GPT-6 Astra or Claude Fable 5.1: which should you buy?

On this page

1. 1.What is different on paper
2. 2.The comparison most organisations should be making
3. 3.What a seat costs, and the catch in Anthropic’s Standard plan
4. 4.Where Fable is cheaper: large documents and repeated material
5. 5.What the evidence supports, and what it does not
6. 6.The model is not the product
7. 7.Which one to choose
8. 8.Governance decides the account before you pick a model
9. 9.How to brief each model
10. 10.Test before you standardise
11. 11.Where this article stops
12. 12.Where to go from here

OpenAI and Anthropic released their top models two days apart. Claude Fable 5.1 arrived on 1 September 2026 and GPT-6 Astra on 3 September. Each costs $10 per million input tokens and $50 per million output tokens through its API, and each takes about a million tokens of input.

The short answer for a UK SME or charity is that neither should be your default model. Most office work belongs on a cheaper model from either vendor.

Where a frontier model does earn its cost, four things decide between the two. They are the tools around the model, what large documents cost, how your agents will run, and which account holds your data.

This guide is for leaders and budget holders making that call. Nearly all the performance evidence is the vendors’ own, and we say where it runs out. We do not sell licences for either.

Our earlier piece on GPT-6 Astra on its own covers that model in more depth.

What is different on paper

The two specification sheets are closer than the launch coverage suggests.

A context window is the amount of text a model can work with in one go, measured in tokens. A token is roughly three quarters of a word. The 50,000-token gap between the two windows will not decide anything for an SME.

Two rows in that table matter more than the rest. Cached input and very large requests are where identical headline prices turn into different bills. We come back to both below.

One caution on “identical” prices. Anthropic’s pricing page says its current tokenizer produces about 30% more tokens for the same text than its earlier models did. OpenAI publishes no comparable figure for Astra.

So the same document may not be the same number of tokens on each model. Nobody has published a cross-vendor count, and only a test on your own documents shows the true cost per page.

The comparison most organisations should be making

The question is rarely Astra or Fable. It is Astra or Fable against a model costing a fraction of the price.

Anthropic makes this case itself. Its model guidance says to start with Claude Opus 5 for most workloads. It reserves Fable 5.1 for demanding reasoning and long-running agent work, or for when tests on Opus 5 at higher effort still fall short.

Opus 5 costs $5 and $25 per million tokens, half of Fable. Sonnet 5 costs $2 and $10, a fifth of Fable. OpenAI’s pricing points the same way: on its Business credit schedule, Astra consumes about 2.5 times the credits of GPT-5.6 Sol for the same tokens.

Routine correspondence, meeting notes, straightforward drafting, social posts, classification and simple FAQ answers do not need a frontier model. A stronger model existing is not a reason to pay for it. Test the cheaper model first, and move up only when it fails on your work.

What a seat costs, and the catch in Anthropic’s Standard plan

The subscription prices match almost line for line. All figures are the vendors’ published US dollar prices, before VAT and before any negotiated terms.

At monthly billing, that gives the following illustrative seat costs before any extra credits.

Nobody should put every employee on Premium. The sensible pattern is Standard seats for most people and Premium, or API access, for the few who do research, analysis or engineering work.

The catch sits in the Standard tier. OpenAI’s Business Standard includes a limited amount of Astra use on the surfaces where it has rolled out. Anthropic’s Team Standard, and its individual Pro plan, do not include Fable 5.1 at all.

Fable use on those plans draws on pay-as-you-go credits from the first use. Fable is included, within published limits, only on Max and Team Premium.

So a team that expects to use Fable often on Standard seats should model that credit spend before signing. A team that wants occasional frontier access across ordinary seats will find OpenAI’s arrangement cheaper to run, on current published terms.

Where Fable is cheaper: large documents and repeated material

Two published pricing rules make Fable materially cheaper for some API workloads. The headline rates match; these rules do not.

The first is caching. An application often sends the same long block of text repeatedly, a policy library or a set of case files. The vendor charges a lower rate for that repeated part.

Astra charges $1 per million cached tokens. Fable charges $0.25, four times less.

The second is the large-request premium. Once an Astra request passes 272,000 input tokens, OpenAI doubles the input and cache rates and charges 1.5 times on output, for the whole request. Anthropic’s pricing page says a 900,000-token request is billed at the same per-token rate as a 9,000-token one.

Our arithmetic shows what that means at standard API prices. These are model-token costs only, with no tool fees, retries or staff time.

