Which AI model should a UK business use? The autumn 2026 decision table

Last checked: 7 Oct 2026. Covering model releases up to 7 Oct 2026. Every figure below comes from the vendor's own pricing or documentation page on that date, not from a comparison site.

The question we get asked most often is a version of this: "Do you use ChatGPT, or Claude, or something else, and why would you pick one over the other for us?" It is a fair question and it is usually answered badly, because most published comparisons rank models on coding tests and then present the ranking as procurement advice.

SpotDev is an AI and digital transformation consultancy that builds real software. We are an OpenAI Select Partner and Certified in the Claude Partner Network, and the table below is the same table we use internally. The working method is a ladder: put the cheapest model that can do the job on the volume work, and reserve an expensive model for the step where being wrong is costly. Most builds end up using two or three models, not one.

Prices are converted from US dollars at £1 = $1.35 on 11 Sep 2026. Vendors bill in dollars and UK VAT applies on top. HubSpot is the exception: it publishes a sterling price list.

The decision table, autumn 2026

Two things the table cannot show. First, included credits: Starter accounts get 500 a month, Professional 3,000 and Enterprise 5,000, with no rollover. Second, the ChatGPT connector bypasses HubSpot validation rules on write, which is a governance decision, not a feature.

Eight criteria that actually decide it

1. Licences you already pay for

Before anything else, list what is already on the invoice. A Microsoft or Google estate, a HubSpot Professional or Enterprise subscription with monthly credits nobody is spending, an existing ChatGPT Business tenancy, or a Cursor subscription, where Grok 4.7 sits in the included usage pool as a first-party model while third-party models on Teams and Enterprise plans carry an extra per-token charge. The cheapest capable option is frequently one you are already funding, and saying so is often the honest answer to "which model should we buy".

2. Residency

This is the criterion most likely to remove options outright, and the position is messier than the marketing suggests. No first-party Claude inference runs in the EU or the UK, and on Amazon Bedrock the options differ by model: Claude Sonnet 5.5 launched without an EU geographic profile and gained one later, while Claude Haiku 5.5 launched with one. An EU profile routes across European regions rather than keeping processing in the UK, so check each model rather than each vendor. OpenAI gives the EU storage and processing but the UK storage only. Gemini 3.8 Flash covers EU multi-regions but not UK in-country. Grok lists residency at Enterprise tier only. If a regulator or an outsourcing policy requires in-country processing, most of this table fails that test today, and the buildable answers are a cloud provider's regional endpoints or keeping the data out of the model entirely.

3. Integration surface

What the model can reach matters more than what it scores. The gap between a read-only connector and one that writes back is the difference between a tool that answers questions about your pipeline and one that owns a stage of it. Right now the ChatGPT connector writes, the Claude connector writes, HubSpot's MCP server reads and writes, and the Gemini connector reads three object types. Grok and Muse have no published HubSpot route at all.

4. Agentic maturity

Judge this on documentation, not demos. Anthropic and OpenAI both publish the specifics: which tools exist, what computer use costs in extra tokens, how effort levels change behaviour. Where a vendor only markets the capability, treat the absence of documentation as an unknown rather than as a weakness, then test it before you commit.

5. Cost at scale, worked

Headline rates mislead until you run a real volume through them. Take 10,000 support conversations in a month, each averaging 8,000 input tokens of ticket history and knowledge-base context and 800 output tokens. That is 80M input and 8M output tokens.

