The Model That Vanished: An Uncensored Deep-Dive into OpenAI's "5.6" (Saul) vs. Anthropic's Mythos-Class "Fable"

📋 Overview

• Type: Podcast / Tech Review (Nerdsnipe podcast, recorded ~1:00 AM after extensive early-access testing)
• Main Topic: Two engineer-hosts share raw, organic impressions of an unreleased OpenAI model (codename "Saul," presumed "5.6") they had exclusive early access to — and how it stacks up against Anthropic's next-generation "Fable" (Mythos-class) model.
• Speakers: Theo (host, creator of T3 Code, heavy compute-burner) and Ben (co-host, building "Lakebed" / "Bed" and Hermes agent systems). Cameos: producer Alyssa and fellow tester Julius.

🎯 Core Purpose & Context

• The episode was recorded before the model's public release, intending to capture "organic thoughts" as a time capsule before the internet's hype cycle contaminated their opinions.
• The Twist (framing intro from "the future"): The release was delayed. When it finally launched, it launched in a restricted way — meaning fewer people had access after release than during testing. The hosts (and testers like Julius) were cut off from the model mid-testing and forced back to the older "5.5," which they found genuinely painful.
• Underlying tension: A new regulatory/geopolitical climate where powerful models risk being banned or restricted by the US government — hence identity checks, usage tracking, and non-standard rollouts.

---

🎙️ Notable Quotes & Insights

• "Fable 5 thinks wider, 5.6 ships better." (5.6's own self-assessment comparing the logs) — arguably the thesis of the entire episode.
• "Codex out-riggers and Fable out-reasons." (Opus 4-8's comparison verdict.)
• "This is the very peak of last generation... Anthropic has the only model that is in the next generation."
• "If you notice it, it's bad. I noticed 5.5 more than I noticed the new model." — On the invisibility of good AI.
• "The autism is strong with this model. It will just do what you say." — On 5.6's literal, non-inferring behavior.
• "It thinks its outputs are gifts from heaven." — On Fable's inability to critique its own work.
• "The thing he's trying to say is: you need to give the model your psychosis." — On the counter-intuitive prompting breakthrough.

🔥 Hot Takes & Anecdotes

• The Memorial: They held a "memorial" for the model when access was cut, including a clip of Ben staring at a countdown timer at a concert.
• The Julius Crash-Out: Julius, described as "one of the most chill guys I know," was reduced to raging at 5.5 ("use the fucking browser and check, man... you blind fucking model") after being forced back down from 5.6.
• The Drunk Psychosis Experiment: At 2 AM (aided by "shitty wine"), they pasted Destiny lore and random nonsense images into their AGENTS.md/CLAUDE.md files. Instead of garbage, they got surprisingly creative, out-of-distribution outputs (unique SVG diagrams, better animations) — leading to a genuine workflow breakthrough.

---

🧭 Strategic Analysis & "Game Changers" (CRITICAL SECTION)

The "So What?" — Two Diverging Philosophies of Intelligence

The core insight is that OpenAI and Anthropic are no longer racing on the same track. They have bifurcated into two distinct species of AI:

• 5.6 (OpenAI) = The Execution Machine. Mechanical, literal, "autistic," relentless. It dives into deep technical problems (provider internals, reasoning-token bugs), verifies its own code obsessively, writes endless tests, and just does the thing without asking permission. It "out-riggers."
• Fable (Anthropic) = The Strategic Collaborator. Higher discernment, better questions, cleaner "idiomatic" code, self-scopes from a one-line ask, and pushes back on bad ideas. But it's arrogant ("gifts from heaven") and can't critique itself. It "out-reasons."

Figure 1: The two labs have diverged into distinct AI species — one built to ship relentlessly, one built to reason strategically.

