Skip to content
How I AIHow I AI

Claude is BACK with Opus 5.5

I’ve been off Claude for months. Not because it got dumb, but because it got annoying. The rambling, the hedging, the preachy little disclaimers on tasks that didn’t need them. I moved most of my daily work to Codex and I didn’t miss it. Then Anthropic shipped Opus 5.5: 40% cheaper than Opus 5, faster, and with what they’re calling a fundamentally different alignment approach. I ran it for a week across real work, including four long-running agentic tasks, a full ChatPRD homepage redesign, an SVG benchmark, and one very firm refusal, and I’m ready to give you the honest verdict. There’s a lot to like. There are still two things that drive me a little crazy. And there’s one capability I genuinely wasn’t expecting. *What you’ll learn:* 1. Why I walked away from Claude entirely, and what it took for me to come back 2. The real cost math on Opus 5.5 and why pricing matters more for agentic work than single prompts 3. What happened when I ran four long-running agentic tasks, including one that tried to manipulate Claude mid-run 4. Why Opus 5.5 is now my go-to for frontend prototyping, and where it still lets me down 5. The one capability I genuinely didn’t see coming, and no other model in my stack can match it 6. The moment Opus 5.5 told me flat-out no, and what that says about where Anthropic’s safety posture actually lands in practice 7. Where Codex still wins, and how I’m splitting my model stack after a full week of testing *In this episode:* (00:00) Why I stopped using Claude (01:02) What Anthropic says Opus 5.5 is (01:54) Cost, speed, and benchmark overview (03:20) Safety, alignment, and the cybersecurity limits (05:02) How I AI bench (05:39) Voice test: is it actually not annoying? (07:54) Long-running agentic task results (10:50) Frontend prototyping (17:23) Writing voice and email (19:41) SVG illustrations (20:46) Video editing (21:42) My verdict: what it’s good at, what it still isn’t *Tools referenced:* • Claude Opus 5.5: https://www.anthropic.com/claude-opus-5-5 • ElevenLabs MCP connector: https://elevenlabs.io/mcp • Codex (OpenAI): https://openai.com/codex *Where to find Claire Vo:* ChatPRD: https://www.chatprd.ai/ Website: https://clairevo.com/ LinkedIn: https://www.linkedin.com/in/clairevo/ X: https://x.com/clairevo _Production and marketing by https://penname.co/._ _For inquiries about sponsoring the podcast, email jordan@penname.co._

Claire Vohost
Sep 22, 202624mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 1:00

    Why Claude fell out of favor: “smart but annoying”

    Claire explains why she stopped using Claude for months despite earlier excitement around Fable and Opus. The core issue wasn’t intelligence—it was the interaction style: rambling, frustrating, and hard to work with conversationally.

    • Loved early Fable’s intelligence, then it was removed
    • Opus/Sonnet felt frustrating to converse with
    • Too much rambling and “Claude slop”
    • Abandoned Claude for most interactive work
  2. 1:00 – 2:01

    Opus 5.5 first impression: faster, cheaper, and finally less annoying

    Anthropic’s pitch is Fable-like performance at lower cost and higher speed. Claire’s bar is simpler: is it still irritating to use? Her early access testing suggests Opus 5.5 is now “minimally annoying,” bringing Claude back into rotation.

    • Claimed ~40% cheaper than Opus 5 and ~30% faster
    • Claire prioritizes usability/voice over raw benchmarks
    • Early access across coding, knowledge work, and chat
    • Conclusion up front: interaction is much improved
  3. 2:01 – 2:31

    What Anthropic says it is: product tour, pricing, and responsiveness

    Claire runs through Anthropic’s positioning and basic specs, including pricing and speed. She notes Opus 5.5 feels noticeably “zippier” than Opus 5, with one UX exception she’ll describe later.

    • “Frontier performance at a fraction of the cost” framing
    • Pricing: $4 input / $20 output (higher in fast mode)
    • Cache discounts follow typical patterns
    • Perceived responsiveness is better than Opus 5
  4. 2:31 – 3:01

    Benchmarks (briefly): comparisons that matter less than lived UX

    She quickly acknowledges benchmark charts showing Opus 5.5 beating Opus 5 and matching/competing with other frontier models. But she downplays benchmark obsession in favor of practical, day-to-day usefulness and “will it blend” usability.

    • Anthropic’s charts show gains vs Opus 5
    • Competitive positioning vs other top models
    • Skepticism about benchmark storytelling and omissions
    • Main test remains: real workflows and annoyance factor
  5. 3:01 – 5:02

    Safety, alignment, and cyber/bio guardrails (and the “Claude scold” vibe)

    Claire covers Anthropic’s emphasis on safety: external evaluations, stronger alignment, and prompt-injection resistance. She also describes the practical downside—Claude’s conservative refusals and cybersecurity task bumping—framing it as a persistent personality trait.

    • First release since “pace the frontier” call; external evals
    • Stronger alignment and prompt-injection rejection claims
    • Cybersecurity requests may be bumped to a restricted model
    • Claude’s “square/scold” personality shows up in practice
  6. 5:02 – 5:32

    How she tested it: real-work bench categories and what she’s measuring

    Claire outlines her evolving “How I AI” bench and where Opus 5.5 was applied: PRDs, prototyping, code auditing, inbox triage, email writing, computer use, and creative tasks. She sets expectations: highlight strengths, show limitations, then give a verdict.

