Internal Access Opus 4.8 arrived barely six weeks after 4.7 with long-running agentic workflows. The acceleration is real, but the £25 per million output tokens pricing means every capability leap compounds the cost crisis for businesses running cloud AI.
Claude Opus 4.8 arrived yesterday, May 28, 2026. Anthropic is calling it a hybrid reasoning model with a million-token context window, built for professional software engineering, complex agentic workflows, and high-stakes enterprise tasks.
It can run for longer with sustained effort. It plans deliberately, uses memory across sessions, and drives long-running work forward with minimal oversight. It is designed to orchestrate complex multi-tool tasks with consistent reliability.
All of that is impressive. But the number that matters most for any business considering it is not a benchmark score. It is £25 per million output tokens.
That is the price you pay for every token the model generates. When you are orchestrating multi-step agentic workflows, the model is generating far more output than input. Every tool call, every plan revision, every verification step burns output tokens. A single complex coding session can easily consume millions of output tokens, and the cost is measured in hundreds of pounds, not pennies.
The release of Opus 4.8 is not remarkable because of what it does. It is remarkable because of when it arrived.
Opus 4.7 was released on April 16, barely six weeks ago. Before that, Opus 4.6 in February. The cadence between releases is compressing, and the gap between versions is shrinking faster than most organisations can adapt their infrastructure to accommodate them.
The acceleration is structural. Each new model has a longer context window, deeper reasoning, and better tool use. That means more tokens per query, more output per session, and more cost per interaction. The capability gains are real, but they come with a cost multiplier that compounds with every release.
For UK businesses, this is not an abstract concern. It is a budget that changes without notice, controlled by a vendor you cannot negotiate with, running on infrastructure you do not own.
The features in Opus 4.8 — sustained reasoning, long-running agentic workflows, memory across sessions — require massive compute. The kind of compute that was not available when Opus 4.7 was released.
There is a growing consensus in the industry that Anthropic held back some of these capabilities because of compute constraints. The company simply did not have enough capacity to run the heavier workloads reliably. That changed with the Colossus I deal with SpaceX — more than 220,000 NVIDIA GPUs, 300 megawatts of power, secured within months.
The compute bottleneck has been cleared. Now the heavier features can ship. But the cost of running them has not gone down. If anything, it has gone up, because the features that require more compute are the ones that burn through tokens fastest.
The community response to Opus 4.8 has been characterised by a mixture of genuine excitement and dark humour about pricing. People are already joking that it is ‘one shorting my entire weekly token allowance’. The joke works because it is closer to the truth than it should be.
The underlying reality is that every company running cloud AI is exposed to the same structural risk. You get the capability, but you never get the predictability. Your costs scale with how much you use the system, which means they grow precisely when adoption is working. It is the worst kind of cost structure for a business.
We have seen this play out already. Microsoft canceled its internal Claude Code licenses because token billing made the cost untenable, even for a company that put £10 billion into OpenAI. Uber burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37% across the board.
None of these companies are failing to use AI effectively. They are succeeding at it, which is why their costs are exploding.
Opus 4.8 is a genuinely powerful model. The sustained reasoning, the long-running autonomy, the improved tool orchestration — these are meaningful advances for anyone building serious AI workflows.
But capability without cost control is not a business strategy. It is a race to see how fast your budget can disappear.
On-premises AI on open-weight models like Qwen3.6-27B sidesteps this entirely. You buy the hardware once. You run the model. Your inference cost is electricity. There are no tokens, no per-seat fees, no usage multipliers, no surprise rate hikes from a provider you did not choose.
This is not just a security argument or a compliance argument. It is an economic one. Stable, predictable AI costs are only possible when you own the hardware.
UK enterprises are already making this calculation, shifting workloads from cloud APIs to on-premises infrastructure through risk register decisions that account for both compliance and cost.
The acceleration is real. The capability gains are real. But so is the cost problem, and it only gets worse with every new release.
The organisations that own their own infrastructure lock in their costs for the life of the hardware. The ones that keep paying tokens are paying more for less predictability, every single week.
JD Fortress AI builds secure, on-premises RAG and agent solutions for UK businesses in regulated sectors. If you are exploring always-on, private AI teammates with predictable costs — no tokens, no usage surprises — get in touch for a confidential discussion.
If you're thinking about secure AI for your business, we'd love to have a conversation.
Get in Touch →