Opus 4.7 ships with cyber safeguards; Claude Design launches"). These are the post-V6 receipts for where the consumer-tier compute went.

Supabase query:

SELECT video_id, title, views, summary_content, summary_key_takeaways, transcript_content
FROM public.videos
WHERE video_id = ANY(ARRAY[
  'MkabEkgGpjA','WiAx9sPw69U','Cc2Vvra9F_c',
  'vUpN_S1iGqI','fjmg7lX4LTY','8De7s6WG7Bo'
]);

The model names Mythos, Mephisto, Project Glasswing / MEOS, and Fable 5 are sourced from the channel's videos and a Claude Code source-code leak referenced in Video 5. The "Mythos 5" string in the V5 leak is a separate model from Fable 5 in V6 — they are not the same model and the article treats them as such. Codenames are unverified against Anthropic's official channels; the channel's reporting is based on X posts from the Claude Code team, community threads, and the channel's own benchmarks. No official Anthropic changelog URLs were cited for the rate-limit decisions, the "extra high" reasoning tier, the 4.7 tokenizer change, the Glasswing / MEOS enterprise-deployment list, or the credits-only pricing cutoff — verify each on Anthropic's official pricing and changelog pages before relying on any of them.