Claude Opus 5 Is Here. The Bottleneck Is Somewhere Else

What is growing is adoption, not maturity

Anthropic released Claude Opus 5, its twelfth model in twelve months. Add the other vendors and you get roughly 25 notable model releases in the same period, about one every two weeks. The number of companies that rebuilt a workflow in those same twelve months is a very different number. That is the point of this article: what is growing right now is AI adoption, not AI maturity.

I am not talking AI down. I work with these tools every day and run more than 15 AI projects in production. That is exactly why I am writing this. In most companies the bottleneck is not the model. It sits in the workflow the model gets dropped into, and no version jump fixes that.

There is a term for this condition that carries more weight than any benchmark table. Claire Vo reviewed Opus 5 independently, calls the model brilliant, and still does not swap out her tools. She describes it as an “intelligence overhang,” meaning more capability than existing workflows can actually draw on (Lenny’s Newsletter, July 24, 2026). That overhang grows with every release. You do not close it with an update. You close it by working on the process.

What Opus 5 does, without the hype

The model costs 5 US dollars per million input tokens and 25 US dollars per million output tokens. That is the same price as its predecessor Opus 4.8 and half the price of Claude Fable 5. It is available immediately in Claude.ai, Claude Code, Claude Cowork, and through the API as claude-opus-5. The main new feature is effort levels, which let you trade intelligence against token consumption per request. There is also a Fast Mode at double the price.

On benchmarks, Anthropic reports Frontier-Bench v0.1 at 43.3 percent versus 33.7 percent for Fable 5, ARC-AGI-3 at 30.2 percent versus 7.8 percent for GPT-5.6 Sol, and an Elo rating of 1861 on GDPval-AA v2. The Register puts Opus 5 at the top of the Artificial Analysis Intelligence Index with 61 points, one point ahead of Fable 5, but also names the weak spots: out of 14 exploit tasks the model solved only 4, while Mythos solved 13. Anthropic itself notes that responses run longer than with earlier Opus models. If you pay by output token, you see that on the invoice.

One small detail shows how fast this is moving. Anthropic and most outlets date the release to July 24, 2026, while The Register and the-decoder write July 25. When even the release date is contested, be careful with rankings that pretend to be accurate to the week.

Roughly 25 models in twelve months

How closely these follow each other is easiest to see in a list. Twelve of them come from Anthropic alone. For the models from other vendors I have no month, only the period, so they sit at the end of the table.

When	Vendor	Model
August 2025	Anthropic	Opus 4.1
Fall 2025	Anthropic	Sonnet 4.5
Fall 2025	Anthropic	Haiku 4.5
November 2025	Anthropic	Opus 4.5
February 2026	Anthropic	Opus 4.6
February 2026	Anthropic	Sonnet 4.6
April 2026	Anthropic	Opus 4.7
May 2026	Anthropic	Opus 4.8
June 2026	Anthropic	Fable 5
June 2026	Anthropic	Mythos 5
End of June 2026	Anthropic	Sonnet 5
July 2026	Google	three Gemini variants on a single day
July 2026	Anthropic	Opus 5
Same window	Google	Gemini 3.1 Pro
Same window	OpenAI	GPT-5.4
Same window	OpenAI	GPT-5.5
Same window	OpenAI	GPT-5.6 family
Same window	DeepSeek	DeepSeek V4
Same window	Moonshot AI	Kimi K2.7 Code
Same window	Moonshot AI	Kimi K3
Same window	Zhipu AI	GLM-5.2
Same window	xAI	Grok 4.5
Same window	Meta	Muse Spark 1.1

I say roughly 25 on purpose, not 24 or 26. Release trackers disagree on individual dates, for example on Grok 4.5 and on the GPT-5.6 family, where some count the announcement and others general availability. One tracker counted seven frontier models in 78 days between February and April 2026 and turned that into a new state of the art roughly every eleven days. llm-stats lists 335 tracked model releases in total.

For a business that creates an awkward situation. The model you tune your workflow to today will not be the best one in three months. Concluding that you must constantly switch gives you a permanent construction site. Concluding that you should wait until things settle down means waiting for something that is not coming.

The counter-check: what actually lands in companies

Start with the number that argues against my thesis, because it deserves to go first. The German industry association Bitkom surveyed 604 companies with 20 or more employees and reported on March 11, 2026, that 41 percent actively use AI, up from 17 percent the year before. Among users, 77 percent say their competitive position improved and 52 percent report a measurable contribution to business results. Anyone claiming nothing is changing is contradicted by that survey.

But the number of users is not the number of changed workflows. And it is less clear-cut than a single quote suggests. Bitkom’s own 2026 study report puts a summer 2025 survey at 36 percent and calls that a doubling from 20 percent in 2024. Germany’s Federal Statistical Office arrives at 26 percent for 2025, broken down into 23 percent of small, 36 percent of medium, and 57 percent of large companies. Three surveys, three sample populations, three numbers. Whichever one you quote, name the source, otherwise you invent a precision that does not exist.

What happens behind adoption is more interesting than the adoption rate anyway. For its study “State of Marketing Europe 2026,” McKinsey surveyed 500 marketing leaders in Germany, France, Italy, Spain, and the UK. 94 percent of organizations sit at low or medium maturity on generative AI, held back by cautious leadership, missing skills, and scattered one-off projects. The remaining 6 percent report around 22 percent efficiency gains over two years and expect roughly 28 percent by 2027. The priority list says the most: brand building ranks first among 2026 trend topics, artificial intelligence ranks seventeenth. (A note on the source: mckinsey.com currently does not respond to requests. The date of November 21, 2025 comes from the URL of the German press release and from two independent secondary sources dated November 22, 2025 and December 16, 2025 that name the study by title.)

