OpenAI should be scared of this one

Summary of Technological/Product Concepts & Key Analysis

Anthropic model lineup update & positioning

• The video focuses on Claude Sonnet 5.5 (Anthropic) as part of the Claude 5.5 family:
  • Fable 5.1
  • Sonnet 5.5
  • Opus 5.5
  • Mentions Haiku 5.5 coming soon.
• The creator argues Sonnet 5.5 feels strategically targeted to compete with OpenAI’s “GPT-6” / “Soul” tiering.
• The creator also claims Anthropic is “screwing over” OpenAI by offering a better fit at the budget/throughput sweet spot.

---

Official claims vs. creator’s critique

Anthropic’s stated claims for Sonnet 5.5

Anthropic claims Sonnet 5.5 is:

• 30% faster
• Up to 30% cheaper for most work
• Strong for well-scoped everyday tasks
• Good for:
  • Bug fixing
  • Producing polished documents
  • Producing slides and spreadsheets
• Has a “sharp eye for design” (the creator is not impressed by the demos shown)

Creator’s counterpoint

• Benchmarks can look good depending on how they’re framed.
• Real-world cost/performance may differ—especially based on:
  • Reasoning settings
  • Token mix

---

Benchmark highlights (and skepticism)

• Terminal Bench 4
  • Sonnet 5.5 shows a dramatic jump.
  • The creator cites approximately:
    • ~70.6% (Sonnet 5.5)
    • ~10.3% (Sonnet 5)
  • Described as:
    • “highest score ever” for Sonnet
    • “7x better”
  • Claimed implication: more viable for code
• Other benchmark notes:
  • Slightly below Opus 5.5 on a coding/general benchmark (GDP Val), which the creator doesn’t care about much.
  • Strong performance on:
    • Long-horizon work
    • Image understanding
      • Mentioned as the first Sonnet model to beat Pokémon Red using screenshots only
  • Collaboration/formatting:
    • Improved clarity in output formatting
    • Less “horrible clauses” (phrasing people hate)
    • Positioned as a better fit for agent workflows
• The creator repeatedly expresses skepticism about benchmarks, but says their own experience aligns with the results (“vibes”).

---

Pricing & cost structure: the key issue = cash reads

Baseline pricing mentioned

• Sonnet 5.5
  • $2 / 1M input
  • $10 / 1M output
  • Plus cash read pricing (creator cites 20 cents per million cash reads)

Major critique

• Even if Sonnet 5.5 is cheaper on nominal input/output tokens, cash reads can dominate the true cost in agentic workflows.
• The creator argues Sonnet 5.5 did not receive the cash-read discounts that earlier recent models (notably Fable 5.1 and Opus 5.5) received.
• Result: overall cost savings may be smaller than expected, and in some “like-for-like” tasks:
  • Sonnet can end up as expensive as Opus
  • Or close to Fable 5.1 depending on conditions
• The creator also calls out what they view as poor/unclear reporting by Anthropic:
  • Mentions missing a dashboard like OpenAI’s
  • Notes confusion in the pricing chart

---

Reasoning configuration: warns strongly against “MAX”

• “Reasoning levels” are framed as budgets, not strictly linear “more thinking.”
  • Low / Medium / High / X-high adjust allowed/required reasoning token ranges.
  • Max is described as problematic because it appears to increase the minimum reasoning tokens, causing large token spikes.
• Creator’s warning:
  • Avoid “max”
  • It can cause ~15x token increases in some comparisons.
  • Even when scores improve, it’s presented as bad value
  • It can sometimes hurt correctness/performance via overthinking
• Additional advice:
  • Avoid low reasoning:
    • “No-reasoning” mode was removed on Opus; Sonnet likely behaves similarly (models may degrade and under-reason).

---

Performance speed vs. token efficiency

• The creator says Sonnet 5.5 is faster (TPS), but not token-efficient.
• Evidence cited:
  • Third-party routing numbers
  • The creator’s own subscription metrics
• Example (“Fish Slop” demo):
  • Sonnet took 43 minutes
  • Opus took 36 minutes
  • The creator suggests generation time and token usage indicate Sonnet used substantially more tokens
• Takeaway:
  • Speed gains may not translate into faster completion when token counts are high.

---

Design/demo evaluation (Witch AI showcase)

• The creator reviews Witch AI design showcase outputs and finds:
  • Several Sonnet-generated UI/design attempts are worse than Opus
  • Sometimes worse than Fable
  • Issues noted include:
    • Scrolling glitches / odd layout behavior
    • Readability problems
    • “Hideous” card example
    • “Awful animations”
    • Weak visual hierarchy
• Conclusion:
  • Sonnet may not be the best go-to model for frontend/design tasks compared to other tiers.

---

Where Sonnet 5.5 does shine: as an agent sub-tool

• While not recommended as the primary “send prompts directly” model, the creator sees major value when orchestrated by agents:
  • Sonnet is framed as a strong choice for helping larger models plan and scope complex work.
• Example: a custom “deep audit” planning benchmark over a large TypeScript/T3 codebase:
  • Sonnet produced a plan at around half the price of Opus
  • Strong scores for architectural breakdown/planning
  • Completed in about ~5 minutes
    • Creator emphasizes “cost per point” and speed
  • Suggestion:
    • Use Sonnet for research/strategy
    • Use Opus/Fable for execution/orchestration

---

“Fish Slop” 3D/animation demo

• Creator praises Sonnet’s output quality in the demo:
  • Better animation smoothness (claimed 120 fps)
  • Fish/coral/seaweed rendering described as “decent” to “cute”
  • Aliens described as the best looking among the demos tested
• Cost comparisons:
  • Mentions Sonnet cost roughly comparable to Opus in that run,
  • with differences attributed to an accidental Opus sub-agent usage
• Rough subscription economics:
  • Creator’s estimate: ~$2,300/week usage on a ~$200 plan (cloud code subscription)
  • Claim: it “stretches pretty far”

---

Key “Tutorial/Guide” Style Takeaways

Recommendation heuristics

• Don’t use Sonnet 5.5 as the default model for every prompt.
• Especially avoid direct use for:
  • Frontend/design heavy tasks
  • Scenarios where token efficiency is most important
• Prefer using Sonnet indirectly:
  • As a sub-agent for:
    • Codebase understanding
    • Investigatory planning
    • Scoping architectural changes

Reasoning settings

• Avoid “MAX” reasoning effort:
  • Token explosion / poor value
• Avoid “low” reasoning:
  • Models degrade without enough reasoning budget
• Medium / High / X-high are generally preferred depending on the cost-quality tradeoff.

---

Main Speakers/Sources (End)

Main speaker/source

• The video author/creator (unnamed in the subtitles) discussing:
  • Anthropic Claude Sonnet 5.5
  • OpenAI competition
  • Benchmarks
  • Demos

Primary external sources referenced

• Anthropic official announcement post: “Introducing Sonnet 5.5”
• Benchmarks/dashboards mentioned:
  • Artificial Analysis Intelligence Index
  • Terminal Bench 4
  • Cursor Bench
  • Frontier Code
  • GBD Val
  • Other named benchmarks
• Demo/benchmark collaborators and comparisons:
  • References to Witch AI (creator: “Dra”)
  • Astra/Opus/Fable comparisons in agent runs