Which Claude model should a UK business use in 2026?

Anthropic replaced its whole mainstream line-up in just over five weeks. Fable 5.1 arrived on 1 September 2026. Opus 5.5 followed on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. For a UK business trying to make one decision and then get on with the work, that pace is the problem: any model comparison written more than a few weeks ago describes a line-up that no longer exists, and the instinct to standardise on the most capable model available is the most expensive instinct in the room.

This guide takes the opposite approach. It starts from the workload, not the model, and it treats the price list as one input among several. Every figure was verified against Anthropic's own documentation on 10 October 2026, and the sterling conversions state their basis, because a Claude cost in pounds is only as good as the exchange assumption behind it. An earlier version of this guide, published in August, covered Opus 5, Sonnet 5 and Haiku 4.5; those models are still available and are now listed as legacy.

The current line-up, and what each model is actually for

Five models matter to a UK business today. Four are generally available and one is not.

Model	USD per million tokens (in / out)	Context / max output	Best fit
Claude Haiku 5.5	$0.10 / $0.50 for prompts up to 100,000 tokens; $0.50 / $2.50 above	1M / 128k	High volume, low ambiguity: classification, extraction, routing, first-pass triage, subagent work
Claude Sonnet 5.5	$2 / $10	1M / 128k	The default for production work: drafting, summarising, document review, most agentic tasks
Claude Opus 5.5	$4 / $20	1M / 128k	Tasks where a wrong answer is expensive: complex reasoning, difficult code, long agentic runs
Claude Fable 5.1	$10 / $50	1M / 128k	The frontier tier, for demanding reasoning and long-horizon agentic work
Claude Mythos 5.1	$10 / $50	Shares Fable 5.1's specifications	Not generally available: see the note below

Prices correct as at 10 October 2026. Treat any price table in any article, including this one, as a dated snapshot rather than a standing fact, and check the current figures before committing a budget.

Three things in that table are worth pausing on. First, every tier got cheaper or held its price. Opus 5.5 is $4 and $20 against $5 and $25 for Opus 5. Sonnet 5.5 holds the $2 and $10 that Sonnet 5 carried. Haiku 5.5 on short prompts is a tenth of Haiku 4.5's $1 and $5. Anthropic’s running-cost estimates use different bases: Opus 5.5 combines lower prices and fewer tokens for an estimated 40 per cent saving against Opus 5; Sonnet 5.5 uses fewer tokens for up to 30 per cent less per task at the same token price; Haiku 5.5’s estimated 75 per cent saving comes from lower prices, partly offset by slightly more tokens per task. Those percentages are the vendor's estimates from its own testing. The list prices are the part you can budget on.

Second, the context window is no longer a differentiator at all. All four generally available models carry a 1M token window and a 128k maximum output, including Haiku, which was limited to 200k in the previous generation. The subscription plans are less clear, because Anthropic's two surfaces disagree: the plan comparison on its pricing page gives the context window as up to 1M on every plan, varying by model, while its Help Center article for the Team plan still states a 200k context window. Check the window in the product for the model and plan you are using before relying on either figure.

Third, Haiku 5.5 is priced by prompt length, and it is the only model that is. A request whose prompt is over 100,000 tokens pays five times the lower rate, and the prompt length counts cached tokens as well as fresh ones. Anthropic says about 90 per cent of requests to the previous Haiku fell under that line. If your Haiku workload feeds it long documents, you are in the other 10 per cent.

A word on Anthropic's own advice, because it differs from ours. Its models overview suggests starting with Opus 5.5 for most workloads and moving to Fable 5.1 where evaluations still fall short. That is reasonable guidance for a team optimising for capability. For a UK SME optimising for cost per useful output, we would start one tier lower, on Sonnet 5.5, and escalate the cases that fail.

On Claude Mythos 5.1, the accurate position is narrow. Anthropic lists it in limited availability to approved partners, and a UK buyer should plan on the basis that it is not part of the available line-up. Our briefing on Claude Fable 5.1 and Enterprise Frontier Safeguards sets out what Anthropic has published.