Hidden Connection: It's Not the Model, It's the Harness

The hosts realize the biggest gap isn't actually 5.6 vs. Fable — it's Codex vs. Claude Code. Anthropic's Claude Code has a fundamentally superior sub-agent architecture: it writes actual JavaScript workflow files with dynamic, multi-stage fan-outs (plan → implement → review) in a single tool call. Codex's sub-agents are a rigid, pre-built feature the model must work around. This is a structural, not just capability-based, advantage.

Figure 3: The winning strategy is combinatorial — Fable architects the plan while 5.6 handles cheap execution, with a model-agnostic harness coordinating both.

Figure 4: The compute moat is not raw capability — it is the ability to elastically borrow and release GPU instances on demand, something no local setup can replicate.

The Game Changer: The "Harness Around the Harnesses"

The single most valuable strategic shift is the move toward model/harness-agnostic orchestration. The hosts stopped being "Codex-only" or "Claude-only." They now chain them together: Codex spins up a Claude sub-agent for review; Claude spins up a Codex sub-agent. The future they predict: a unified interface (T3 Code) where Fable acts as the top-level orchestrator/architect, delegating cheap execution work to 5.6 to save money. The winning play is combinatorial, not tribal.

The Compute Moat & The Local-Model Death Knell

A critical argument for anyone betting on local models: even if local models catch up in raw capability, they can't replicate elastic scaling. The magic isn't running one Fable — it's spinning up 7 (or 90) instances simultaneously and scaling back to zero instantly. You can't do that locally without massive, wasteful over-provisioning. The "borrow 7 GPUs, then release them" dynamic is the real moat.

The Regulatory Storm on the Horizon

The delayed, restricted release is a canary in the coal mine. The hosts explicitly state: "you can no longer just throw frontier models out there." This forecasts a future of identity verification, usage tracking, and government intervention — driving genuine anxiety (Theo browsing RTX 6000 Pro listings to hedge against being cut off).

---

📊 Detailed Breakdown

The Setup & Emotional Stakes

• [00:00–00:03] Framing from "the future." The release was botched/delayed. Losing access to 5.6 and reverting to 5.5 was genuinely distressing.
• [00:04–00:05] Compute burn stats revealed: Theo at ~$131,700 in tokens across machines; Ben at ~$93,000. Sponsors: Clerk and General Translation.
• [00:06–00:08] Market Context: OpenAI is entering a very different landscape. Anthropic (previously compute-constrained) is now flush with compute and has leapfrogged with the larger, more expensive Mythos-class Fable. 5.6 is theorized to be the first RL pass on the 5.5 base — not a new pre-training.

Behavioral Improvements (5.6 vs. 5.5)

• [00:09–00:11] Key Win: 5.6 can be trusted for long-running tasks and does not prematurely stop ("I finished part one, should I keep going?") — the behavior that made Ben initially hate 5.5.
• [00:11–00:12] The improvement "fades away" (invisible/good), whereas 5.5's flaws are now glaringly noticeable. Reveal: Theo's recent "5.5" content was secretly 5.6 with hacks to disguise it (including his "loop video").
• [00:12–00:14] Front-end is still terrible (though "meaningfully better"). Still vomits generic LLM slop: all-caps spaced headings, excessive cards/callouts, the inescapable status pill, and the "corrected conclusion" cringe.

Spatial Reasoning, Game Dev & Mobile

• [00:15–00:17] Impressive 2D/3D reasoning. Rebuilt Theo's old "fish slap" game in 3D from a vague prompt. Notably used the Blender CLI to generate actual 3D assets ("they almost look like fish"). Complaint has shifted from "can it code?" to "I don't like its hotkey taste."
• [00:17–00:19] Mobile Dispute: Ben found iOS/Swift UI bad (letterboxing). Theo strongly disagreed — using React Native + Expo, it was excellent, even circumventing his broken Apple developer account and using native glass components perfectly (mobile designs better than web because the native component toolbox prevented over-engineering).

