Claude Opus 4.7 and GPT-5.5 for UK SMEs: April 2026 deployment guide

Reviewed 14 October 2026. This is an April 2026 release comparison. For a current buying decision, use our Claude and ChatGPT plan comparison and UK cost guide.

April 2026 brought two frontier model releases inside a fortnight. Anthropic released Claude Opus 4.7 on 16 April 2026, with measurable gains on long-horizon agentic work, vision, and long-running coding tasks, plus task budgets in public beta as an advisory token target. OpenAI launched GPT-5.5 on 23 April 2026, alongside a new Codex-only seat type on ChatGPT Business and Enterprise that took effect on 2 April 2026, when the standard ChatGPT Business seat price also reduced by 5 US dollars per month. This was the April 2026 arrangement; the Codex seat structure has since changed, as described below. The commercial implication for UK SMEs is not that one provider has won. It is that the deployment decision is now better expressed function-by-function and across both providers, rather than as a single vendor choice. This guide sets out what changed, where each model fits, what the total cost looks like in GBP at typical SME seat counts, and the UK compliance footnote that still has to be respected regardless of which model is chosen.

What is Claude Opus 4.7?

Claude Opus 4.7 was Anthropic's frontier model at its April 2026 release, generally available from 16 April 2026, designed for long-horizon agentic work, complex coding, and document reasoning. It is the successor to Claude Opus 4.6 and is available through claude.ai, the Claude API, AWS Bedrock (in US East N. Virginia, Asia Pacific Tokyo, Europe Ireland, and Europe Stockholm at launch), and Google Cloud Agent Platform with EU multi-region routing. API pricing is unchanged from Opus 4.6 at 5 US dollars per million input tokens and 25 US dollars per million output tokens, with prompt caching reducing repeated input to 0.50 US dollars per million and batch processing cutting standard rates by 50 per cent.

The change against Opus 4.6 is concentrated in three areas. First, long-horizon agentic performance: Anthropic's release notes describe the model as working coherently for hours on multi-step tasks rather than degrading after a few minutes, which matters for any agent loop a business actually depends on. Second, vision: Opus 4.7 accepts input images up to 2,576 pixels on the long edge, more than three times the previous limit, which improves comprehension of complex documents, charts, and multi-page PDFs. Third, the General Finance benchmark that Anthropic publishes alongside the release moved from 0.767 on Opus 4.6 to 0.813 on Opus 4.7. That is a meaningful gain on a finance-heavy task set, although as with any single benchmark a UK SME should still pilot it on its own actual workload before treating it as decisive.

Anthropic task budgets, checked on 14 October 2026, provide an advisory token target across an agentic task. They are a soft signal to the model, not a hard cap. Enforce limits on steps, tokens, time and spend in application code, with alerts and a stop condition.

One caveat is worth pricing in. Anthropic shipped a new tokeniser with Opus 4.7. Anthropic states that the same input can use approximately 1.0 to 1.35 times as many tokens, depending on content type, so the headline price is unchanged but the effective cost per request can rise. Treat token count as an input to the estimate, not a list-price ceiling. Run a sample workload before committing to volume.

What is GPT-5.5 and the new Codex seat?

As checked on 14 October 2026, OpenAI’s GPT-5.5 API model card lists text and image input, text output, a 1,050,000-token context window and a 128,000-token maximum output. It does not list native audio or video input. Standard prices are USD 5 per million input tokens, USD 0.50 for cached input and USD 30 for output. Prompts above 272,000 input tokens attract higher rates and regional processing has a 10% uplift. These API specifications should not be treated as the limits or modalities of every ChatGPT feature.

GPT-5.5's headline benchmark claim is Terminal-Bench 2.0 at 82.7 per cent. Terminal-Bench evaluates real command-line workflows, including planning, iteration, and tool coordination in a sandboxed terminal, and it is the closest published proxy for how an agent performs against a developer's daily working environment. Anthropic reports Claude Opus 4.7 at 69.4 per cent on the same benchmark, so GPT-5.5 currently leads that specific test by more than thirteen points. The practical implication is narrower than the gap suggests, because Terminal-Bench measures a particular slice of agentic coding rather than enterprise document or finance work, but it is a real signal for terminal-heavy and codebase-heavy workflows.

The April launch introduced a usage-based Codex seat option. Current Business workspaces use Standard or Premium seats that include Codex; older workspaces may retain a legacy Codex seat arrangement. Check the applicable workspace billing documentation before modelling engineering usage.

For procurement, distinguish access fees from metered usage and included allowances. Record the workspace’s current seat types, usage limits and billing controls instead of applying an April launch arrangement to every workspace.

Function-by-function fit

Single-vendor framing breaks at first contact with how UK SMEs actually use AI. Different functions have different ratios of long-horizon reasoning, vision, code, text and image input, and cost sensitivity. The table below maps the six functions where The AI Consultancy's clients deploy AI most often, against which model is the better fit at the time of writing and why.