    • Bench includes PRD writing, prototyping, codebase auditing
    • Adds knowledge-work tasks like inbox triage and email-as-me
    • Also evaluates agentic voice and computer-use workflows
    • Will share selected examples rather than exhaustive scores
  7. 5:32 – 8:03

    Voice test: clearer, more GPT-like brevity—plus a silence/latency tradeoff

    She demonstrates improved conversational formatting: concise bullets and fewer annoying flourishes. The main downside is “quiet time” on long turns—less narration, but more uncertainty about whether the model is still working.

    • Outputs are straightforward and skimmable (bullets, clarity)
    • Feels tuned to be less verbose and less “Claude-y”
    • Perceived latency worsens when it goes silent for minutes
    • UX lesson: verbosity and responsiveness shape user perception
  8. 8:03 – 10:36

    Long-running agentic tasks: reliability, step depth, and real catches

    Claire reports that all four long-running agent tasks succeeded (inbox triage, backend feature build, research, computer use). She focuses on the meaningful signals: many-step endurance, catching outliers, and resisting prompt injection, with some remaining gaps like voice matching.

    • All four long-running tasks completed successfully
    • 25–82 steps per run shows endurance on multi-turn work
    • Ignored a prompt injection during inbox triage
    • Found mislinked tickets/outliers and applied policy constraints
    • Backend work caught edge cases; still benefits from cross-checking
  9. 10:36 – 12:37

    Frontend redesign win: ChatPRD homepage refresh and conversion-oriented layout

    This is where Claude shines: Claire shows a redesigned homepage that brings more content above the fold and improves visual hierarchy. She notes remaining “slop” issues like misalignment and persistent border quirks, but plans to ship the update.

    • Stronger above-the-fold layout and more visual structure
    • Improved value prop highlighting, logos placement, and reviews
    • Interactive integration filtering component stands out
    • Minor slop remains: misalignment, icon failures, border oddities
    • Net: a meaningful upgrade worth shipping
  10. 12:37 – 14:40

    Prototyping across app types: complex UIs vs “consumer app” weakness

    Claire reviews multiple one-shot prototypes—some impressively complete but overly dense (“chaos rain”). Opus 5.5 performs best on SaaS/dev-tool workflows and complex enterprise-like UIs, but struggles with consumer aesthetics (e.g., plant app) and falls back into “Claude tan/orange” styling.

    • One-shot prototypes are interactive and detailed
    • Highly dense designs can be hard to reason about
    • SaaS dashboards look clean with good rhythm and color usage
    • Consumer-style apps come out sloppy and off-brand
    • Common style artifacts: Claude tan/orange, sidebars, rounding
  11. 14:40 – 17:12

    SVG illustrations: a standout creative capability

    Claire highlights a surprisingly strong area: generating cute, consistent SVG illustrations (plants and character sets). Compared to other models, Opus 5.5 produced cleaner designs with fewer structural bugs, making it a top pick for code-based illustration workflows.

    • Generates recognizable plant SVGs (fern/cactus/snake plant)
    • Creates consistent character sets suitable for animation
    • Outperforms other models in style coherence and correctness
    • Useful new bench category: SVG writing/illustration generation
  12. 17:12 – 20:43

    Writing voice & email assistant behavior: good mimicry, but still scolds

    In agent-assistant mode, Claire finds the writing quality fine and generally non-annoying, but she dislikes the model’s tendency to moralize or refuse (“told me no”). Email voice imitation is strong—short, direct, and close to her preferred style.

    • Agent-assistant writing is acceptable and familiar for Claude
    • Can be brown-nosey and overly corrective in tone
    • Refuses risky dev instructions (e.g., YOLO to prod)
    • Email-in-my-voice output is short, clean, and usable
  13. 20:43 – 21:44

    Video editing benchmark: works, but taste and execution lag behind competitors

    Using an ElevenLabs connector and MCP workflow to cut selfie videos into shorts, Opus 5.5 underperforms. Claire cites poor taste in edits, bad color grading, weak jump-cut pacing, and mediocre overlays—suggesting this task needs specialized skill/prompting and other models currently do better.

    • Can complete the workflow, but quality is low
    • Weak editing judgment: pacing/jump cuts and overlay design
    • Color grading choices are poor
    • Other models did better on short-form editing in her tests
  14. 21:44 – 24:52

    Final verdict: where Opus 5.5 fits vs Codex and what she’ll keep testing

    Claire summarizes Opus 5.5 as “Claude is back”: less annoying, strong on agentic tasks, and exceptional for frontend and SVGs—though still conservative and sometimes slower in perceived latency. She’ll still reach for Codex for harness/app UX and computer use, while using Opus 5.5 for PR review, architecture, and frontend work, with more benchmarking to come.

    • Pros: less annoying, cheaper/faster, reliable agentic runs
    • Best-in-class for frontend design and SVG generation
    • Cons: conservative/scold vibe; perceived latency on long turns
    • Claude vs Codex split: Codex for computer use and app UX
    • Availability + increased usage limits; full bench comparison upcoming

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.