Claude Opus 5 for Business: Costs, Capabilities and UK Use Cases (2026)

Claude Opus 5 is Anthropic’s new flagship model, released on 24 July 2026, and it costs exactly what Opus 4.8 cost: $5 per million input tokens and $25 per million output tokens. The headline isn’t the sticker price, which didn’t move. It’s that early enterprise users report the same or better results using roughly a quarter fewer tokens, a third fewer steps and 60% less time. Your price per token is flat; your price per finished job falls. This is the business guide: what it costs in practice, what it’s genuinely good at, and five UK use cases worth building this quarter.

By Toni Dos Santos, Co-Founder, Spicy Advisory — we help mid-market and enterprise teams actually use the AI tools they’ve bought, tool-agnostically, across the UK and EU.

What is Claude Opus 5?

Claude Opus 5 is Anthropic’s flagship model for complex agentic coding and enterprise work, released on 24 July 2026 as the successor to Claude Opus 4.8. Anthropic’s own documentation now names it as the model to start with unless you specifically need the highest available capability, in which case you go up to Claude Fable 5 — at twice the price.

The Claude line-up as of late July 2026, and the one-line job of each:

The specs worth writing down

That 1M-token context window deserves a moment. It holds roughly 555,000 words — an entire ITT pack, a full year of board minutes, a mid-sized codebase — and Anthropic prices a 900,000-token request at exactly the same per-token rate as a 9,000-token one. There is no long-context tax. For a lot of UK businesses, that single fact removes the main reason they were chunking documents and losing context in the process.

What Claude Opus 5 actually costs

Most companies meet Opus 5 through one of two doors, and they have completely different cost shapes. Getting this wrong is the most common budgeting mistake we see.

Door 1: seats (the Claude apps)

If your team uses Claude in the browser or desktop app, you’re buying seats, not tokens. As published at the time of writing, Claude Team seats sit around $25–30 per user per month (roughly $20 on annual billing), with premium seats that include Claude Code at a considerably higher rate; Claude Enterprise is quoted per seat with usage billed at API rates on top. Check claude.com/pricing for today’s numbers — they move.

The important structural point: on a seat, the model upgrade is free. Opus 5 arrives inside the plan you already pay for. Your cost doesn’t change; only what your people can do with it does. Which means the entire return on this launch, for seat-based teams, is an adoption question rather than a procurement one.

Door 2: tokens (the API, Bedrock, Google Cloud, Foundry)

If you’re building — agents, document pipelines, internal tools — you pay per token, and there are four levers that change the bill by an order of magnitude:

• Prompt caching (up to 10× cheaper on the repeated part). Cache reads cost $0.50 per million against $5 standard. Cache writes cost 1.25× base for a 5-minute cache, 2× for an hour. On Opus 5 the minimum cacheable prefix drops to 512 tokens, half what Opus 4.8 required — so prompts you’d previously written off as too short to cache now qualify with no code change.
• The Batch API (50% off). Anything that doesn’t need an answer this second — overnight document runs, monthly reporting, bulk classification — halves in price. Batch and caching discounts stack.
• Effort. The effort parameter (low through max) controls how hard the model works per turn. It defaults to high. Anthropic’s own migration guidance is unusually direct here: low and medium are unusually strong on this model, so effort settings inherited from a previous model are probably wrong and worth re-testing.
• The multipliers people forget. Regional or multi-region endpoints on Bedrock and Google Cloud carry a 10% premium over global. Pinning inference to the US via inference_geo applies a 1.1× multiplier across every token category. Web search inside a request is $10 per 1,000 searches on top of tokens.

The number that actually matters: cost per finished job

Here’s the argument to take to your CFO. Opus 5 costs the same per token as Opus 4.8. But three independent organisations reported the same pattern at launch — same or better output, materially less work to get there:

• Harvey (legal AI) reported that Opus 5 achieved similar performance to Opus 4.8’s maximum-reasoning mode “while generating 26% fewer tokens on average.”
• Fundamental Research Lab reported that on hard financial-modelling tasks the model averaged nine percentage points higher accuracy “while using roughly one-third fewer turns and tool calls and 60% less time.”
• Cognition (the team behind the Devin coding agent) reported that on FrontierCode 1.1, “Claude Opus 5 approaches Fable-level performance at half the cost,” with particular strength in debugging and root-cause analysis.

The benchmark story points the same way. On Frontier-Bench v0.1, an agentic terminal-coding evaluation, Opus 5 scores 43.3% — more than double Opus 4.8’s 18.7%, and comfortably ahead of the more expensive Fable 5 at 33.7%. On ARC-AGI 3, which tests novel problem-solving, Anthropic reports Opus 5 scoring roughly three times the next best model.

