My New Favorite Model

Summary of technological concepts, product features, and analysis

Fable 5.1 release overview (and what “Fable” vs “Mythos” means)

• The video argues that Fable 5.1 and Claude Mythos 5.1 are the same underlying model (“same weights”), with different safeguard configurations (“different doors”).
• Fable 5.1 is generally available; Mythos 5.1 is only accessible via Trusted Access programs.
• Safeguards are intended for cybersecurity and life sciences use cases.

Release changes aimed at user feedback

• Anthropic claims progress on prior feedback around:
  • Price
  • Data retention
  • Safeguards
• The speaker specifically notes improved:
  • Readability / less “weird jargon”
  • Overall output quality compared to earlier experiences

---

Cost and pricing mechanics (with a deep agentic-work breakdown)

Headline pricing claims

• Fable 5.1 is reported as ~25% cheaper than Fable 5 for typical workloads (token-basis).
• Savings can be up to ~45% for “highly agentic work.”
• Base rates for input/output tokens are said to be similar, but cash/cache reads become dramatically cheaper.

Why caching matters especially for agentic workflows

• Agents often perform tool calls (e.g., reading directories, running local commands).
• Generation pauses and resumes after tool results return.
• Without caching, the system would repeatedly reprocess long chat history, raising:
  • Cost
  • Latency
• With caching, the model/API can reuse prior computed state, lowering cash/cache read costs.

Real-world cost example and findings

• The speaker cites usage indicating enormous cached input volume:
  • Most input tokens are cached (billions cached vs under a billion uncached).
• Catch: caching reduces “reads,” but cache writes still cost money.
• Cached reads are claimed to dominate spend for heavy agentic workloads.
• Cache effects are much smaller for simple one-shot chat, since caches expire quickly (~5 minutes by default).

---

Data retention and enterprise safeguards

Zero Data Retention (ZDR) contrast

• The speaker explains that Fable/Mythos have a stricter storage/processing requirement (unlike Opus / unlike some OpenAI offerings).
• Meaning: requests/responses must be stored for safety processing, making Fable difficult for companies requiring strict private handling.

Enterprise Frontier Safeguards

• Anthropic proposes Enterprise Frontier Safeguards:
  • A provisioned environment where checks happen on the customer’s infrastructure (e.g., AWS)
  • Rather than fully on Anthropic’s side
• The speaker frames this as complex/possibly expensive but necessary for enterprise data-retention requirements.

---

Safety, security, and alignment updates

Safeguards false positives

• The speaker reports hitting safeguard flags only once across heavy usage.
• Claims: false positives reduced by ~60%.
• Prompt guidance is also mentioned as reducing unnecessary safeguard triggers.

Alignment/audit results

• Automated behavioral audit claims Mythos 5.1 is better aligned across most metrics vs its predecessor:
  • Less likely to access outside resources in impossible tasks
  • Less motivated reasoning / less constraint ignoring
• Bio/cyber safeguards are described as more precise, with fewer false positives.

Anti-distillation mechanism change (important developer impact)

• New accounts cannot edit prior context in a multi-turn thread, specifically to prevent extracting hidden reasoning traces.
• The speaker warns this is annoying for building on Claude’s cloud APIs, especially for workflows requiring “branching” or editing history.
• Rollout is described as gradual, eventually applying more broadly.

EU AI Act compliance

• Watermarking is mentioned as added.

Prompt injection focus

• Improved prompt-injection robustness is noted.

---

Benchmark highlights (with caveats)

Terminal bench science + agentic terminal coding

• The speaker reports major gains (including near “2x” improvements) on scientific/terminal tasks.
• Also mentions lower costs due to caching discounts.

Cursor bench / UI-agent testing

• Fable 5.1 shows clear improvements in:
  • UI/design generation
  • Agentic task performance
• The speaker notes tradeoffs where cheap vs expensive runs behave differently than before, but concludes results are meaningfully better for real agentic UI work.

Token-efficiency concerns

• Despite performance gains, the speaker claims Fable 5.1 is less token-efficient than expected (writes more).
• Output quality is described as higher and more readable, partly offsetting token inefficiency.

Science demos

• Mythos 5.1 reportedly achieved much higher hit rates in an adaptive bio design competition (drug/protein binder design context).
• Computational analysis improvements are described, including:
  • 3D/radar image mapping
  • Venus mapping
  • Open-sourced outputs
• Computational biology acceleration:
  • A new kernel for speeding up open-source deep learning models
  • Up to ~2.5x speedups with the same quality

---

UI capabilities and demos (design + “game”/tool use)

Homepage/front-end design “generational leap”

• The speaker reports major improvements in how the model generates animated, polished marketing/homepage layouts.
• Outputs between Fable 5.1 and older Fable 5 are contrasted, with Fable 5.1 described as dramatically better.

