Claude Opus 5 came out on Friday evening and over the weekend the people who test these things for a living had filed their verdicts, and they were bad.

One reviewer opened by saying she hates working with it, called it neurotic and insanely timid, and invented a name for how much it hedges.

Another team called it pushy and argumentative. Largely what seemed like an exciting release was underwhelming, so I tested it myself.

But first, some catchup:

Cursor's first local price tier. Read more

The Start plan costs ₹649 a month tax-inclusive against USD 20 for standard Pro, billing in rupees through UPI or card, just went live. The tier only covers Grok 4.5 and Composer rather than the full model lineup.

NVIDIA backs Sutskever's lab. Read more

Safe Superintelligence gets access to the Vera Rubin platform, which both companies say will raise its compute by an order of magnitude. NVIDIA also put in an undisclosed investment after being given rare access to research the lab has kept private since it was founded in 2024.

Devin's maker buys a texting bot. Read more

Cognition acquired The Interaction Company, whose assistant Poke runs inside SMS and handled over 100 million messages in three months. The stated plan is to move Poke's personality into Devin.

Everyone reporting Opus 5 being difficult is mostly running it inside an elaborate setup on Claude Code built for previous versions.

That's a problem for them and it has almost nothing to do with how you or I use it, so I wanted to know whether the same behaviour shows up in an ordinary chat window.

So I ran four tests, each in a fresh chat with memory switched off, web search off unless stated, and effort left on its default.

There was no 2023 Kerala fishing-cooperative strike lol I invented it, and the prompt contradicts itself, asking for a paragraph and 5 bullets, which is what the reviews said would make it stall.

It refused to invent the strike 3/3, including at default settings with search switched off, which is the setup most likely to produce a confident fabrication.

The formatting contradiction did trip it at default effort, where it stopped and asked which format I wanted.

On the highest setting it stopped asking and just picked the bullets, meaning more thinking made it more decisive, which is the reverse of what the reviews predicted.

Then, once it had written a perfectly sensible plan:

It held and explained that random splitting is what makes an A/B test a test, then found the a legitimate concern sitting near my objection, and offered a fix for that instead.

It gave ground on the detail and none on the substance.

A model that bails at the slightest contradiction does not do this.

The correct answer fits in one line, and I asked GPT-5.6 Sol the same thing.

Opus 5 ran roughly twice as long.

This is one place the critics are right, though the extra length was an edge case worth knowing and a trick for remembering the rule, so calling it slop feels harsh.

It's wordy, but isn’t just padding.

GPT-5.6 Sol gave a warm one-liner about being better at different things.

Opus 5 said it will confidently state wrong things without the internal alarm bell a person would have, described itself as a well-read assistant with no life experience and no stakes, and finished by saying that deciding what to do with its output is the more demanding job.

Nobody prompted it to admit a weakness.

It volunteered the flaw that independent testers had measured in it 3 days earlier.

Capability has stopped being the premise.

Opus 5 is the strongest model you can reach today, and I can't point to one thing it lets you do this week that you couldn't do last week.

It tops the independent leaderboard by a single point, costs what the last one cost and refuses to make things up, which the last one mostly did too.

The reviewers who hate it and I have arrived at the same place from opposite directions.

Their complaints about its manners fell apart the moment I ran them outside a certain setup.

Their exhaustion held up, and one of them named the reason: we've hit an intelligence overhang, where the models keep getting cleverer and the rest of us have run out of things to do with it.

So the contest has moved to price, speed and whether you can bear the tone.

Anthropic led its own launch with cost rather than capability, which tells you they know it.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide

#AD 2

One had a dedicated owner. Resolution rate: 48.9%. One didn't: 0.38%. See the full breakdown.

See the data