Function	Better fit	Why	UK SME deployment note
Finance month-end	Claude Opus 4.7	Higher General Finance benchmark (0.813 against 0.767 on Opus 4.6); stronger long-document reasoning across reconciliations and multi-period accounts.	Most useful via the API or claude.ai with explicit non-training settings on a Team or Enterprise tier. Use application-enforced limits on steps, tokens and spend.
Contract and policy review	Either, with a tilt to Claude	Both handle long contracts well. Claude's vision improvement helps with scanned PDFs and printed contracts. For recorded calls or voice notes, assess the separate transcription or audio service and its data flow.	Most UK firms already pilot legal review on a single platform; switching providers per matter creates more friction than it removes.
Sales follow-up and email drafting	GPT-5.5	Text and image input, with connected apps subject to workspace access and controls.	Standard ChatGPT Business seat at 20 US dollars per user per month annual is the natural fit.
Internal data Q&A (RAG)	Either	Both serve the model layer of a RAG architecture well. Choice is usually decided by what the rest of the stack already integrates with rather than the model itself.	Pilot with the provider whose enterprise admin and DPA are already in place. Switching is cheap if the retrieval layer is owned.
Operations playbooks and runbooks	Claude Opus 4.7	Long-horizon coherence over multi-hour playbook execution; application-enforced limits are needed to cap agentic spend.	Useful for finance ops, compliance ops, and back-office automation. Always require human approval at irreversible steps.
Knowledge work (drafting, summarisation, research)	Either, with a tilt to GPT-5.5 for breadth and Claude for depth	Test the required document formats, source citations and integration permissions on the actual workflow.	Most UK SMEs pick one for the firm-wide standard and let specialists request the other. Price difference at this scale is noise.

The honest reading of the table is that Claude is the stronger choice for finance, ops, and long-document work, GPT-5.5 is the stronger choice for sales, multimedia workflows requiring separate components, and broad knowledge-work surface area, and the rest are draws decided by existing stack integrations. For a longer-form companion to this analysis, see our Claude vs ChatGPT Enterprise for UK SMEs comparison, which sets out the current plan comparison and the total-cost-of-ownership method.

Claude cost examples at 10 and 50 seats; quotes for larger teams

The subscription prices below were checked on 14 October 2026. Anthropic’s UK pricing page shows pounds sterling including 20% VAT. API prices are in US dollars. Confirm each quote’s tax basis and your own VAT recovery position before comparing net costs; do not relabel a dollar price as pounds.

Claude Team Standard is £18 per seat per month billed annually, or £22.50 on monthly billing, including 20% VAT. Premium is £90 annually billed per month, or £113 monthly, including 20% VAT. Claude Enterprise lists USD 20 per seat per month billed annually plus model usage, with seat minimums. Obtain a quotation for ChatGPT Enterprise and confirm current ChatGPT Business pricing and tax treatment directly with OpenAI.

At 10 Claude Team Standard seats on annual billing, the annual subscription total is £2,160 including 20% VAT. At 50 seats it is £10,800 on the same basis. These totals exclude implementation and any separately billed API usage. Compare the same billing period, usage allowance and tax treatment in an OpenAI quotation.

Seat fees and API consumption are separate costs. Claude Enterprise adds model usage to its seat fee. Use the current UK cost guide for plan prices, and model API demand separately. Advisory task budgets do not prevent an agent from exceeding a financial limit; the application must enforce one. Obtain a quotation for ChatGPT Enterprise.

At 200 seats, compare a written ChatGPT Enterprise quote with Claude Enterprise seat and usage costs on the same tax and billing basis. Include integrations, administration and training. An assumed Enterprise price band is not a substitute for a quote.

Decision checklist

Five questions resolve most procurement decisions in 2026. The headlines are the same as twelve months ago; the answers have moved.

Pick Claude as the firm-wide standard when the dominant use case is long-document reasoning, finance work, ops playbooks, or any agent loop with a multi-hour execution window. Pick ChatGPT as the firm-wide standard when the dominant use case is sales, marketing, text and image workflows, or broad knowledge work across many casual users. Run both when the firm has functionally distinct teams (for example, finance ops on Claude and sales on ChatGPT) and the admin overhead is not a binding constraint. Use the API directly via AWS Bedrock (Europe Ireland or Stockholm for Opus 4.7 at launch) or Google Cloud Agent Platform EU multi-region when the workload is automation-heavy rather than chat-heavy, when seats would be a wasteful price model for the actual usage, or when a stricter EU residency boundary is required than the consumer surfaces provide.

The most common procurement mistake in 2026 is treating the seat price as the comparison axis. At 10 to 50 seats, the seat price is within procurement noise. At 200 seats it dominates the seat line but is still not the binding constraint, because usage costs and admin overhead can swing the total by more than the seat differential. Make the comparison on function fit, compliance fit, and total cost across seats and usage, not on which seat is a few pounds cheaper.

UK compliance footnote

The compliance baseline does not move because the models did. Three rules still apply, regardless of provider.

Under UK GDPR, any processing of personal data by an AI vendor on the firm's behalf requires a Data Processing Addendum compliant with Article 28, executed before the data is sent. Both Anthropic and OpenAI offer DPAs on business and enterprise tiers. The consumer tiers do not, which is why the consumer-tier ban is non-negotiable for any business workflow involving personal data, even for individual experimentation by employees.

As checked on 14 October 2026, AWS lists Claude Opus 4.7 in-region in Ireland (eu-west-1) and Stockholm (eu-north-1) on bedrock-mantle, not in London. Google Cloud’s Claude Opus 4.7 EU multi-region endpoint is an EU route, not a UK endpoint. Anthropic’s first-party API offers US or global inference. ChatGPT Enterprise UK residency is storage at rest; its documented inference regions are Europe, the US and UAE. Confirm the chosen service, endpoint, storage and contractual terms against our dated residency matrix. See the Claude residency matrix.

Two adjacent rules round it out. ICO guidance on AI and data protection still applies, including the requirement to complete a Data Protection Impact Assessment for higher-risk processing. The EU AI Act's Article 50 transparency obligations bind UK firms whose AI output is used in the EU from 2 August 2026, regardless of which model they use behind the scenes; its high-risk obligations were deferred by Regulation (EU) 2026/1744 to 2 December 2027 and 2 August 2028.