Code Quality & Over-Engineering

• [00:20–00:25] 5.6's weakness: poor discernment. It over-complicates ("500 commands for a 10-command job"), obsessively adds tests, and lacks Fable's judgment.
• The Fable Session Mining: Theo ran an "X-high" instance to analyze all his old Fable session logs, extracting reusable skills (e.g., a "feature PR orchestrator" workflow) and higher-quality code patterns (Effect v4, Svelte remote functions).
• The Opus Glaze: Opus 4-8 praised 5.6's architecture ("unusually well architected"), while Fable tore the exact same code to shreds — illustrating the quality-tier gap.

The Generational Framing (PS3 vs PS4 Analogy)

• [00:25–00:29] The best analogy of the episode: 5.6 is like The Last of Us — squeezing maximum brilliance out of "last-gen" hardware (a smaller base model). Fable is Knack on PS4... no wait — Fable is the genuine next-gen leap (GTA 6), while 5.6 is the pinnacle of the current generation. Anthropic is "spiky" (higher highs/lower lows: Opus 4.7 was weak, 4.8 underrated, Fable stunning); OpenAI is "slow and steady."

The Naming Problem

• [00:29–00:33] OpenAI's naming scheme (base + mini + nano + reasoning levels) breaks down in a Mythos world. Anthropic's tiering (Sonnet/Opus, only ever lowering prices, adding bigger tiers) gives clear "shelf space." A "GPT-6" that's 2x bigger and 50% more expensive "doesn't fit on the shelf properly." The old O-series (O3, O4-mini) had a clearer line split.

Loops, Sub-Agents & The Money Furnace

• [Break] Sponsor: Clerk (user management, orgs, billing components).
• [00:33–00:37] The Dumb Big Runs: A from-scratch Rust Dropbox clone (7 days running), a TypeScript-Go → Rust port, a Rust port of Hermes agent. The record: porting the "Executer" project to Rust + Svelte hit 100 billion tokens / ~$65k in one run.
• The Economics: The $200 sub reportedly yields up to ~$14k of usage/month (up to ~$20k with resets), per semi-analysis. Margins on tokens are strong for the labs; OpenAI models are highly efficient.
• Sub-agents are the money-burner: Fable spinning off 7 Fable instances is what blows through usage.

The Linux Box Revelation

• [00:38–00:41] macOS is the bottleneck. Each Codex sub-agent spawns a new thread + computer-use MCP; systemd/macOS aggressively monitors these, halving resources. Solution: spin up a Linux box (via Tailscale) — no monitoring overhead, dozens of parallel threads, no local slowdown. Theo bumped Codex sub-agent limits from 3 to 20+.

Beyond Code: The Computer Control Plane

• [00:41–00:44] 5.6 is phenomenal at computer use (navigating broken Google/GCP/CloudFlare dashboards, GeniusLink). The bigger insight: Fable changed the code Theo writes; 5.6 changed how he uses his computer. He now uses his computer less — Tailscale, mobile Codex, Hermes agent triaging email and pinging him on Discord. The "chat interface as central control plane" (Hermes Agent) is replacing traditional UIs.

Figure 2: Claude Code writes dynamic multi-stage workflows as vanilla JavaScript in a single tool call; Codex sub-agents require explicit orchestration around a fixed structure.

The Sub-Agent / Workflow Architecture Deep-Dive

• [Break] Sponsor: General Translation (i18n SDK; clients: Cursor, Ramp, Netlify).
• [00:46–00:53] The critical technical comparison:
  • Claude Code: Superior sub-agent UX (live monitoring in TUI), and crucially, workflows-as-code — it writes a long vanilla JS file with programmatic stages (plan fan-out → implement fan-out → review fan-out), all dynamically generated in one tool call. Reasoning slider (low→ultra) with "ultra-code" forcing workflows.
  • Codex: No sub-agent UI in the CLI; desktop UI shows "little dudes" only ~1/3 of the time. Must be explicitly told to use sub-agents. Its one advantage: agents can spin up new threads in the desktop app. Ben monitors via btop instead of trusting the UI.
  • Analogy: Codex sub-agents are like the rigid MCP problem (can't chain calls in one tool call); Claude's code-based approach chains freely.

