Context
Claire Vo hosts How I AI and runs ChatPRD, an AI product manager tool. She had stopped using Claude models entirely for months, not because they were unintelligent but because she found them annoying to work with: rambling, hedging, and adding scolding disclaimers to tasks that did not need them. She moved most of her daily work to Codex instead. This episode is her hands-on verdict on Opus 5.5, Anthropic's new model, tested across real PM and engineering work: PRD writing, prototyping, code review, agentic tasks, inbox triage, email writing, and creative work like SVG illustration and video editing.
The Big Idea
Opus 5.5 is the first Claude model in months that Claire does not find annoying to use, and that alone is enough to bring her back for specific tasks, even though she still prefers Codex's harness for daily driving.
Her verdict is about the working experience, not just raw intelligence. She rates it as strong at long running agentic tasks and exceptional at front end design and SVG generation, while still showing some of Claude's old scolding tendencies and a slower feel than the GPT models she has grown used to.
Key Insights
The complaint was tone, not intelligence
Claire is explicit that she left Claude because it was annoying, not because it was dumb. She describes the older models rambling, making little sense at times, and adding disclaimers she did not ask for. She would repeatedly ask it to talk to her like a human. That frustration, not a benchmark score, is what pushed her fully onto Codex for months, keeping Claude around only for a couple of background tasks.
Opus 5.5 fixed the annoying voice
Claire tested a prompt asking for fun ways to use a new feature in her product, and Opus 5.5 explained its plan first, went and read the relevant docs, then answered in clean bullet points without extra text. She calls the new voice more GPT-like, and means that as a compliment. Her one caveat: in cutting the obnoxious narration, the model sometimes went quiet for eight or nine minutes on longer turns, so it felt slow even though it was technically faster, because she could not tell if it was still working.
Four long agentic tasks all succeeded
Claire ran inbox triage, a backend feature build, long running research, and a computer use task, each running between 25 and 82 steps per prompt. All four completed. Specific catches stood out: the model ignored a prompt injection during inbox triage, caught a fake help desk ticket that had been linked to the wrong company during the computer use test, declined a customer request that did not match the support playbook, and noticed that 41 of 44 support tickets in a research task came from a single customer, then adjusted its ranking approach in the resulting strategic memo accordingly.
Cheaper long tasks changed her workflow
Because token cost is lower on Opus 5.5, Claire says she can now run it alongside ChatGPT's Codex in parallel rather than picking one. She has also started using it as an adversarial reviewer of GPT's output and vice versa, sometimes bumping GPT-written code to Opus 5.5 for a second check and finding real errors that way.
Front end design stands out
Claire had Opus 5.5 redesign the ChatPRD homepage and says she will probably ship the result. It moved far more content above the fold, added a clean interactive filter for browsing integrations by use case, and improved the visual hierarchy of the reviews section. She still spotted small slop: minor misalignment around an eyebrow icon and stray borders it could not get rid of. Across several other prototypes (a dev tools incident center, a scheduling app, a SaaS renewals dashboard), she found the same pattern: strong use of whitespace, gradients, and color for professional tools, but a tendency toward visually dense, complex layouts that are technically complete but harder to read at a glance.
Consumer app design lags
Claire says she cannot get Opus 5.5 to design a good consumer app. Her test case, a plant care app, came out with what she calls Claude's signature look: tan and orange colors, paper backgrounds, and rounded sidebars. She calls it the sloppiest output she saw from the model in this round of testing.
SVG generation is newly strong
Claire asked for SVG illustrations of small plants and separately of characters with different facial expressions, and calls both sets the best she has seen from any model so far, both in style and technical consistency (clean code, no broken details like legs coming out of a character's head). She frames this as a new benchmark category worth testing if a team has not tried it yet.
It still scolds and says no
When Claire jokingly asked to skip tests and push straight to production, Opus 5.5 refused outright, warning her that skipping tests turns a rough day into an all-nighter. She calls this a genuine refusal, not just hedging, and says she still needs to check whether other current models would have said no to the same prompt. She sees it as a real signal about how the model is tuned to behave, beyond a single joke exchange.