• GPT-6 Luna: about £9 (GPT-5.6 Luna: about £19)
• Claude Haiku 5.5: about £9, the same as Luna, because each 8,000-token prompt sits under Haiku's 100,000-token threshold
• Gemini 3.8 Flash: about £67 at the introductory rate, about £133 from 1 Jan 2027
• Grok 4.3: about £89
• Muse Spark 1.3: about £99
• Grok 4.7, or Grok 4.6 at the same price: about £154
• Claude Sonnet 5.5 or GPT-6.1 Sol (GPT-6 Sol is the same list price): about £178
• Claude Opus 5.5: about £356
• GPT-6 Astra or Claude Fable 5.1: about £889
• HubSpot customer agent: about £4,500, or about £4,050 on annual credit pricing

Read that honestly. The token figures are raw inference only. They exclude the retrieval layer, the platform, the engineering to build it and the human review step. The HubSpot figure buys a finished, supported product where you pay only on a resolved conversation. The comparison is not 45p against a fraction of a penny, it is "buy the outcome" against "build the thing". A hundred-fold spread between the top and bottom rungs is the whole argument for a ladder, though. One client's per-record cost fell by roughly 90% when we had an overnight process pre-summarise account history instead of making the agent re-read everything live, with no model change at all.

List price is also not the bill. Claude Sonnet 5.5 and GPT-6.1 Sol cost the same per token, yet on Artificial Analysis's run of its Intelligence Index at maximum effort, checked 30 Sep 2026, GPT-6.1 Sol averaged about 53p per task and Sonnet 5.5 about £5.63, more than Claude Opus 5.5 at about £4.43, because Sonnet 5.5 generated far more tokens. That is one index at one setting, and Anthropic itself says Sonnet 5.5 complements Opus 5.5 best at lower effort. The practical point is that the effort setting now moves cost as much as the model choice does, so test at the setting you will actually run.

The bottom rung has the same trap. Claude Haiku 5.5 and GPT-6 Luna charge identical rates below 100,000 tokens, but on the same index, both at maximum effort, Haiku 5.5 scored 43 to Luna's 38 and generated about 440M output tokens to Luna's 140M, so its extra capability shows up in the bill per task. Prompt length decides the rest. A 150,000-token prompt costs five times as much on Haiku 5.5 as on Luna, because Haiku's higher rate starts at 100,000 tokens and Luna's surcharge only starts at 272,000. Haiku 5.5 suits the harder short tasks. Luna suits the cheapest possible volume and anything with a long context. We set out the full comparison in whether Claude Haiku 5.5 or GPT-6 Luna should run your volume work.

6. Vendor churn risk

Anthropic replaced three of its four current Claude models in just over two weeks: Opus 5.5 on 22 Sep 2026, Sonnet 5.5 on 28 Sep 2026 and Haiku 5.5 on 7 Oct 2026. Opus 5 and Sonnet 5 have moved to legacy. Haiku 4.5 is still available, with retirement listed as not sooner than 15 Oct 2026 and no deprecation notice yet. OpenAI has shipped three Sol models in under three months: GPT-5.6 Sol on 9 Jul 2026, GPT-6 Sol on 22 Sep 2026 and GPT-6.1 Sol a week later on 29 Sep 2026, with the older models still listed. Google shipped three Flash releases in six weeks, and its own model card for 3.8 Flash says the safety evaluation found no meaningful new capabilities against 3.7 Flash. The OpenAI Assistants API was sunset on 26 Aug 2026 and is no longer available. SpaceXAI replaced its flagship after 38 days: Grok 4.6 shipped on 14 Aug 2026 and Grok 4.7 on 21 Sep 2026, at the same price, with 4.6 still listed. This is not a reason to pick one vendor over another. It is a reason to keep the model name in configuration, behind an interface, so swapping it is an afternoon and not a project. If you already run Opus 5, what changes when you move to Opus 5.5 covers the breaking changes and the new default effort setting.

7. The shape of the pricing model

Three shapes are on offer. Per-token pricing suits variable and unpredictable volume. Per-seat pricing suits universal adoption where everyone uses it daily. Per-outcome pricing, which HubSpot is alone in offering here, suits low-volume high-value tasks and makes budgeting trivial. The shape usually matters more than the rate, because it decides whether your bill scales with usage, with headcount or with results.

