Video summary

FABLE IS BACK! (And Sonnet 5 is here too)

Main summary

Key takeaways

Technology

Tech/Product/Analysis Summary

1) Major news: Claude/Fable model access (policy + deployment impact)

  • Fable 5 was unbanned after export controls were lifted by the US Secretary of Commerce (the video claims Fable would likely become available by the time viewers watch).
  • The prior ban (June 12) was tied to jailbreak/hacking concerns, especially around security issue discovery capabilities.
  • Anthropic states it coordinated with the US government and committed to:
    • proactively detect and address security risks,
    • work on protocols/standards for releases,
    • inform the government of malicious activity.

Most important for developers: the withdrawn controls mean export/reexport/in-country transfer is no longer restricted.

  • The speaker suggests this likely includes hosting model APIs (e.g., running a service similar to “T3 chat”) where users access the model via API.

Implication: developers can likely offer hosted model functionality again without the previously required licensing constraints.

2) Claude Sonnet 5: “agentic” improvements, especially tool use + sub-agent behavior

Anthropic’s positioning (as quoted/discussed):

  • Sonnet 5 is built to be highly agentic, described as capable of:
    • making plans
    • using tools (browser + terminal)
    • running autonomously
    • performing work that previously required larger models

Benchmarks mentioned (and what the speaker thinks they mean):

  • Improvements vs Sonnet 4.6, including:
    • SWE-bench Pro (noted as “compromised/contaminated”)
    • Terminal Bench (improvement ~ high 60% → 80% range)
    • “Una’s last exam” (meaningful bump)
    • Computer use (slight bump)
    • Knowledge work: claimed to improve and even score slightly higher than Opus 4.8 (per the speaker’s reading of the charts)

Reasoning/tool use vs safety:

  • The speaker claims Sonnet 5 is not a special safety risk compared to some other models (they explicitly say GLM 5.2 is a more meaningful security risk).
  • However, they later criticize dual-use refusal performance, suggesting Sonnet 5 may refuse benign/suspicious tasks more often than before—leading to incorrect refusals.

3) Cost/value critique: Sonnet 5 is “agentic,” but often expensive and sometimes slower than alternatives

Key cost arguments repeated throughout:

  • Sonnet 5 can be very costly per real task—the speaker reports benchmark runs costing $300+, plus added cost from failures.
  • They argue Sonnet 5 is:
    • often not the best price/performance
    • sometimes more expensive than alternatives like GPT-class models at comparable effort
  • They reference cost-per-task comparisons from third-party reporting/benchmarks (e.g., Artificial Analysis intelligence index, Cursor Bench), claiming Sonnet 5 is far less efficient than some competitors.

Important detail: token efficiency + runtime

  • The speaker argues Sonnet 5 is a token hog (and thus slow in practice).
  • In one real attempt (rebuilding a “fish web game” from a repo), they compare:
    • Opus 4.8: ~26–27 minutes, produced a playable port with reasonable gameplay/economy
    • GLM 5.2: ~354 minutes, poor economy and rendering; also no vision (so it couldn’t use browser/visual inspection)
    • Sonnet 5: ~2 hours (possibly 2.5), with bugs and awkward UI interactions; also took longer due to many sub-agent spawns

4) The speaker’s “what’s actually new”: orchestration + sub-agents (but not reliably smart enough)

Core technical take:

  • The speaker is most impressed that Sonnet 5 shows behaviors previously associated with Fable 5, especially:
    • sub-agent usage
    • orchestrating parallel investigations
    • breaking work into smaller pieces
    • staying on task more than expected

But they also claim:

  • Sonnet 5 is not smart enough to do this at the level required to be efficient, so it can:
    • run in circles
    • split work incorrectly
    • take forever
    • raise costs dramatically

They emphasize that the right approach is model selection/orchestration:

  • Use Sonnet 5 as a tool/agent component called by an orchestrator (like Fable 5), rather than replacing Opus for everyday coding.
  • They believe better results come when Fable/Mythos (or future orchestrators) decides when to route to the right “tier” model.

5) Negative performance notes and quirks (benchmarks + real tasks)

Examples called out:

  • Cursor Bench: Sonnet 5 (most variants) looks worse than a GPT 5.5 equivalent in both performance and cost (per the speaker’s interpretation).
  • Computer use bench anomaly: the speaker is “confused” why Anthropic included it because Sonnet 5 lags behind Opus there.
  • “Skatebench” (skate tricks from descriptions):
    • Sonnet 5 (especially X-high) scored low (~37%)
    • Sonnet 5 Max scored better (~59%) but became extremely expensive
      • speaker cites ~15 cents/question, with some questions costing up to a dollar
    • The speaker claims Sonnet 5 “can’t figure it out,” leading to runs in circles and increasing token usage
      • average tokens rising from ~1,600 → 6,000 when moving X-high → Max
  • Real report about “thinking trace” leakage:
    • the speaker mentions a viewer report that on a cloud platform, Sonnet 5 may be leaking internal “thinking traces” / messy reasoning before the final answer.

6) Suggested use cases (practical guidance)

The speaker’s conclusion on “what to use Sonnet 5 for”:

  • Best role: an agentic sub-component used by a more capable orchestrator.
  • Not ideal as a direct Opus replacement for day-to-day coding.
  • The speaker plans to stick with:
    • Sonnet 5.5-style medium variants and Opus for now
    • switch back to Fable 5 when it’s available again for orchestration workflows

Tutorials/Guides mention

  • The only explicit “how-to/product setup” appears in the sponsor segment about Devon (agent software/product):
    • Spin up a cloud Linux environment for agents to run work.
    • Use agents to test website pages for:
      • mobile responsiveness
      • UI bugs
      • client-side errors
    • Demonstrate sub-agent parallelization across multiple pages.
    • Scheduled regression checks (daily) with minimal setup.

Main speakers/sources

  • Primary speaker: the video’s narrator/reviewer (mentions “my podcast co-host Ben”).
  • Key external sources referenced:
    • Anthropic (Sonnet 5 positioning + system card claims)
    • US Department of Commerce / Secretary of Commerce (Fable export restriction changes)
    • US government / BIS (Bureau of Industry and Security) (evaluation referenced)
    • Artificial Analysis (intelligence index + benchmark/cost discussion)
    • Cursor Bench
    • Ben (podcast co-host mentioned regarding Sonnet 5 impressions)
    • Devon sponsor (product: Devon agent scheduling/orchestration workflow)

Original video