Mental Models & Frameworks
Perceived speed versus raw model speed
Claire separates two things that get conflated when judging a model: how fast it actually generates tokens, and how fast it feels to a user waiting on a long turn. Opus 5.5 is faster on paper, but its quieter narration style meant she sometimes could not tell if it was still working during a long task, so it felt slow despite being technically quicker. Use this distinction when evaluating any agent's user experience: verbosity and formatting choices shape the felt latency as much as raw throughput does.
Test on real work, not benchmarks
Claire's structure for this review was deliberately not a benchmark walkthrough. She says published benchmarks tend to bury the areas where a model underperforms, so she instead ran the model across real personal and product tasks (PRD writing, prototyping, code review, inbox triage, email writing, computer use, and creative tasks) and reported what she actually observed on each one.
Trade-offs & Nuance
Claude versus Codex depends on task
Claire is not moving all her work back to Claude. She still prefers Codex's harness, desktop app, and computer use experience overall, and expects that preference to be hard to unlearn. But she now splits her work: Opus 5.5 for PR review, architecture questions, and front end design, and Codex for computer use and video editing. Her split is not permanent; she says she will keep re-testing as new models (including upcoming Grok and Meta Muse models) come out.
Safety gating trades caution for friction
Opus 5.5 ships with cyber and biosafety guardrails that route cybersecurity related tasks to a different, more restricted model rather than handling them directly. Claire treats this as expected Claude behavior rather than a flaw, but notes it adds a layer of friction (getting bumped to a different model mid-task) that a user has to learn to work around.
Practical Application
Watch for silent long-running turns
If a team adopts Opus 5.5 for agentic work, expect it to go quiet for several minutes at a time on long tasks instead of narrating each step. Build a way to check task progress (logs, step counters) rather than relying on the model's own commentary to know it is still working, since Claire found herself repeatedly wondering if the model had stalled.
Second reviewer on GPT output
Route PRs or written work from one model through Opus 5.5 as an adversarial check, and vice versa, rather than trusting a single model's self-review. Claire found real bugs this way that the originating model had missed, in both directions.
Default for front end and SVG
For prototypes, dashboards, and illustration or SVG generation, Claire recommends reaching for Opus 5.5 first, based on the ChatPRD homepage redesign and the character and plant SVG tests. Skip it for consumer-facing visual design specifically, where she found it defaulted to a distinctive but low-quality look (tan and orange tones, paper textures, rounded sidebars).
Mine long agentic task signals
Before trusting a model with a long running agentic task in production, run it through several multi-step scenarios that include an adversarial element, such as a prompt injection or a mismatched record, the way Claire's inbox triage and computer use tests did. A model that only looks good on a short single-turn prompt may behave differently across 25 to 82 sequential steps.
Bottom Line
Claire Vo left Claude for months over its tone, not its intelligence, and Opus 5.5 is the first version that fixed that enough to bring her back for specific work: long agentic tasks, front end prototyping, PR review, and SVG illustration. She still prefers Codex for computer use and daily driving, and Opus 5.5 still scolds, occasionally refuses instructions, and struggles specifically with consumer app design.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Claude Opus 5.5 | Anthropic's frontier model, this episode's subject | Tested across agentic tasks, prototyping, writing, and creative work; the focus of the whole episode. |
| Codex (OpenAI) | Coding and agent harness | Where Claire moved most of her daily work during her months away from Claude, and still prefers for computer use. |
| ChatPRD | Claire's AI product manager tool | The product whose homepage she had Opus 5.5 redesign as a live test. |
| ElevenLabs MCP connector | Voice and video tooling via MCP | Used in Claire's test of cutting selfie video into short-form clips, which Opus 5.5 handled poorly. |
Notable Quotes
"I stopped using Claude because it was annoying." (Claire Vo)
"Pushing to prod while tests are failing is how so done today turns into up all night." (Opus 5.5, quoted by Claire Vo)
"Claude can come back in my house." (Claire Vo)