8. Speed

Latency is a product decision, not a benchmark. A customer waiting on a live chat reply and a nightly batch job have opposite requirements. Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and ranks Fable 5.1 slower within its own lineup. Sonnet 5.5 defaults to Medium effort in the Claude apps and Claude Code and to High on the API. Anthropic calls Haiku 5.5 its fastest model to date at standard speed, and it defaults to Medium effort on the API. Google positions Flash as the speed tier. Both OpenAI and Anthropic charge more for faster or higher-effort modes. Decide the acceptable wait first, then shop within it.

What the leaderboards are actually measuring

Public benchmark leaderboards are coding and reasoning tests repackaged as business comparisons. They are real tests, honestly run, and they were never designed to answer "which of these should handle our quote chasing". Nothing on them measures whether a model can find the right record in your CRM, respect a permission boundary, or hand over to a person at the right moment.

The gap between vendor-reported and independent scores is the clearest illustration. OpenAI reported 99.9% for GPT-6 Astra on ARC-AGI-3. ARC Prize's own published result shows 62.7% under its standard harness and 99.9% only through OpenAI's provider adapter, which preserves opaque reasoning state between requests and compacts long conversations so the model can reuse prior work. Both numbers are true. They measure different things, and only one of them is the model.

Then look at the ranking itself. On the Artificial Analysis Intelligence Index, checked 30 Sep 2026 at the effort settings shown in its headline chart, Claude Opus 5.5 leads at 58, with Claude Sonnet 5.5 second at 56, ahead of Claude Fable 5.1 and GPT-6 Astra, which both score 53. GPT-6.1 Sol scores 52, Muse Spark 1.3 48, Grok 4.7 46 and Gemini 3.8 Flash 41. Claude Haiku 5.5, checked on its own model page on its launch day, scores 43 at maximum effort, ahead of GPT-6 Luna at 38. The vendor tables tell you what vendor tables always tell you. Anthropic's Sonnet 5.5 table shows it close to Opus 5.5 and well ahead of GPT-6 Sol on the knowledge-work tests it chose, while Anthropic says Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment". OpenAI's GPT-6.1 Sol post shows it beating Opus 5.5 on some professional-work tests at a fraction of the cost per task. Neither compares its new model with the other's release of the same week. Anthropic's Haiku 5.5 table does compare directly with GPT-6 Luna, and shows Haiku ahead on every test where Luna has a score. SpaceXAI's table shows Grok 4.7 ahead of Grok 4.6 on all seven benchmarks it chose to publish. On ARC-AGI-3, GPT-6 Sol now scores 4.6% under ARC Prize's standard harness and 23.0% through the provider adapter, GPT-6 Luna scores under 1% under both, and Opus 5.5, Sonnet 5.5, Haiku 5.5, GPT-6.1 Sol and Grok 4.7 have no result yet. A six-point spread across the top five is not a procurement shortlist. Use benchmarks to exclude models that are obviously unfit, then run your own evaluation on twenty real examples from your own data. That evaluation takes about a day and it beats every leaderboard for your specific decision.

How this page is maintained

This table is updated within three working days of each major model release. Prices, context windows, residency positions and connector capabilities are re-checked against the vendor's own page at each update, and the "last checked" date at the top moves only when that has actually been done.

• 7 Oct 2026: Claude Haiku 5.5 added on its launch day, at the same price as GPT-6 Luna below 100,000 tokens. Claude Sonnet 5.5 cache reads halved, and its Bedrock residency now includes an EU profile. GPT-6 Astra Ultrafast recorded. Worked costs, cost per task, vendor churn, speed and leaderboard figures re-checked.
• 30 Sep 2026: Claude Sonnet 5.5 replaces Claude Sonnet 5 and GPT-6.1 Sol replaces GPT-6 Sol. OpenAI EU residency now covers Flex and Batch. Bedrock launch position for Sonnet 5.5 recorded. Worked costs, cost per task, vendor churn, speed and leaderboard figures re-checked.
• 22 Sep 2026 (second update): Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna added on their launch day. GPT-6 Luna replaces GPT-5.6 Luna as the volume rung. Worked costs, vendor churn and leaderboard figures re-checked.
• 22 Sep 2026: Grok 4.7 replaces Grok 4.6 in the table at the same price. X Search billing recorded as per post from 21 Sep 2026. Leaderboard figures re-checked. Cursor first-party route noted.
• 12 Sep 2026: first published, covering releases to 8 Sep 2026.

