Fable 5.1 and Mythos 5.1: what do the new Claude models change at work?

Two names. What is the actual difference?

Anthropic announced both models on 1 September 2026. Fable 5.1 offers broadly available access, while Mythos 5.1 has safeguards adapted for approved cybersecurity and life sciences work. Asking which is smarter can therefore send this comparison in the wrong direction.

Our conclusion: for a presentation, customer proposal or website project, the Mythos name alone does not establish an extra benefit. Value comes from task execution, available tools and verification.

• Anthropic: introducing Fable 5.1 and Mythos 5.1
• Fable: availability in Claude plans
• Mythos: access programs and restrictions

What can you use in Latvia today?

Fable users have two different routes: working in the Claude app or integrating the model into a system through the API. An app subscription and an API token bill are separate charging models. The calculator below explains API usage specifically.

Mythos access is provided through programs for vetted participants. The life sciences program starts as an invite-only beta; Mythos access through the Cyber Verification Program has been announced for a future rollout. Applying does not mean approval. Mythos powering the Claude Security product also does not mean every user can freely call the model through the API.

• Fable: availability in Claude plans
• Mythos: access programs and restrictions

What is new, and what do the big numbers mean?

The official Fable 5.1 API specification lists a one-million-token context window, up to 128,000 output tokens, and text and image inputs with text output. Adaptive thinking is always on; the effort setting controls its intensity. These are API specifications, not a promise of identical file limits in every app.

Anthropic highlights improvements in long-running coding, research, and work with documents, spreadsheets and presentations. API beta features include changing effort during a conversation and readable progress updates between tool calls. The documentation describes multilingual performance as comparable to Fable 5; the version number alone does not establish a leap in Latvian-language quality.

Reading an image and producing a working file are also different things. If the task requires a spreadsheet with calculations, provide an environment in which the file can be created and checked. Text arranged like a table in a chat is not yet a verified Excel model.

• Fable 5.1: official model specifications
• Fable 5.1: new features and behavior changes

Find a useful first task

A good starting point is work whose quality you know how to assess. Sorting 20 emails by topic is different from finding a contradiction across 20 documents. The second may need deeper analysis; the first may depend more on speed and cost per item.

Anthropic's selection guide suggests Opus 5 as a starting point for complex work, with Fable 5.1 for particularly demanding tasks or cases where evaluations with Opus still fall short. Below is our practical interpretation for four types of work.

• Anthropic: model selection guidance

Example: turn three proposals into a decision

Fictional teaching scenario: a small Latvian team is comparing three website maintenance proposals. A costs EUR 180 a month and includes two hours of content changes. B costs EUR 145, with changes charged at EUR 45 an hour. C costs EUR 210 and promises ‘priority support’ without specifying a response time. Assume the prices use the same tax basis and exactly two hours of content changes are needed each month.

A weak prompt asks for the best option. A useful prompt asks for a comparable cost basis and a separate list of unknowns. In this example, A is EUR 180 and B is EUR 235. C's full cost still needs checking because we do not know whether changes are included. ‘Priority’ is not a measurable response time.

Compare the three attached maintenance proposals. Goal: prepare a one-page decision brief for a team planning 2 hours of content changes per month.

First establish whether the prices use the same tax basis. For each figure and contractual condition, name the source file and section. Show the calculation, not just the total.

If a service, allowance or response time is unspecified, write ‘not stated’. Separate facts, assumptions and your recommendation. Do not invent missing contract terms.

Deliver in Latvian: a comparison table, the three most significant differences, unanswered questions and a conditional recommendation. Do not contact the suppliers.

This is an author-created task, not a quotation of Fable's output. You can give the same material to your current model. If both handle it equally well, this task does not yet justify buying a more expensive model.

Five adjustments that help with Fable 5.1

The new model's documented behavior is more useful than a universal ‘perfect prompt’. Anthropic notes less frequent search-tool use at low effort, denser prose, and a tendency to rewrite an entire file for a small change. The instructions below are our examples of defining the result you need.

1. If a fact must be current, request verification

   For example: ‘Check the price on the provider's website; include the link and date checked.’ If no search tool is available, ask the model to request a source instead of answering from memory.

2. Define the scope of work

   ‘Fix the application form's validation and check the error message on a phone.’ Add a specific acceptance criterion. Identify actions that require your decision, such as publishing or sending an email.

3. Name the reader and required format

   ‘Reply to the customer in Latvian, in 120–160 words. One idea per paragraph. Explain each technical term when first used.’ This creates an editorial target you can check.

4. Request an intermediate result when it helps a decision

   ‘After reviewing the sources, briefly identify contradictions; then prepare a recommendation.’ A progress update is useful when it lets you correct an assumption before more work builds on it.

5. Specify the extent of a file change

   ‘Change only the price calculation function and its related error handling. Show the change and verification result.’ Afterwards, inspect the file differences instead of relying only on a summary.

• Fable 5.1: prompting and workflow guidance

When does more thinking pay off?

The effort setting changes the intensity of analysis. The API documentation uses high as its starting point. Anthropic's announcement also specifies high as the Claude Code default, while the Claude app and Cowork default to medium. Talking about ‘default quality’ without naming the product and setting is therefore incomplete.

