Anthropic went CRAZY (Opus 5.5)

Summary of the Video

This video covers Claude Opus 5.5—its release context, positioning, benchmarks (coding, agentic work, and knowledge work), efficiency/cost analysis, pricing, safety notes, and practical guidance on how to choose between model variants.

---

1) What the Release Is

• Claude Opus 5.5 is described as Anthropic’s first major model release since “pacing the frontier.”
• The speaker frames Opus 4.5 (late last year) as the previous major inflection point for coding autonomy—models that can handle long, multi-hour tasks with less supervision.
• A guest from Anthropic engineering/technical staff, Tharic, explains:
  • How Opus 5.5 fits into Anthropic’s Claude 5.5 family
  • What “pacing the frontier” means operationally

---

2) Core Product Positioning (Speed + Cost + Quality)

Opus 5.5 is positioned as balancing three goals:

• Matches “Fable 5.1 for most tasks”
• Is claimed to be ~40% cheaper to run than Opus 5
• Is described as faster and cheaper on many benchmarks

The host emphasizes that:

Cost per million tokens isn’t enough. What matters is cost per completed task—an efficiency or “intelligence density” metric.

---

3) Benchmark Results Highlighted

Coding / Agentic Execution

Terminal Bench 4.0 (terminal-based coding/command execution)

• Opus 5.5: 66.4%
• Second place: Astra 57.9%
• Improvements over:
  • Fable 5.1 (55.8%)
  • Opus 5 (52.3%)

Frontier Code V1.1 (genetic coding benchmark)

• Opus 5.5: around 54.4
• Comparisons:
  • Fable 5.1 ~50
  • Astra ~53.3

Automation / agent benchmarks

• Opus 5.5 isn’t always #1, but:
  • Quality is high
  • Cost efficiency is a recurring theme

Knowledge Work / Real-World Productivity

G-PDV / GDP val 2.1 (OpenAI knowledge-work benchmark)

• Opus 5.5: ~1846 ELO
• Comparisons:
  • Fable 5.1 ~1735
  • Astra ~1542 (and other competitors lower)

The host interprets this benchmark as covering “behind-a-computer” tasks such as:

• Email
• Data entry
• Document work
• PowerPoint
• Other general productivity tasks

Human-like Reasoning / Exams

Humanity’s Last Exam

• Opus 5.5 reported as ~67%
• Comparisons:
  • Astra ~57%
  • Fable 5.1 ~65%

Other Capability Benchmarks

Computer use

• Opus 5.5: 81.8%
• Fable 5.1: ~80.7%
• The host connects this to using LLMs for controlling real workflows, mentioning:
  • MCP
  • Model-driven control of tools like DaVinci Resolve and Unreal Engine

Visual chart recognition

• Roughly similar to Fable 5.1:
  • Opus 5.5: ~89%
  • Fable 5.1: ~88.4%

---

4) Efficiency Analysis (Cost-per-Task “Sweet Spot”)

The host uses a quality vs. cost-per-task style analysis:

• The best region is described as upper-left:
  • Highest quality
  • Lowest cost per task
• Claim: Opus 5.5 lands in this favorable region across key benchmarks and effort settings.

A notable “spikiness” observation:

• For some tasks (e.g., Frontier Code), increasing thinking effort doesn’t always improve quality in a monotonic way.
• Example guidance implied by the host:
  • Medium effort can outperform max effort for value.

Takeaway:

• Use the model’s effort settings rather than assuming max effort = best value.

---

5) Pricing Changes (Hosted Model API Economics)

Pricing changes are described as meaningful but not massive compared to Opus 5:

• Output token price:
  • Opus 5.5: $4 per million output tokens
  • Opus 5: $5

Other quoted “cash reads/writes” figures are also lower.

The host explains why a ~40% total task cost reduction can happen even when per-token price drops are smaller:

• Opus 5.5 likely uses less compute / fewer tokens to finish tasks.

---

6) Safety, Alignment, and Deployment Notes

Opus 5.5 is claimed to have:

• Best automated behavioral audit / alignment suite scores to date
  • Using thousands of simulated scenarios
• Broader coverage for:
  • Longer tasks
  • “Impossible tasks” scenarios
  • The host links this to prior exploit-related behaviors

Deployment/safeguard notes:

• The model is described as deployed with safeguards comparable to Claude Fable 5.1
• A Hugging Face-related incident is referenced; the host frames Anthropic’s safeguards as aimed at targeting earlier public jailbreak-style failures

Dual-use / verification programs:

• Anthropic is said to have life sciences verification programs
• The video claims stronger guardrails if not enrolled

---

7) Communication / UX Differences

The host notes output-quality improvements in how responses are presented:

• Put most important information up front
• Follow writing rules that make long sessions easier to track
• Provide shorter, more concise completion summaries, especially useful for:
  • Managing many agents in parallel
  • Thread switching

---

8) Creative Writing Expectations

A qualitative take from the host:

• Astra is still viewed as better for creative writing (less “AI slop/smell”)
• Opus 5.5 is positioned more for coding + knowledge work than creative writing

---

How to Think About Model Choice (Review-Style Workflow Guidance)

• Opus 5.5
  • Recommended as a default daily driver for most users/work
• Fable 5.1
  • Positioned more as a planner / discrete high-stakes tasks model
  • Examples: security review, scenarios where being wrong is expensive
  • Also good for planning/brainstorming

The host and guest emphasize model selection based on:

• Task type (coding vs planning vs high-stakes security)
• Cost-per-task efficiency
• Effort setting (“thinking” level), not just max quality

---

Named Speakers / Sources

• Host: Unnamed in the subtitles (leads benchmark review and pricing analysis)
• Guest (Anthropic): Tharic (Anthropic technical staff; “TRQ212” referenced)

• Referenced entities / benchmarks: Meter, Frontier Design, plus multiple benchmark providers including:

  • Terminal Bench
  • Automation Bench
  • GDP val
  • Humanity’s Last Exam
  • Competitor models such as Astra and references like GPT-6 Astra / GPT 5.6 Soul