Anthropic Actually Fixed Opus

Summary of Tech/Product Analysis (Opus 5.5)

Context: Opus 4.5 → Opus 5.x

• The speaker argues that Opus 4.5 fundamentally changed “coding with AI” workflows (they note it launched around November of the previous year).
• They describe later releases—especially Opus 5 and Sonnet 5—as “iffy.”
  • Opus 5 is said to sometimes produce sloppy or incorrect explanations.
  • It may also omit important details, making it feel unreliable for everyday coding.

Main news: Anthropic “Opus 5.5” feels like a fix

• The speaker expected Opus 5.1, but received 5.5 instead, and says Anthropic set a high bar.
• Early personal experience indicates:
  • Better instruction following
  • Improved day-to-day usefulness
  • Stronger writing/communication
• They call it possibly a “Fable killer,” but note they can’t prove it yet.

---

Claimed Improvements & Features in Opus 5.5

Cost & throughput

• Lower cost than Opus 5
  • The speaker cites roughly 40% lower serving cost on typical workloads at default settings.
• Token pricing changes
  • Input: $4 per million tokens
  • Output: $20 per million tokens
  • Described as ~20% less than before.
• Cache read optimization
  • ~60% cheaper cache reads
  • Approximately $0.20 per million (highlighted as a major benefit)
• Speed
  • Estimated ~30% faster (based on their timing observations)
• Important nuance
  • Opus 5.5 may not be more token-efficient overall, especially when using high reasoning settings (see “Benchmarks & cost-vs-performance” below).

Communication / safety angle

• The speaker’s biggest praise is how Opus 5.5 writes and communicates:
  • More natural and clearer
  • Easier to follow
  • Places important information up front
  • Feels like a reliable “work partner” during long sessions
• Readability as a safety benefit
  • Clearer output may help both humans and agents understand what’s happening and catch problems sooner.

Reasoning & session behavior

• Fix: reasoning effort placement
  • Previously, an “effort level” appeared at the top of the system prompt.
  • Changing reasoning levels mid-session caused billing issues described as “bust cash.”
• Now
  • You can adjust reasoning levels mid-session without triggering extra costly accounting.
• Recommendation
  • Avoid max reasoning due to:
    • cost
    • risk of looping

Data retention / enterprise enablement

• The speaker claims Opus supports zero data retention (ZDR) policies (unlike the “Fable line”).
• They argue this makes Opus 5.5 more deployable in enterprise/regulated environments where data retention restrictions are strict.

---

Benchmarks & Cost-vs-Performance (As Presented)

Performance results

• The speaker says Opus 5.5 is leading most benchmarks they reviewed (with some exceptions in certain agentic/terminal/science/workflow benches).
• They reference multiple “artificial analysis” charts, emphasizing:
  • Terminal Bench 4
    • Opus 5.5 Medium reportedly scores higher than prior Fable max
    • Example cited: ~$2.94 vs ~$20 for Fable 5.1 on max
  • Cursor Bench
    • Astra was removed due to OpenAI banning Cursor from using OpenAI models, limiting comparisons.

Cost vs score visualization

• The speaker criticizes the readability of existing charts and creates alternatives.
• They prefer log-scale cost visualization to show cost gaps clearly.
• Main claim:
  • Opus 5.5 delivers strong cost/performance on medium/high effort
  • It “sweeps the Pareto frontier” for score ranges 50+

Efficiency nuance: tokens can be expensive at “max”

• A key tradeoff is highlighted:
  • At high reasoning, Opus 5.5 may generate far more output tokens than competitors.
• Example token/task comparison cited:
  • Opus 5 (max): ~73k tokens/task
  • Fable: ~78k tokens/task
  • Opus 5.5 (max): ~120k tokens/task
  • GPT-6 Astra: ~27k tokens/task
• This is framed as an approximately 4x-ish efficiency gap, where speed doesn’t matter if generation length explodes.

Practical guidance inferred by the speaker

• Don’t use “low” (too weak for cost)
• Be cautious with “max”
  • can loop
  • can take extremely long
  • can burn huge tokens for little score improvement
• Recommended practical range:
  • Medium / High / XH high

---

“Real Use” / Day-to-Day Workflow Observations

Personal daily driver impressions

• The speaker claims heavy usage didn’t burn too much weekly quota across their accounts.
• Example workflows:
  • generating and porting a custom visualizer
  • using computer use to validate/collect data
  • finding bugs, then having the model update a repo and open a PR with fixes and verification via an agentic workflow

Comparison to other Anthropic options

• Opus 5.5 is positioned as far better value than cheaper “Claude code”-style alternatives, relative to how quickly those options might drain usage.

---

Front-End / UI Design Capabilities

Opus 5.5 vs Fable 5.1

• The speaker feels Opus 5.5’s UI/web demos are:
  • sometimes “okay,” but
  • often sloppier and more like generic templates (Tailwind-esque/pastel styles)
• They believe Fable 5.1 remains stronger for:
  • subtle design taste
  • animations and layout polish
• Counterpoint:
  • Another tester (“Mia”) reportedly generated 100 HTML pages and found Opus 5.5 to be the best she tested.

---

Demos & Agentic/3D Capability

3D generation trend

• Multiple creators are showing models generating interactive experiences, including:
  • a Dark Souls-like demo clone
  • flight simulator improvements
  • 3JS/browser-based game demos
  • a skate/Tony Hawk-style C++ clone (plus “SkateBench” vs “Skate Game Bench” discussion)
  • the speaker’s own fish-based 3D game (“Fish slop”)
• The emphasis is on browser/3D stacks like 3JS, with attention to animation quality and gameplay loop generation.

Model quirks during long agent runs

• The speaker notes occasional edge behaviors, such as:
  • “paranoia” about context window/completion
  • confusing statements about autocompaction and crash safety
• They argue benchmarks may miss these issues because many finish before compaction-related problems appear.

---

Sponsor / Related Tooling Mentioned (Not Core Model)

• The video includes ads for agent infrastructure:
  • Work OS
    • Described as an “OMD” standard for enabling agents to sign in and take actions on a user’s behalf
    • Mentions integrations with Cloudflare and Firecrawl
    • References replacing painful enterprise SSO/config setups (e.g., Okta/ADP/Duo) via an admin portal and IT link
  • Parallel
    • Fast web search + APIs/extract endpoints
    • Monitoring capabilities
    • MCP integration for agents

---

Main Speakers / Sources

• Primary speaker/source: The video narrator/reviewer (consistent throughout; not named in subtitles).
• Referenced/third-party sources:
  • Anthropic official release materials (e.g., Opus 5.5 official claims)
  • Artificial Analysis (benchmarks/charts/data)
  • Work OS, Parallel (sponsor segments)
• Named individuals mentioned in passing: “Edwin,” “Lydia,” “Sam Altman,” “Gabriel” (ex-OpenAI Sora lead), “Alex,” “Matt’s editor/assistant Alex,” “Bjon,” “Mia,” “Max,” plus others credited for demos.