Now take a repeated job: a 200,000-token reference pack, 10,000 new tokens per case, 2,000 tokens of output, run 100 times. Astra costs about $0.40 a run, or $42.50 for the batch including the cache write. Fable costs about $0.25 a run, or $27.50.

Fable is around 35% cheaper on that job.

Below 272,000 tokens and without heavy reuse, the two cost the same. A sole consultant running dozens of tasks a month will not notice either figure against their own time. An automated system pushing hundreds of thousands of tokens through many stages will.

What the evidence supports, and what it does not

Four days is not long enough for independent testing. Each vendor reports strong results on coding, research, computer use and professional tasks. OpenAI compares Astra directly with rival models on some of them.

Those are vendor claims. One independent model catalogue we checked labels its Astra figures as OpenAI self-reported rather than verified.

That leaves a short list of things the evidence does support:

• each model has the features, limits and prices above;
• Fable’s caching and large-request pricing are cheaper by published rule;
• Anthropic classifies Fable as slower than its other models;
• Anthropic recommends starting most work on the cheaper Opus 5.

Everything else is a hypothesis for you to test.

Claims you will hear this month that nobody can yet support:

• that one is the best coding model;
• that one writes better business documents;
• that one is more accurate, or makes things up less often;
• that one is faster at completing real business tasks;
• that one needs fewer tokens per finished job;
• that one runs browser and computer tasks more reliably;
• that Anthropic’s claim of up to 45% cheaper agent workloads applies against Astra. That figure compares Fable 5.1 with its own predecessor, not with OpenAI.

The model is not the product

A model name on a price list tells you less than it seems.

GPT-6 Astra is a model with a context window, prices and an API. ChatGPT adds Work, memory, connectors and account controls around it, and Codex adds a separate software environment. OpenAI’s own help pages describe different availability and usage allowances across Chat, Work and Codex for the same model.

The same holds for Anthropic. Fable 5.1 the model is not Claude.ai, Claude Research, Cowork, Claude Code, the Microsoft 365 integrations or Projects. What each one can do, how much you can use, and how your data is handled depend on the product and the plan.

A benchmark result says nothing about what a member of staff can do on your chosen plan. It does not tell you whether they can securely search SharePoint, produce a PowerPoint, operate a browser or write to the CRM. Buy the model, the application, the plan, the data controls and the workflow as one system, and test that system.

Which one to choose

These are our recommendations, built on the vendors’ published positioning and prices. Each is a starting point for a test, not a verdict.

Choose GPT-6 Astra when:

• you already run ChatGPT Business or Enterprise and want to extend into Work or Codex without a second AI estate;
• you are building a custom application that needs the breadth of OpenAI’s Responses API tools in one place: web search, file search, code execution, computer use and structured outputs;
• you want occasional frontier access across ordinary Standard seats, because OpenAI includes some Astra use where Anthropic charges credits from the first token;
• a controlled test shows Astra fails materially less often than GPT-5.6 on your hardest tasks.

Choose Claude Fable 5.1 when:

• your workloads reuse large blocks of material, or regularly exceed 272,000 tokens in a request, so the caching and large-request pricing bites;
• a Claude, Cowork, Claude Code or Microsoft-centred environment suits how your people work;
• a controlled test shows the cheaper Claude models fail on your long-running reasoning or agent work.

Either will do when the work is drafting, research synthesis, report preparation, policy first drafts, data interpretation or multi-file analysis. No independent evidence yet says one writes a better board paper or grant application than the other. Decide on the tools around the model, the governance controls and your own measured results.

Use a cheaper model when the work is routine email, social content, meeting notes, classification, ordinary summaries, extraction, FAQ answers or first-pass CRM drafts. That is most of the work in most organisations.

Use both vendors when a second opinion is worth paying for on a high-value decision, or when you want resilience against one vendor’s outage. Software teams sometimes find each better on different repositories. Do not pay twice for every low-stakes task and call it governance.

Deploy neither without extra controls when the task is a hiring, disciplinary, safeguarding or eligibility decision. The same applies to legal submissions, payments, contract acceptance, broad write access to customer or financial records, and autonomous public communications. Each of those needs a person accountable for the outcome.

Governance decides the account before you pick a model

For organisational use, start with the account and the contract, not the model.

OpenAI says data from its Business, Enterprise and API offerings is not used to train its models by default. It describes encryption, data processing agreements, single sign-on at enterprise level and audit access. Anthropic says qualifying API and Claude Code Enterprise customers can obtain zero data retention arrangements, with exceptions for legal and safety functions.

Neither statement extends to a personal Plus, Pro or Max subscription. An employee’s own account is not an approved place for client, donor or beneficiary data just because the model inside it is the same.

