By Thomas | financial enthusiast
My AI diary: special entry.
I've been running Claude Opus 4.8 for a week now. Proper daily use. Not a benchmark. Not a promo video. Just me, my projects, and a growing API bill.
Here is what I found.
First reaction: this thing thinks differently.
I don't mean "thinks" in the mystical AI hype sense. I mean it visibly pauses on hard problems, decomposes them, then gives you an answer that actually accounts for the edge cases you forgot to ask about.
I threw it a valuation model with inconsistent assumptions. Previous models either ran with the bad numbers or gave me a generic warning. Opus 4.8 flagged the specific contradiction, suggested two repair paths with different trade-offs, and then asked which one matched my intent before proceeding. That's not magic. But it's genuinely useful. First time I felt like the model was watching my back rather than just completing the task.
What it does well.
Reasoning chains. Long, multi-step problems where earlier steps constrain later ones — this is where Opus pulls away from Sonnet or Haiku. If you're debugging complex pipelines, writing nuanced analysis, or building something that requires holding a lot of context simultaneously, the gap is noticeable.
Code too. Not "write me a function" code. Refactoring decisions. Trade-off analysis between approaches. It caught a subtle bug in my publish pipeline that I'd missed for three sessions. Spotted it in the middle of answering an entirely different question.
Writing with stakes. I used it to draft a few articles that needed careful framing — nothing went sideways. The nuance holds up even in long pieces.
What you'll hit first: speed.
Out of the box, Opus 4.8 is slower. That's the honest truth. If you're used to Haiku's near-instant replies or Sonnet's snappy turnaround, Opus feels heavy. You submit a prompt and there's a real pause.
Anthropic knows this. They shipped a Fast mode specifically for Opus — you get the full model, not a smaller one, but with optimised output speed. In Claude Code you toggle it with /fast. On the API you route to the same model ID (claude-opus-4-8) but with the flag set. The difference is meaningful for interactive work.
Still not Haiku-fast. But usable without wanting to close the tab.
The cost conversation.
This is the part nobody loves. Opus 4.8 sits at the top of Anthropic's pricing tier. Input tokens run roughly $15 per million, output closer to $75 per million. For a single article draft or a one-off analysis, that's cents. For an agentic workflow running hundreds of calls a day, it adds up fast.
Damned.
My honest take: use Opus 4.8 for the things that actually need it. Complex reasoning, high-stakes writing, architectural decisions. Run Sonnet 4.6 for everything else — it's $3/$15 per million tokens and handles probably 80% of tasks without you noticing the difference. Haiku for anything that needs to be cheap and quick.
The model tiering exists for a reason. The mistake people make is defaulting to Opus for everything because it feels safer. Your wallet will disagree.
The competitive picture.
Anthropic is in a genuine three-way race with OpenAI and Google right now. Opus 4.8 benchmarks well — top tier on reasoning evals, strong on code, competitive on long-context tasks. Whether it beats GPT-4o or Gemini Ultra depends heavily on the task.
What I notice in daily use: Opus tends to be more careful. It will push back on a bad premise rather than executing it. Some people find that annoying. I find it saves me from stupid mistakes at 11pm when I'm tired and rushing.
That said, it's not a silver bullet. On purely creative tasks — writing that needs flair, not precision — I've had Sonnet outshine it for my taste. Opus optimises for correctness. Sometimes you want something messier.
My actual setup after a week.
Opus 4.8 for planning, architecture, and anything that involves trade-off reasoning. Sonnet 4.6 for article drafts, code generation, and general daily work. Haiku when I need something instant and don't need nuance.
I didn't expect to care this much about which model handles which job. But the cost difference is large enough that it changes how you think about routing calls. If you're building anything on the API, model selection is now a product decision, not just a tech one.
Worth a trial. Just watch the dashboard.
Have you switched to Opus 4.8 yet — or are you still on an older tier and wondering if the upgrade is worth it?