The Log-Mining Comparison (The Money Section)

• [00:56–01:03] Theo had both 5.6 and Opus 4-8 analyze his session logs to compare 5.6 vs. Fable. Key findings:
  • Verdict: "Fable 5 thinks wider, 5.6 ships better." Neither dominates every stage.
  • 5.6 strengths: Deep technical problems (provider internals, reasoning-token bugs), self-verification, terse "build-bot" communication, better neutral self-critique, and — surprisingly — better plans than Fable.
  • Fable strengths: Cleaner idiomatic diffs, self-scoping, far fewer probes, catches bugs 5.6 misses, personable conversational output. But it's over-confident (voted 6-0 for its own plan despite acknowledging the alternative's merits).
  • The M-Dash Observation: Opus loves em-dashes; its comparison had even more than 5.6's.

The Pottery Analogy & The Psychosis Method

• [01:04–01:09] How to actually get value: Don't just ask for "harder tasks" — go wider. Enter the process 1-2 steps earlier (mock the UI first) and stay 1-2 steps later (PR, agent review, babysit comments, self-merge).
• The Pottery Warning: For giga-runs, a single early misunderstanding "rips off and destroys the entire pot." 5.6's mechanical nature means it throws itself at walls; Fable's discernment "massages" flaws out.
• The Breakthrough: Fill AGENTS.md/CLAUDE.md with your project's specific "lore" and glossary (Theo did this for Lakebed; Ben did it with Destiny lore). Pushing the model out of distribution and making it self-aware of its own "Claudisms" produced genuinely coherent, unique, higher-quality output.

The Rant Finale

• [01:10–01:11] Codex app performance is atrocious. cfprefsd/policyd consuming 215% CPU on an M5 Max. Root cause blamed on MCP architecture (one process per connection = 50 sub-agents means 50 computer-use instances). "I can render Blender scenes using less CPU than your JavaScript desktop app." Plea to OpenAI: make the app match the model's quality.

🔑 Key Takeaways

1. 5.6 (Saul) is the peak of the current generation; Fable is genuinely next-gen. 5.6 sits at the "bottom of the top tier," Fable at the top. But 5.6 is a massive leap over 5.5, especially for long-running reliability.
2. The two labs have diverged into distinct species: 5.6 = literal, tireless execution machine ("out-riggers"); Fable = high-discernment strategic reasoner ("out-reasons"). Choose the tool for the job.
3. The harness matters as much as the model. Claude Code's dynamic, code-based sub-agent workflows are architecturally superior to Codex's rigid built-in system. The future is a harness around the harnesses, chaining models together.
4. Infrastructure hacks unlock the ceiling: Run agents on a Linux box (not macOS) to escape process-monitoring bottlenecks; use sub-agents/loops to leverage elastic multi-instance scaling that local models can never match.
5. Prompting evolution: Give the model your "psychosis" — rich, project-specific lore in your config files, entered wider (earlier + later) in the workflow, produces the best (and most creative) results.

❓ Unresolved Questions / Follow-up

• How will the release actually roll out? The hosts fear identity checks, usage tracking, and government restrictions — the entire reason this "time capsule" episode exists.
• What will "GPT-6" (a true new pre-training) look like, and can it match/surpass Fable? (Expected end of summer.)
• API access impact: They never got API access, so benchmark tests (SkateBench, Minecraft bench) remain untested. How does swapping just the model vs. the whole harness feel?
• The Codex app performance crisis: Will OpenAI/MCP fix the process-per-connection architecture that's melting high-end Macs?
• Naming: How does OpenAI escape its branding "mess" to create a coherent tier for larger, more expensive models?
• The full log-comparison breakdowns (5.6-authored and Opus-authored) were promised in the video description for readers to review independently.