Retention is now a selection criterion

The Covered Model distinction makes retention a selection criterion as well as capability and price. Check the current model-specific terms and any approved temporary arrangement. Claude Fable 5.1 is what Anthropic calls a Covered Model: prompts and completions are retained for at least 30 days by default on every surface, and zero data retention is not available where Covered Models can be accessed, other than for eligible customers under a time-limited arrangement.

Opus 5.5 is not on that list. As we read the Covered Models page on 10 October 2026 it contains Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1 only, and Anthropic's launch posts state that Opus 5.5 and Sonnet 5.5 are available with zero data retention. Anthropic also says Opus 5.5 performs at the level of Fable 5.1 on most work, while noting that benchmark margins at this level are a less reliable guide to real differences than they used to be.

For a UK firm in regulated or client-confidential work, that changes the default answer at the top of the range. If a 30-day retention window is a problem for your DPIA, Opus 5.5 gives you capability Anthropic describes as close to the frontier inside the retention position you already hold. Two cautions: zero data retention is an agreement you hold with Anthropic or your cloud provider rather than a default, and Anthropic says it will update the Covered Models list as designations change. Record the model, the position and the date you checked.

The same prices in sterling

Anthropic publishes API prices in US dollars and bills in dollars, so a sterling figure has to be derived. These conversions use HMRC's published monthly exchange rate for October 2026, USD 1.3456 to GBP 1, published on 16 September 2026 and valid for the calendar month. We use HMRC's rate rather than a spot rate because it is official, dated and fixed for a known period, which is what a budget needs.

Model	GBP per million tokens (in / out)	Same workload through the Batch API
Claude Haiku 5.5, prompts up to 100,000 tokens	£0.07 / £0.37	£0.04 / £0.19
Claude Haiku 5.5, prompts over 100,000 tokens	£0.37 / £1.86	£0.19 / £0.93
Claude Sonnet 5.5	£1.49 / £7.43	£0.74 / £3.72
Claude Opus 5.5	£2.97 / £14.86	£1.49 / £7.43
Claude Fable 5.1	£7.43 / £37.16	£3.72 / £18.58

These figures exclude tax and are indicative rather than invoiced amounts: API billing is in dollars, so what lands on a card statement also depends on the issuer's own rate and any non-sterling fee. Our sterling worked example for Claude pricing sets out the full method, including the subscription side, where Anthropic does publish prices in pounds to UK visitors.

Choosing by workload: five patterns and the model each one wants

The useful unit of decision is a workload, not an organisation. Most UK deployments contain three or four distinct workloads with different requirements, and the cost difference between matching them properly and standardising on one model is large.

High volume, low ambiguity. Classifying inbound email, extracting fields from invoices, routing enquiries, tagging documents. Haiku 5.5 at $0.10 and $0.50 is the right tool, provided the prompts stay under 100,000 tokens. At that rate the input cost is a twentieth of Sonnet 5.5's, which is a large enough gap to justify testing Haiku first on anything mechanical. Keep long documents out of this tier or split them, because the higher rate applies to the whole request.

Drafting and summarising with a human reviewer. Client correspondence, meeting notes, first drafts of reports. Sonnet 5.5 is the default. A human is reading the output before it goes anywhere, so the marginal value of a more capable model is largely absorbed by the review step you were going to do anyway.

Document review across a large corpus. Here the 1M token window earns its place, and on Sonnet 5.5 and Opus 5.5 it carries no pricing premium. The question is only whether the reasoning is hard enough to justify Opus. Start on Sonnet 5.5, measure, and escalate the cases that fail rather than the whole workload.

Work where a wrong answer is expensive. Complex analysis, difficult refactors, anything feeding a decision that is hard to reverse. Opus 5.5 at $4 and $20 is the honest choice, and the sensible framing for a finance conversation is not the token price but the cost of the error being avoided. Reach for Fable 5.1 only where your own evaluation shows Opus 5.5 falling short, and only after settling the retention question above.