The Information Commissioner’s Office (ICO) guidance on AI and data protection covers accountability, lawfulness, fairness, transparency, accuracy, security, data minimisation, individual rights and automated decisions. The ICO says parts of that guidance are being updated following the Data (Use and Access) Act 2025. Until the updates land, the existing guidance and data protection law apply.

Whether a particular use needs a data protection impact assessment is a question for your data protection lead, against current ICO guidance. It is not something to infer from model capability.

The Charity Commission’s published position is that trustees remain responsible for decisions, and that decision-making must not be delegated to AI. It highlights convincing but inaccurate, biased, plagiarised or copyright-infringing output, expects human oversight, and calls for extra care where information concerns vulnerable beneficiaries.

A more capable model loosens none of that. The more an agent can retrieve and act on records, the more damage over-permission can do.

The practical baseline for a UK SME or charity is short:

• use the correct organisational account;
• send the model the minimum data the task needs, and redact where practical;
• settle what the data is for, the lawful basis and how long you keep it before scaling up;
• restrict connectors and credentials, and start agents read-only;
• require a person’s approval before payments, deletions, contractual changes, hiring decisions, beneficiary outcomes and publication;
• log what agents do;
• treat web pages, emails and retrieved documents as untrusted input. OpenAI reports better prompt-injection resistance in Astra. Better is not immune.

One distinction earns its place in any workshop. An agent can read, draft, recommend, write or commit, and each step carries more risk than the last:

• reading is low risk;
• drafting can mislead people;
• recommending raises accountability;
• writing changes your records;
• committing, which means sending, paying, deleting, contracting or submitting, is the highest-risk class.

Controls should tighten at each step.

For charities, that gives a simple ladder:

How to brief each model

Both vendors have moved away from prompt formulas. The good prompt for either is a brief. It states the outcome and the evidence that governs it. Then it sets the hard constraints, what the model may decide alone, what needs approval, and how success will be checked.

The differences are behavioural, and each vendor documents its own.

Astra follows instructions closely and asks a clarifying question when missing information could change the result. Tell it when to assume and when to stop. Our Astra article covers the rest.

Fable wants long documents before the question, not after. Anthropic reports that ordering improving some long-document tasks by up to 30% in its own tests. It responds well to a few examples where format matters.

Fable can read “could you suggest changes?” as a request for advice, so say “make the changes” when you mean it. In long agent runs, tell it to finish all authorised work and to run independent steps in parallel.

At low effort it may skip searching, so tell it to search for anything recent.

Test before you standardise

Do not change your default model on a launch announcement. Run the test we use before recommending any model change, for any vendor.

1. Take 20 to 50 real tasks from your own recent work, with personal data removed and a clear pass mark for each.
2. Run each on the cheapest plausible model, then on Astra, then on Fable, with the same brief.
3. Record task success, material errors, minutes of human correction, elapsed time, tokens including cached and reasoning tokens, tool calls, retries, and cost per accepted result.
4. Where an agent is involved, add three more: did it stop early, did it repeat work, and did it cross an approval point it should have stopped at.
5. Decide per type of work, not per organisation. Most work will stay on the cheaper model.

For a software-heavy SME, run both models on representative issues from your own repositories and count tests passing, regressions and scope creep. For a research or professional services firm, run the 20 to 50 task comparison blind.

For an operations-heavy SME, decide first whether the process needs an agent at all. A fixed trigger, field mapping and notification flow is usually better as ordinary automation, with a model only at the fuzzy step.

Where this article stops

Everything above rests on the vendors’ documentation, pricing pages and plan guidance as published by 5 September 2026, plus the ICO and Charity Commission positions. Fable 5.1 had existed for four days and Astra for two. Independent benchmark replications and months of practitioner experience cannot exist yet.

Launch benchmarks are not comparable across vendors. Tool access, prompting, effort settings, retries and benchmark versions all differ, so a few points of difference mean nothing without a common test.

Plan entitlements are moving. OpenAI’s Astra rollout across ChatGPT surfaces was incomplete at the cut-off, and Anthropic had just revised its Fable plan access.

Enterprise pricing, retention terms and regional processing are contractual for both vendors. Check your quote, order form and data processing agreement, not a pricing page, before you commit.

Where to go from here

For any AI purchase, our buyer’s guide to choosing an AI tool sets out the criteria and the questions to put to vendors. If you use ChatGPT today, ChatGPT for business covers the governance basics that do not change with the model.

If you want an independent view of which model, plan and account fit your work, our AI consulting and advisory service runs exactly this comparison. We take no vendor commissions. Book a free discovery call.

Back to top