“Stop comparing models on price per million tokens. Compare them on cost per completed job. A model that costs the same per token but needs a third fewer turns to finish the work has just cut your unit economics by a third — and no line on the pricing page will tell you that.”

— Toni Dos Santos, Co-Founder, Spicy Advisory

Worked example: a contract review agent at a 200-person UK firm

Assume a professional services firm reviewing 500 contracts a month. Each run sends about 30,000 tokens of stable context (review playbook, clause precedents, worked examples) plus 10,000 tokens of the contract itself, and produces about 6,000 tokens of output. All figures in USD, converted at $1.30 = £1 for illustration — the API bills in dollars, so your sterling cost moves with the rate.

Forty-one pounds a month to review 500 contracts with a frontier model. That is roughly eight pence per contract. Whatever is expensive about this project, it isn’t the model — a point worth making loudly in any business case, because AI budgets are still routinely built as though inference were the dominant cost. It isn’t. People, process and governance are. We put numbers on that in measuring AI training ROI in a UK business.

Worked example: a 60-person Manchester agency on seats

Twenty-five people who’d realistically use Claude weekly, on standard Team seats at roughly $25/user/month, comes to about £480 a month, or £5,800 a year. Now set that against one hour saved per person per week: 25 people × 45 working weeks = 1,125 hours, at a blended internal cost of £45/hour, is roughly £50,000.

The licence pays for itself if each person saves about seven minutes a week. That is not a hard bar. Which is exactly why the licence is never the problem — the problem is that across large organisations roughly 90% have invested in AI tools while only about 20% of employees use them weekly. A new model does nothing to that gap. Redesigned workflows and a measured first win do. We explain the mechanism in why AI adoption fails in companies, and the UK-specific numbers are in our UK SME AI adoption statistics.

What Opus 5 is genuinely good at

Cutting through the launch language, here is what changed in ways a business team will feel.

• Long, hard, multi-step work. This is the headline capability. Multi-file features, larger refactors, end-to-end deliverables — it finishes things rather than leaving stubs and placeholders. The gap over previous models is smaller on quick single-step edits and largest on the genuinely difficult end of your workload, so evaluate it there.
• Finding real problems. High precision and high recall on code review and bug-finding, and it stays accurate at lower effort settings — which makes a cheap fast pass at commit time plus a thorough pass later a practical pattern rather than a luxury.
• Office documents that survive contact with a client. It generates and edits complex multi-sheet Excel files with non-trivial formulas, and PowerPoint decks that follow actual slide-design conventions. It can be held to your template when you give it one.
• Charts, scans and screenshots. Stronger on chart, document and diagram understanding, and on replicating a UI from an image. The highest-leverage change here is giving it tools to crop and re-examine its own work — on this model, tool use beats simply telling it to think harder.
• Coordinating other agents. It runs teams of sub-agents without them overwriting each other’s work, and uses writer-verifier patterns effectively.
• A fresher view of the world. Knowledge reliable to May 2026, against January 2026 for Fable 5. For anything touching recent regulation, market conditions or tooling, that matters more than a benchmark point.

Five UK use cases worth building this quarter

These are patterns we’d actually scope for a UK mid-market client this quarter, matched to where Opus 5’s specific strengths land. Treat the numbers as worked illustrations with our assumptions — swap in your own.

1. Clause review for law firms and in-house counsel

The Harvey evidence lands directly here: same quality as maximum-reasoning Opus 4.8, materially fewer tokens. A 40-partner regional firm can run first-pass review against its own playbook, flagging deviations rather than drafting, with a named solicitor signing off every output. The 1M context means the precedent library and the contract sit in the same window. Start with our guide to AI training for UK legal teams.

2. Financial modelling and analysis in accountancy and financial services

Nine points of accuracy on hard modelling tasks, with a third fewer turns, is the single most quotable result of this launch — and multi-sheet Excel generation with working formulas is the capability that makes it usable. For FCA-regulated firms, pair it with the governance expectations we cover in the FCA Mills Review and AI training for UK financial services.

3. Legacy migration and code review for engineering teams

Cognition’s “Fable-level performance at half the cost” on FrontierCode 1.1, plus the debugging and root-cause strength, points at the work most UK engineering teams have been deferring: the COBOL-adjacent migration, the untested legacy service, the dependency upgrade nobody wants. Where this fits against Copilot and Claude Code is covered in when to use Claude, Copilot or code.

4. Bid, tender and proposal work

A 500-page ITT pack, the framework agreement, three past submissions and your capability statements fit in one context window with room left over. For construction, consultancy and public-sector suppliers, this removes the chunking step that used to lose the cross-references that actually decide a bid. More on the sector pattern in how UK professional services firms are using AI to win more business.

5. Back-office document processing at batch prices

Invoice coding, supplier onboarding checks, expense review, records reconciliation. Overnight, on the Batch API, with vision handling the scanned material — this is where the 50% batch discount plus caching makes frontier-model quality cost less than the offshore alternative. See AI workflows for finance teams.