How we would use this with you

Model choice is the last decision, not the first. In most engagements the binding constraint is data quality, a permissions model or an integration that does not exist yet, and the model is a line item. Our approach across AI implementation is to settle the workload, the residency position and the approval step first, then pick the cheapest rung that clears the bar. Where a single vendor is the right answer we go deep on it, through OpenAI implementation, Claude implementation, SpaceXAI implementation or HubSpot AI. For a two-vendor comparison in more depth, see Claude vs ChatGPT for business. Claude’s agent surface also changed on 16 Sep 2026, when Anthropic merged Cowork into the main product: see what the Claude Cowork merge means for your AI policy.

If you want a view on your own situation rather than a table, start with a diagnostic, or Request a Quote.

Frequently asked questions

Which AI model is cheapest for a UK business in autumn 2026?

On raw API pricing, GPT-6 Luna and Claude Haiku 5.5 tie as the cheapest options in this table, at about 7p per million input tokens and about 37p per million output tokens for prompts up to 100,000 tokens. Above that, Haiku 5.5 rises to about 37p and about £1.85, while Luna keeps its price up to 272,000 input tokens, so Luna is cheaper for long documents. Haiku 5.5 scores higher on benchmarks but tends to use more output tokens per task. Gemini 3.8 Flash is next at about 56p and about £2.78, though that is an introductory rate that doubles on 1 Jan 2027. Prices are converted from US dollars at £1 = $1.35 on 11 Sep 2026 and UK VAT applies on top.

Can I keep AI processing of our data inside the UK?

Not with most first-party model APIs today. OpenAI's UK region offers storage only, with processing in the US or the EU. Google lists Gemini 3.8 Flash for EU multi-regions but not for UK in-country regions. Anthropic offers no first-party EU or UK region, only US-only or global routing, with EU hosting available through Amazon Bedrock or Google Vertex AI for most models, though Bedrock's EU profiles route across European regions rather than keeping processing in the UK, and Claude Sonnet 5.5 launched without one. If in-country processing is a hard requirement, design for it explicitly rather than assuming a vendor tier covers it.

Which AI assistants can write back to HubSpot?

As of 30 Sep 2026, the ChatGPT connector and the HubSpot connector for Claude can both create and update CRM records, and HubSpot's remote MCP server gives any MCP-compatible client read and write access. The HubSpot connector for Gemini is read-only and covers contacts, companies and deals. There is no published HubSpot connector for Grok or Muse.

Should we standardise on one AI vendor?

Fewer vendors means fewer data-residency questions to answer and fewer contracts to govern, which is a real benefit. It is also usually more expensive, because you end up running volume work on a flagship model. A reasonable middle position is one approved vendor for anything touching customer data, with a cheaper second model allowed for internal, low-risk, high-volume steps.

How much does HubSpot's Agent Hub cost compared with building an agent?

HubSpot charges per completed outcome in credits, at £9.00 per 1,000 credits monthly or £8.10 annually. A resolved customer-agent conversation costs 50 credits, about 45p. A custom build's inference cost for the same conversation is a fraction of a penny, but that excludes the platform, the retrieval layer and the engineering. Agent Hub is usually better value at low volume or where the work sits entirely inside HubSpot. A build wins at high volume or where the process crosses systems.

Do benchmark scores tell us which model to buy?

No. Public leaderboards are coding and reasoning tests, and vendor-reported scores can differ sharply from independent ones. OpenAI reported 99.9% for GPT-6 Astra on ARC-AGI-3, while ARC Prize's own standard harness gives 62.7%. Use benchmarks to rule out obviously unfit models, then evaluate the shortlist on twenty real examples from your own data.