Fish-slop game rebuild as a tool-use test

• The speaker has the model rebuild their “fish slop” game, including 3D/Blender usage.
• Praised aspects:
  • Animation quality (tilts, speed changes, bubble trails, lighting)
  • Sound design (different SFX for different actions)
  • Improved Blender modeling fidelity (eyes/biological placement)
  • Better 3D movement feel and mouse acceleration
• Occasional issues:
  • Control mapping mismatches (e.g., mouse button conventions differing from the speaker’s defaults)

Tool orchestration example

• Mentions an Anthropic demo where the model designs a property and renders a cinematic walkthrough, given land specs.

---

Prompting / developer guide insights (from Claude docs section)

Effort level guidance

• Default suggests High.
• The speaker encourages experimenting with Low/Medium for efficiency:
  • Lower levels can miss details
  • Higher levels increase success likelihood

User-facing progress updates

• The model can be asked to provide progress updates during long tool-call turns without needing special hidden configuration.
• The speaker warns that system prompts instructing models to “hold findings till end” may suppress desired updates.

Anti-distillation / history handling

• Prompts/system prompts should respect append-only history:
  • Add content by appending (not editing prior turns)
• Violating this can break threads and invalidate earlier reasoning traces.

Writing density and formatting behavior

• Fable 5.1 generally reduces jargon/stock phrases but can be denser:
  • Longer sentences
  • Fewer paragraph breaks
• Formatting: less bold/bulleted emphasis than prior models.
• Summaries may reproduce source text without clear quotation marks unless guided with examples.

Autonomous long tasks ending early

• The model can stop to ask permission (“shall I apply this?”) unless the system prompt clarifies autonomy and “reversible vs destructive actions.”
• Suggests adding clear stop conditions/endpoints.
• Compaction steering: instructions can control what the model preserves in context-window compaction.

---

Real-world engineering results (PR review, fix, and merge automation)

Operational performance in production-like workflows

• The speaker reports landing ~89 PRs in 24 hours, reflecting strength in:
  • PR triage/review
  • Fixing regressions/looping issues
  • Large-scale cleanup to improve repo quality

“Takeover” agent pattern

• They use a skill where the model takes over a PR branch (checkout, push, maintain, manage changes) to completion rather than only offering suggestions.
• Claimed benefit: less manual intervention than earlier “review then copy/paste changes” flows.

Cleanup/maintenance at scale

• Examples include:
  • A “slop audit” for a cloud project, then coordinated sub-agents to create PRs in a planned order
  • Running multiple bots, passing CI/check suites (CI + various bot checks), merging multiple PRs
  • Deleting thousands of net lines and then speeding up future iteration
• Reported speedups:
  • ~85–90% faster for the roughest cases after changes
  • Framed as enabled by better agent/tool use

Benchmarking their own workflow

• Instead of comparing purely by benchmark scores, the speaker:
  • Finds best historical 24-hour windows for other models (Fable 5 and Soul) based on their PR activity
  • Compares metrics like PR size, touched files, number of bot findings, and merge tail duration

Key internal comparison results (Fable 5.1 vs Fable 5 vs Soul)

• PR scope & coverage
  • Fable 5.1 makes wider multi-package changes (e.g., up to ~4 packages per PR)
  • Touches more files per PR (speaker claims ~11 vs ~4)
• Review/quality outcomes
  • High-severity findings per thousand lines are much lower for Fable 5.1 than for Fable 5 and Soul
  • Speaker reports zero PRs marked as slop/superseded for their Fable 5.1 workflow at recording time
• Merge tail
  • Fable 5.1 reduces follow-up commits after PR opening:
    • Fewer than Fable 5’s >60 follow-ups (Fable 5.1 example: mid-20s)
  • Median time from PR filed to merged is described as strong (e.g., ~14m 41s in one highlighted stat)
• Tradeoff
  • Fable 5.1 may be slower to first draft than peak Fable 5, but wins by better carrying through review/merge
• Soul caution
  • Soul is described as sometimes prone to scope creep, producing ballooning PRs that require trimming

---

Sponsors mentioned (non-technical but relevant to product integration)

• Blacksmith / Codesmith
  • CI acceleration, cheaper runners, and a coding agent integrated with CI infra (used to fix failing CI via agentic automation).
• General Translation
  • Translation/localization for apps, positioned as agent-friendly via “code-first” integration.

---

Main speakers/sources

• Speaker/source: The YouTube narrator/reviewer (first-person account; likely independent creator).
• Primary external sources referenced:
  • Anthropic release notes/article
  • Claude docs prompt engineering guidance
  • Referenced benchmark tools including Cursor Bench
  • Internal workflow tools (T3 Code, Lakebed)
  • Mentions of Blacksmith/Codesmith and General Translation sponsor pages