Long agentic runs with many tool calls. Model choice matters here for a reason that is not on the price list: tool-use overhead. Anthropic publishes a per-model system prompt cost for tool use, and the spread is wide. For a tool choice of auto or none it is 286 tokens on the Claude 5.5 models, against 354 on Sonnet 5, 496 on Haiku 4.5 and 675 on Opus 4.7. A tool-heavy workload therefore gets cheaper by moving to a newer model even before the lower token price is counted. Cache reads matter most here too, because an agent re-reads its context on every step, and they are cheapest in relative terms on Sonnet 5.5, Opus 5.5 and Fable 5.1.

The levers that move a bill more than the model row does

In our experience the model choice is usually the third or fourth largest determinant of a Claude bill. These are the levers that tend to matter more.

Batch processing. The Batch API applies a 50 per cent discount to both input and output tokens in exchange for asynchronous turnaround. Anything overnight, anything scheduled, anything where nobody is waiting: this is the single largest available saving and it requires no change to the prompt.

Prompt caching. Writes cost 1.25x the base input rate for the five-minute cache or 2x for the one-hour cache. Reads cost 0.1x as standard, and less on the newer models: 0.05x on Opus 5.5 and Sonnet 5.5, which is $0.20 and $0.10 per million tokens, and 0.025x on Fable 5.1, which is $0.25. Anthropic states the standard break-even explicitly: caching pays off after one cache read at the five-minute duration, or two reads at the one-hour duration. For any workload with a stable system prompt or a repeated document, this is the cheapest saving to implement. The two discounts stack: batch and caching can be combined.

The effort parameter. Effort takes five values (max, xhigh, high, medium, low) and is set as output_config.effort. It affects every output token: the response text, the tool calls and the thinking. Three points from Anthropic's documentation matter in practice. The default is not the same everywhere: it is high on most models and medium on Opus 5.5 and Haiku 5.5, so a request that omits effort runs one level lower on Opus 5.5 than it did on Opus 5. The levels have been recalibrated on Sonnet 5.5, so a setting carried over from Sonnet 5 does not buy the same amount of thinking. And changing the top-level effort between requests invalidates the prompt cache, so hold it constant within a cached conversation, or use the per-message effort change that the current models support in beta, which preserves the cache.

Effort now reaches the cheapest tier. Haiku 4.5 did not support effort. Haiku 5.5 supports all five levels, which makes effort a lever on high-volume work for the first time. On thinking, the models differ. Opus 5.5 has adaptive thinking always on, and a request that tries to disable it returns a 400 error at every effort level. Sonnet 5.5's lowest setting is between_tools, which turns off up-front thinking and is rejected at xhigh and max. Haiku 5.5 accepts disabled thinking at high effort or below. An integration that set thinking off on the previous generation needs checking before it moves.

Fast mode, and when not to buy it. Fast mode is a research preview offering up to 2.5 times faster output on Opus 5.5, Opus 5 and Opus 4.8. On Opus 5.5 it costs $8 input and $40 output per million tokens, twice the standard rate and still below Fable 5.1's standard $10 and $50. It is available on the first-party Claude API only, not on Claude Platform on AWS or the partner clouds, and not with the Batch API. Buy it for a latency requirement you can name, not for general speed.

The tokeniser change, and why you cannot do this on a calculator

Anthropic's documentation states that Claude 4.7 and later models use a newer tokeniser which produces approximately 30 per cent more tokens for the same text, and that Sonnet 4.6 and earlier use the previous one. Haiku 5.5 has moved to the updated tokeniser as well, and Anthropic says it uses slightly more tokens per task than Haiku 4.5 did.

The trap is treating 30 per cent as a conversion factor. Anthropic's own wording is that the exact increase depends on the content and workload shape, and its advice is to measure with the token counting endpoint against the model you intend to run. A move from Haiku 4.5 to Haiku 5.5 therefore combines a price that is up to 90 per cent lower with a token count that is somewhat higher, and the net effect on the bill is a measurement question rather than an arithmetic one. Anyone who multiplies the two percentages together has produced a number no source supports.

Two second-order effects catch engineering teams rather than finance teams. If the same text occupies more tokens, a prompt that sat comfortably under Haiku 5.5's 100,000-token line on the old count may sit above it on the new one. And an output ceiling tuned on the previous generation may truncate an equivalent response, particularly now that thinking counts towards max_tokens on models where it cannot be turned off. Neither announces itself as a cost problem, and both become one when somebody raises the limits to fix them.

