Opus 4.7’s Visual Acuity

When users rhapsodize about Opus 4.7, in early trials, they often reference a dramatic enhancement of its computer vision, which is one core aspect of all AI.

“The model (4.7) has substantially better vision,” write spokespersons in Anthropic’s announcement. “It can see images in greater resolution.”

Although an early reviewer, Gabriel Anhaia at Dev, also touts other 4.7 powers, he starts with this, too.

“The benchmarks are genuinely wild,” Anhaia writes. “98.5% visual acuity (up from 54.5%), 3x resolution on image inputs, 21% fewer document reasoning errors, and it resolves coding tasks that neither Opus 4.6 nor Sonnet 4.6 could crack. I've been using it since this morning in real projects, and this is the first model update in a while that made me stop and rethink my workflow.”

54.5% to 98.5% is a major jump, indeed. Fewer reasoning errors and better coding suggests that the new Opus does have some of that magic sauce that people were looking for in new models.

I didn’t see many more zero-day reviews on Reddit or elsewhere, but here are some of the benchmarks.

Metrics for Success

Anthropic provides a list of benchmarks for Opus 4.7 against prior versions.

Among them, a SWE-bench verified agentic coding score of 87.6% for 4.7, up from 80.8% from Opus 4.6.

The new model, the company reports, gets a boost on multidisciplinary reasoning on Humanity’s Last Exam: 1.4% with tools, and 6.9% without tools.

A Trial Run

Anthropic bigwigs are also suggesting that Opus 4.7’s release will help with the overall game plan to get Mythos safely out in the world, as the announcement notes:

“We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses. What we learn from the real-world deployment of these safeguards will help us work towards our eventual goal of a broad release of Mythos-class models.”

That makes sense: even though Opus 4.7 is a less powerful model than Mythos, the principle is the same. Anthropic can assess, and see what happens.

There’s also a new Cyber Verification Program at Anthropic, presumably to tell the black hats from the white, red and grey, and a Cyber Use Case Form that would-be users can fill out.

Peeling Potatoes

As for OpenAI’s offering, colloquially known as “Spud,” there’s no real word about release, although Sam Altman has said the model is through training.

Interestingly, I asked GPT about this.

My question:

“Is Sam Altman vagueposting?”

The model answered pretty unequivocally, as follows:

“Yes - that’s a fair description, at least in the common internet sense of hinting at something big without giving concrete details. So if people are calling his posts or comments ‘vagueposting,’ they usually mean he’s signaling momentum and significance without pinning down exactly when, what the final name is, or how it will launch. That fits the usual meaning of the term.”

Touche.

Watering it Down

Oddly, some of the common sentiment over on Reddit, notwithstanding my earlier trouble finding it, is that Anthropic was “nerfing” Opus 4.6 to try to make 4.7 look better.

For the uninitiated, “nerfing” means you’re watering something down, weakening it, making it not-as-good, for whatever reason. The term apparently comes from the gaming world, where game makers might lower damage for a weapon, or, say, limit usage. The “modbot” on a forum characterizes a series of posts this way:

“There's a strong belief that Anthropic is ‘nerfing’ the current model to free up compute for training 4.7, a pattern users are calling the ‘Anthropic Cycle’ (release, impress, degrade, repeat). A few of you are also mourning the neglect of Haiku, which is basically a skeleton at the bottom of the ocean at this point.”

That’s a strange critique, but one making its rounds among critics of Anthropic’s strategy.

Choose Your Own Adventure

Part of the takeaway here is that if you don’t like Opus 4.7 for whatever reason, just wait a minute: or peruse other currently available models, which each have their own strengths. What does it mean that 4.7 has eagle eyesight? Or that Mythos slam dunks on cybersecurity? It means that users have options, to tailor use cases to what they want to do with AI. And there’s a lot being done. Stay tuned for more.