The MIT NANDA report “The GenAI Divide: State of AI in Business 2025” points the same way, finding no measurable impact on the bottom line for 95 percent of enterprise AI projects. The authors explicitly blame missing integration into workflows and structures rather than model quality. That 95 percent has been circulating as a headline for months, and it is disputed. Paul Roetzer criticizes that the study builds its core finding on just 52 interviews, defines success too narrowly, and does not disclose how the 300 evaluated projects were collected (Marketing AI Institute, August 26, 2025). So I treat the figure as a direction, not a measurement.

Then there is the point that causes the most trouble in practice. According to Bitkom, only 23 percent of companies have written down any rules for using AI, while more and more employees bring private accounts into work (press release, October 21, 2025). Barely a quarter with rules in place. That is the real maturity number.

The objection worth taking seriously

There is a good counterargument, and leaving it out would be dishonest. METR has been measuring for years how long a task can be before a model can no longer finish it on its own. That time horizon doubles roughly every 196 days on the long trend, every 131 days since 2023, and every 89 days since 2024 (METR Time Horizon 1.1, January 29, 2026). Progress is measurably accelerating, and that is not a vendor marketing number.

If the horizon keeps doubling every three months, the answer to what you can hand off to a workflow eventually shifts too. Tasks that fail on length today will run through. Opus 5 supplies an example itself. On Frontier-Bench the model reconstructed a machine part from a drawing by programming its own computer vision pipeline, something other models failed at after five attempts. The source for that is Anthropic’s own benchmark, which is worth saying out loud.

Both things are true at once. Models are improving fast, and workflows are not keeping up. If you are thinking about AI agents, the practical side of this is in my article on CEO-Bench, where autonomous agents were supposed to run a startup and a simple rule-based heuristic beat all of them.

A day with Opus 5 on my own website

I spent July 25 working entirely with Opus 5, not in a chat window but with a team of specialized agents on my own company repository and on this website. One finding from that day sums up my point nicely.

The trigger was a price increase for AI work. Update one number, I thought, half an hour of work. In reality that number sat in 88 places across 22 pages. On top of that came a shared component embedded on 222 pages that would have contradicted the new figure if it had stayed as it was. By hand I would have missed part of it, and specifically the inconspicuous part that would then have gone unnoticed for months.

Now the honest punchline, which matters more to me than the numbers. I cannot tell you whether that was the model. It may just as well have been the fact that my workflow was properly written down for the first time: clear responsibilities, fixed rules, documented context. It was probably both, and I cannot separate the shares. That uncertainty is my thesis in short form. Resolve it immediately in favor of the model and you buy a subscription. Resolve it immediately in favor of the process and you miss where this is going.

What I recommend

The recommendation is not to ignore the update. It is this: first measure what your workflow costs, then swap the model. In that order.

1. Pick one workflow that costs real money. Quoting, product copy, inbound request handling. One, not five. Scattered one-off experiments are exactly why 94 percent never get past medium maturity, according to McKinsey.

2. Write down the current state. How many minutes per run, how often per month, what goes wrong regularly. Without that baseline you cannot prove the value of AI afterwards, and every project turns into a matter of belief.

3. Describe the context instead of tuning the prompt. What the model needs to know about your company, which rules apply, what a good result looks like. That is the work context engineering refers to, and it survives every model change. Pure prompt engineering does not. I described what a growing knowledge base for this can look like using Andrej Karpathy’s LLM wiki as an example.

4. Write down the rules. What may go in, what may not, who checks the output, what happens to customer data. If barely a quarter of companies have done this, it is the cheapest advantage available to you. In the EU, transparency obligations under the AI Act start applying on August 2, 2026 anyway.

5. Only now choose the model. Once those four steps are in place, moving to a new model is one line in a configuration rather than a rebuild. Then you can run Opus 5 against the previous one and see in your own numbers whether it pays off.

Conclusion

Claude Opus 5 is a strong model at an unchanged price, with real progress and with weaknesses that Anthropic partly names itself. If you already work with Claude, try it. That costs you almost nothing.

What it is not is the answer to why AI is not yet delivering what you hoped for in your business. That answer sits in your workflows, and in most places little has happened there across twelve months and twelve models. The overhang between what the models can do and what your processes draw on grows with every release. It does not close by itself.

If you have a specific workflow in mind and want to know whether AI pays off for it, I am happy to look at it. As an AI consultant I always start with the process rather than the tool, and the AI solutions page gives an overview of what is possible. For recurring processes, workflow automation is the right entry point. Rates for AI work are listed openly on the pricing page. How I work with these tools myself is described in my article on my two Anthropic certifications for Claude Code.

Sources and further reading

• Anthropic: Claude Opus 5
• The Register: Anthropic debuts Opus 5 at half the price of its Fable sibling (July 25, 2026)
• the-decoder: Anthropic introduces Claude Opus 5 (July 25, 2026)
• Lenny’s Newsletter: Claude Opus 5 review (July 24, 2026)
• METR: Time Horizon 1.1 (January 29, 2026)
• Bitkom: Companies and AI (March 11, 2026)
• Bitkom: AI study report 2026
• Bitkom: Employees use shadow AI (October 21, 2025)
• Federal Statistical Office of Germany: Use of artificial intelligence in companies
• McKinsey: State of Marketing 2026 (November 21, 2025)
• Tom’s Hardware on the MIT NANDA report
• Marketing AI Institute: criticism of the MIT study (August 26, 2025)

All links retrieved on July 25, 2026.