What gets trickier

An honest rollout names the costs as well as the wins. Five things to plan for:

• It talks more. Opus 5 writes longer user-facing responses and longer files than its predecessor, and narrates more during agentic runs. Turning down effort does not reliably fix this — a short conciseness instruction in the system prompt does, cutting response length by around 20% in Anthropic’s own testing.
• It delegates eagerly. This is a reversal: Opus 4.8 under-used sub-agents and needed prompting to delegate; Opus 5 reaches for them freely. Every sub-agent re-establishes context and reports back, which multiplies both cost and latency. If you added “delegate more” guidance for 4.8, take it out and add a cap.
• It checks its own work — so stop telling it to. Instructions like “double-check your answer” and separate verification steps in your harness now cause over-verification rather than accuracy. Deleting them reduces waste with no loss of quality. This inverts a prompting habit most teams have been taught, so flag it explicitly in training.
• It can widen the brief. It occasionally adds steps nobody asked for. A short scope-discipline line in the system prompt largely eliminates this.
• Rate limits are a separate pool. Opus 5 does not draw on the shared Opus 4.x limit. Moving traffic across neither frees headroom on the old bucket nor inherits it — check your tier’s Opus 5 limits before you shift volume. Our guide to protecting Claude usage limits covers the day-to-day tactics, and AI spend visibility covers how to see the bill before it surprises you.

None of these are blockers. All of them are prompt-and-policy work — roughly an afternoon for someone who knows what they’re doing, and roughly a quarter of unexplained overspend for someone who doesn’t.

Opus 5, Sonnet 5, Fable 5 or Haiku: how to choose

The practical rule we give teams: default to Sonnet 5, escalate to Opus 5 when the task is genuinely hard, and only reach for Fable 5 when you’ve measured that Opus 5 isn’t enough. Most organisations reach for the biggest model reflexively and pay for it. For the wider cross-vendor question, see Claude vs ChatGPT for business and our benchmark of generalist AI assistants for work.

A 30-day rollout that won’t embarrass you

1. Week 1 — audit your model strings. Find every place a model ID is hard-coded. Anything still pointing at claude-opus-4-1 is on borrowed time: it retires 5 August 2026. While you’re there, note every call that omits the thinking parameter, because behaviour changes on Opus 5.
2. Week 1 — pick one hard workload. Not a demo. The migration nobody wants, the review queue that’s always behind. Opus 5’s advantage is largest on difficult work and smallest on easy work, so an easy pilot will understate it.
3. Week 2 — sweep effort. Run the same workload at medium, high and xhigh against your own evaluation set. Do not inherit the setting from your previous model. Record cost, latency and quality for each.
4. Week 2 — turn on caching and batching. These two changes are usually worth more than any prompt engineering you’ll do all quarter.
5. Week 3 — tune the three known habits. Add a conciseness instruction, a sub-agent cap and a scope-discipline line. Delete your old “verify your work” scaffolding.
6. Week 4 — measure and decide. Cost per completed job, before and after. If it hasn’t moved, the model was never your constraint — and that’s a genuinely useful finding.

Governance: what UK companies need to have in place

Opus 5 ships across Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry on day one, which means for most UK enterprises adopting it is not a new vendor relationship — it’s a new model inside a cloud estate you already govern. The questions that gate adoption are cloud-governance questions you can mostly answer with the frameworks you have: data residency, access control, retention, DLP.

Three specifics worth flagging. First, on commercial plans and the API, Anthropic does not train on your business inputs and outputs by default — but you still need a documented policy on what staff may enter. Second, if you need inference pinned to a geography, both the inference_geo parameter and regional cloud endpoints exist, and both carry a ~10% premium; budget for it rather than discovering it. Third, Opus 5 ships with elevated cybersecurity safeguards, which means legitimate security and life-sciences work can occasionally trip a refusal — worth knowing before someone concludes the model is broken.

The detail is in our guides to AI data residency for UK enterprises and AI governance and what the ICO expects. If you operate in the EU as well, the Article 4 AI literacy obligation already applies to you.

Where Spicy Advisory fits

We help teams get from “we have the newest model” to “it changed how we work”: the default profiles, the effort-and-spend policy, the workflow redesign on real deliverables, and the enablement that closes the gap between 90% of companies investing and 20% of employees using. We’re tool-agnostic and work in English or French, on-site and hybrid across the UK and EU — see AI training in the UK, AI consulting for UK SMEs and mid-market, Claude training and Spicy Advisory for Enterprise. New to Claude? Start with the getting-started guide for teams and the broader Claude for companies playbook. For the previous two launches in this line, see our guides to Claude Opus 4.8 and Claude Fable 5.