Claude Opus 4.7 lifts vision, holds capability back

The release of Claude Opus 4.7 shows two things happening at once. Anthropic has made its frontier model more capable on professional work. At the same time, it has held capabilities back for safety reasons. Both matter for UK SMEs, and neither story is getting the attention it deserves.

Anthropic released Claude Opus 4.7 on 16 April 2026. It replaces Opus 4.6 as Anthropic's generally available top-tier model. Pricing stays the same at five dollars per million input tokens and twenty-five dollars per million output tokens.

What sits behind this release is more interesting than the benchmark numbers. Anthropic has a stronger model, Mythos Preview, that it is deliberately keeping limited. Opus 4.7 was trained with efforts to reduce cyber capabilities. It ships with safeguards that detect and block prohibited cybersecurity requests.

That tells you something about where we are. The frontier is no longer just "how good is the model". It is "how much of the model are we safe to release".

This is the capability debate in plain sight

For two years, the industry's public conversation has been about raw capability. Bigger models, higher benchmarks, faster releases. Opus 4.7 is a public inflection point.

Anthropic is saying, on the record, that it has a more capable model and is holding it back. It is trialling safeguards on a deliberately less capable model first. The company's own alignment assessment describes Opus 4.7 as largely well-aligned and trustworthy, though not fully ideal in its behaviour.

Read that again. The company releasing the model says it is not fully ideal.

That is the honest framing UK business owners should want from their AI suppliers. The alternative is a race where every provider ships their strongest model as fast as possible. Consequences get dealt with later. If you are buying AI tools for a 20-person team, the supplier's attitude to its own limits matters. It matters more than one point on a benchmark.

Why literal instruction following changes how you prompt

The most practical change in Opus 4.7 is one many businesses will feel immediately. The model follows instructions more literally than any Claude before it.

That sounds good. It is good. But there is a catch in the release notes. Older prompts sometimes produce unexpected results. Previous models interpreted instructions loosely or skipped parts entirely. Opus 4.7 does not.

If you have an AI assistant running on prompts written six months ago, those prompts may now behave differently.

This is the story we have been telling clients since gecco started. Prompts are not a one-off setup task. They are a living part of how your business runs. As models get stronger, loose prompts expose their weaknesses. A prompt that said "summarise this customer email" got away with it on Opus 4.6. On Opus 4.7, the same prompt might produce output so literal it misses context an older model inferred.

At gecco we use the GRAFT methodology: Goals, Role, Audience, Format, Tone. Every assistant starts with those five elements defined explicitly. That is how you write prompts that survive model upgrades. A well-constructed prompt tells the model what you want, not what you hope it figures out.

The practical takeaway is simple. Any business running AI assistants on current-generation models should review their prompts when upgrading. Not because the prompts are broken, but because better models reward better prompting.

Better vision and better professional work

Two areas where Opus 4.7 shows its biggest real-world gains are the ones most UK SMEs will use daily.

Vision has moved forward. The model now accepts images up to 2,576 pixels on the long edge, roughly 3.75 megapixels. That is more than three times the resolution of previous Claude models. XBOW, a security testing firm, reported 98.5% on their visual-acuity benchmark versus 54.5% for Opus 4.6.

For a UK SME, that changes what you can reasonably ask AI to do with an image:

- If you are already running AI assistants, schedule a prompt review. Pick your three most-used assistants and test them against the new model behaviour. Look for outputs that have become oddly literal, overly verbose, or slightly off-topic.

- If you use image input for invoice extraction, contract review, or any visual task, retest your hardest cases. The vision improvement is significant enough that workflows that were not reliable before may now work.

- If you build client-facing documents with AI, retest your templates. Better professional output means you can often reduce the scaffolding and let the model do more.

- If you are not yet running AI assistants, the news here is simpler. The gap between human-quality professional output and AI-assisted output has closed again on specific tasks. Presentations, dashboards, reports, data extraction from images. This is the territory custom assistants operate in.

Professional work quality has also lifted. Anthropic's own testing shows the model produces higher-quality interfaces, slides, and docs. Aj Orbach, co-founder and CEO of Vercel v0, called it the best model for building dashboards and data-rich interfaces.

That matters for the work gecco builds every day. Custom assistants that draft client-ready documents, create presentations, or generate management reports. The gap between AI output that needs cleaning up and AI output that is professionally usable has narrowed again.

What this means for UK SMEs right now

If you are already running AI assistants:

If you are not yet running AI assistants, the news here is simpler. The gap between human-quality professional output and AI-assisted output has closed again on specific tasks. Presentations, dashboards, reports, data extraction from images. This is the territory custom assistants operate in.

A note of realism

Better models do not fix bad foundations. Without core documents sorted, a stronger AI model produces stronger versions of the wrong thing faster. Brand voice, audience definitions, products and services, SEO keywords, business overview. Those are the foundations.

At gecco we start every engagement with the Five Core Documents for exactly this reason. The model is only as good as the grounding you give it.

Where this leaves us

Opus 4.7 is a useful upgrade. It is better at what businesses actually use AI for. It also tells us something important about where AI is heading. The next phase is not about making models more powerful. It is about which capabilities to ship, which to restrict, and how to be honest with customers about both.

gecco builds custom AI assistants for UK SMEs using Claude, ChatGPT, Gemini, and Copilot. Our GRAFT methodology and Five Core Documents approach means the assistants we build work reliably across model upgrades.

If you want a clear read on where your business sits with AI adoption, take the AI readiness assessment. It takes a few minutes and you will get a tailored report specific to your organisation.