Shelf life: how long is this decision good for?

Better than the release cadence suggests. Anthropic commits to at least 60 days' notice before retiring a publicly released model, notifying customers with active deployments by email and in the documentation. The published tentative retirement dates for the current line-up are not sooner than 1 September 2027 for Fable 5.1, 22 September 2027 for Opus 5.5, 28 September 2027 for Sonnet 5.5 and 7 October 2027 for Haiku 5.5.

The previous generation is on a shorter clock. Opus 5 and Sonnet 5 carry horizons of July and June 2027. Haiku 4.5's is not sooner than 15 October 2026, the nearest of any active model, although no deprecation notice had been issued when we checked. Claude Sonnet 4.5 has one: it was deprecated on 30 September 2026 and retires on 30 November 2026, with Sonnet 5.5 as the recommended replacement. If anything in your estate still calls Sonnet 4.5, that is the date to plan around.

Two caveats belong in a procurement conversation. Retirement dates published by Anthropic apply to Anthropic-operated platforms; Amazon Bedrock and Google Cloud set their own schedules, so a model's lifecycle status can differ there. And a breaking change is already in place: on Claude 4.7 and later models, setting temperature, top_p or top_k to a non-default value returns a 400 error. Anthropic's recommended replacement is to omit them and use prompting instead. That is an API-level change, not guidance, and it is the sort of thing that turns a model upgrade into an afternoon of debugging.

Where residency and cost meet

For UK organisations with data-location requirements, the position on the first-party API is narrow and should not be overstated. Anthropic documents two values for inference_geo on Claude 4.6 and later models: global, the default, at standard pricing, and us, which pins inference to United States capacity at a 1.1x multiplier applied to input, output, cache writes and cache reads alike, including Haiku 5.5's higher long-prompt rates. No UK or EU residency tier is documented with a published price. Amazon Bedrock and Google Cloud set their own regional pricing.

The practical consequence is that residency constraints can push a UK deployment onto a route with different economics and a different retirement schedule, which is why the two decisions should be taken together. Our guide to Claude data residency for UK organisations covers what each route offers, and if your rollout involves personal data, the DPIA screening test is the right next step before switching anything on.

A short method you can run this week

List your workloads, not your teams. Assign each one a default of Sonnet 5.5. Move the mechanical, high-volume ones down to Haiku 5.5, check how many of their prompts exceed 100,000 tokens, and measure whether quality holds. Move up to Opus 5.5 only where you can name the cost of a wrong answer, and to Fable 5.1 only where Opus 5.5 demonstrably falls short and the retention position is acceptable. Then, before anyone argues about model choice again, turn on batching for everything asynchronous and caching for every stable prefix, and re-measure your token counts on the model you will actually run rather than trusting last quarter's numbers. Most of the saving is in those last two sentences.

Where The AI Consultancy fits

Choosing a model per workload, measuring where the tokens actually go, and building a unit-cost model that survives the next release is the kind of work our Claude implementation engagements cover, with the advisory side handled through Claude consulting. If you are choosing between plan routes rather than models, our Claude Team versus Enterprise guide covers the feature split, and our briefing on the Claude 5.5 family covers what changed in the releases and what to test before migrating.

All prices and model specifications verified on 10 October 2026 against Anthropic's published pricing documentation, models overview, model deprecations page and effort documentation, the plan comparison on claude.com/pricing, the Claude Help Center article on Covered Models, and the launch announcements for Claude Opus 5.5 (22 September 2026), Claude Sonnet 5.5 (28 September 2026) and Claude Haiku 5.5 (7 October 2026). First published 12 August 2026 covering Claude Opus 5, Sonnet 5 and Haiku 4.5, and rewritten on 10 October 2026 for the Claude 5.5 family. Sterling conversions use HMRC's monthly exchange rate for October 2026, USD 1.3456 to GBP 1, published 16 September 2026. Prices exclude tax. Percentage savings attributed to Anthropic are its own estimates. Anthropic publishes API prices in US dollars only, so all sterling figures are indicative conversions. This is general information, not procurement, tax or data-protection advice.