Try two settings on the same task. Check correctness first, then record the whole time commitment: waiting, reading, editing and asking again. A longer first answer may be a worse deal if editing it takes more time.

• Anthropic: introducing Fable 5.1 and Mythos 5.1
• Fable 5.1: prompting and workflow guidance

Cheaper cache reads are not 75% off the entire job

Fable 5.1's base API rates are USD 10 per million input tokens and USD 50 per million output tokens. Cache reads cost USD 0.25 per million, compared with USD 1 for Fable 5. The reduction applies to the reused portion of the input. If a long output makes up most of your bill, the total saving will be much smaller.

Caching lets you reuse an unchanged beginning of a prompt, such as instructions and a reference document. Writing it initially costs more than ordinary input: USD 12.50 per million with a five-minute lifetime, or USD 20 with a one-hour lifetime. A subsequent request must match the stored prefix; a successful read refreshes its lifetime. A changed prefix or expired cache may require another write.

• Claude API: token and tool pricing
• Claude API: prompt caching

Calculate the cost of your API scenario

Change the shared context size, output and number of successful cache reads. Notice the two comparisons: the benefit against working without caching, and the difference against Fable 5 at identical token usage. They answer different questions.

Tokens are units of text or other data processed by the model; they are not a word count. Base a budget for your Latvian text on API usage records. Include the relevant tool instructions and results in the inputs, and all billed output tokens, including thinking, in the output field.

The starting example uses 10 requests, each with 32,000 shared input tokens, 2,000 fresh input tokens and 3,000 output tokens. One five-minute cache write and nine successful reads cost USD 2.172. The same tokens without caching would cost USD 4.900. Fable 5 with the same cache usage would cost USD 2.388. Here, the new version's pricing reduces the bill by about 9%, not 75% overall.

• Claude API: token and tool pricing
• Claude API: prompt caching

Compare models without an impressions contest

Public benchmarks help you decide what to try, but they do not test your particular Latvian-language work. In Anthropic's published evaluations, the tools, effort settings and safeguard interventions matter; in some cases, another model completes a task. We therefore avoid presenting a misleading universal winner chart here.

For your pilot, collect, for example, ten previously completed tasks with verifiable results. Include a routine case, conflicting documents, a missing fact, Latvian terminology and an incorrect figure. Ten tasks are a starting sample for investigating a workflow, not statistical proof that one model is superior.

1. Fix the task before testing

   Keep the same sources, prompt, tool access and evaluation criteria. Record the model version and effort setting. If one model can search and the other cannot, you are testing two workflows rather than an isolated model capability.

2. Evaluate anonymously

   Rename the answers A and B and vary their order. The reviewer first records evidence, errors and necessary corrections. Reveal the names afterwards.

3. Count completed, accepted work

   Record accepted deliverables, critical errors, total API costs and human review minutes. If no answer is usable, the cost per accepted result is not zero: there is no accepted result.

4. Repeat the borderline cases

   One good answer can happen by chance. Repeat tasks where the models differed and add the following week's work to your sample. If you change the prompt, preserve the original comparison separately.

• Anthropic: introducing Fable 5.1 and Mythos 5.1

Build an evaluation sheet for one result

This is a simple rubric suggested by Hexa Academy, which you can adapt to your work. Check the answer against the sources before scoring it. A critical factual error or unauthorized action prevents acceptance even when the prose is polished.

This tool's acceptance rule is at least 8 out of 10, full marks for facts and scope compliance, and no critical error. It is a working rule, not a scientifically validated quality index.

Delegate a stage of work with a clear handover

In a complex project, AI is most useful when it can finish a meaningful stage: examine sources, prepare a solution and verify a specific result. ‘Work until everything is done’ leaves too much interpretation when done has not been defined.

For a website, the handover could include a preview of the changes, a mobile verification result and a short list of unresolved questions. For data analysis, it might be a file with formulas, source dates and a checked total. In both cases, the next action is understandable to a person.

Distinguish a recommendation from an action. Permission to draft an email does not establish who should receive it or when it should be sent. When implementing an agent, enforce this boundary through tool permissions and approval mechanisms as well as prompt text.

Data and migration: two checks before adoption

The API documentation specifies 30-day data retention for Fable 5.1 and Mythos 5.1, with exceptions requiring explicit authorization. The announced Enterprise Frontier Safeguards approach puts monitoring data in customer-controlled infrastructure and is scheduled for a phased autumn rollout. Eligible customers may have authorized zero retention for Fable in the meantime; it is not an automatic condition for every account.

For an initial trial, use public, fictional or appropriately anonymized material. Before entering company data, establish which product and agreement apply, what data will be sent and who can access tool results. The model name alone does not answer those questions.

• Claude API: data retention conditions
• Enterprise Frontier Safeguards: announced rollout
• Migrating to Fable 5.1: developer guide

Answer one useful question in the first week

Choose one repeatable job and define what finished means. Prepare the sources, save the original prompt, and try Fable 5.1 alongside your existing solution. Use the same rubric and include review time.

At the end of the week, aim for a specific finding: in this task, the new model produces more accepted results, reduces correction work or enables analysis you could not previously carry out. If the evidence is not there yet, the pilot has identified the next check. That is still far more useful than another impressive demonstration.