Claude Fable 5.1 vs Claude Opus 5: 2 Memo Paths 2026

Claude Fable 5.1 vs Claude Opus 5 for accounting knowledge memos is a ledger-adjacent ranking, not a chat ranking. A firm that writes tax research, PBC explanations, and variance memos needs a model that can hold the binder, cite the rule, and still leave the journal entry to a person. Claude Fable 5.1 is Anthropic's 1 Sep 2026 public flagship for long knowledge work. Claude Opus 5 is the cheaper, default Anthropic workhorse at $5 input / $25 output per 1M tokens. This vs page names exactly those two products. US Tech Automations is not a third model; it is the hold between the memo and the books.

Anthropic's own model guide still tells most workloads to start on Opus 5 and move to Fable 5.1 when Opus at higher effort still falls short. That sentence is the buying fork for a firm: stay on Opus for routine memos, or pay Fable list $10 / $50 to chase the extra Briefcase Elo.

TL;DR

• Pick Claude Fable 5.1 when AA-Briefcase and GDPval-style memos are the ranking: Fable Briefcase Elo: 1,694 versus Opus 5 at 1,685, effectively tied, with Fable ahead on analytical quality.

• Pick Claude Opus 5 when the token bill is the constraint: list $5 / $25 and cache hits $0.50 per 1M versus Fable $10 / $50 and Fable cache reads: $0.25 / 1M.

• Do not treat the 1,694 Elo as a pure Fable-only run; the AA eval used Anthropic's default safety fallback with about 4% of output tokens routed to Opus.

• Orchestrate memo-to-ledger only after unique client IDs, a reviewer, and a journal-entry owner exist.

Quick-answer FAQs up top

Which Claude model wins knowledge-memo Elo?

Claude Fable 5.1 wins the published AA-Briefcase absolute Elo at 1,694 versus Claude Opus 5 at 1,685, a gap Artificial Analysis calls effectively tied. Fable is ahead on analytical quality (2,025 vs 1,980) and behind on presentation (1,495 vs 1,572). If your partners care about the analysis, Fable. If they care about the slide polish, Opus is not the weaker presentation model.

Is Opus 5 cheaper than Fable 5.1 for firm memos?

Yes on list input and output: $5 / $25 versus $10 / $50. Yes on uncached work. Not always on cache-heavy binder loops: Fable 5.1 cache reads are $0.25 per 1M versus Opus 5 at $0.50. A PBC binder that is reread all week can flip the cache column. A one-off research paste cannot.

Can a firm stay on Opus 5 after Fable 5.1 launched?

Yes. Anthropic still positions Opus 5 as the default for most workloads. Fable 5.1 is the step up for long-horizon agentic coding, multistep research, and document/spreadsheet/slide work when Opus at higher effort still misses the rubric. Leaving Opus is optional.

Do these models post the journal entry?

No. Neither Claude Fable 5.1 nor Claude Opus 5 is a general ledger. The memo is evidence. The JournalEntry is a separate object with a reviewer. If you let the model post, you have built an unreviewed write into the books.

When does a workflow layer belong between memo and ledger?

When the same client ID must move from the knowledge memo to a PBC tracker, an engagement binder, and a journal — and a person must sign the last step. Native Claude chat is enough for one partner, one PDF, one email. US Tech Automations is the hold, not the CPA.

Will Fable 5.1 break our tool caller?

It can. Forced tool_choice of type any or tool returns HTTP 400. auto and none still work. Thinking is always on. Earlier Claude models cannot read Fable 5.1 thinking blocks. Editing earlier turns invalidates thinking. Opus 5 does not carry that particular 400.

Who this is for

This page is for accounting firm partners, CAS leads, and tax managers who already draft knowledge memos in Claude and are deciding whether Claude Fable 5.1 is worth 2× Opus 5 list on the same binder. Typical shape: a multi-person firm with Karbon, TaxDome, or a shared drive, plus QuickBooks or Sage as the ledger, where PBC text and variance explanations currently live in Word.

Red flags: skip this comparison if the firm has no written memo standard and is really shopping for a chatbot; skip it if the general ledger is closed to any model write and you only needed a summarizer — Opus 5 is enough; skip it if AWS Covered Model retention for Fable 5.1 is a client-file blocker and you have not secured EFS/ZDR through 31 Dec 2026; skip it if you need a third product in the title.

according to BLS, $79,880 is the median annual wage for accountants and auditors on the May 2023 Occupational Outlook estimates. That wage is the labor a memo model is supposed to draft against, not replace. according to AICPA, 16 hours is the length of the Uniform CPA Examination, which is a reminder that the credential still sits on people, not on a context window.

How we evaluated knowledge memos

We scored Claude Fable 5.1 and Claude Opus 5 the way a firm uses a model after busy season: memo quality on independent knowledge-work Elo, reconstructable token price including cache, whether the API caller will 400 on forced tools, and whether a finished memo can sit next to a journal without posting it.

How the automation works

Worked example

When a firm pins the engagement binder with Anthropic cache_control, Claude Fable 5.1 can reread that binder at $0.25 per 1M cache hits while Claude Opus 5 rereads it at $0.50, and the same run can draft a knowledge memo that a reviewer files before anyone touches a QuickBooks JournalEntry. Official cache field: Anthropic prompt caching. Three figures on that recipe: Fable 5.1 cache reads $0.25 per 1M, 5-minute cache writes $12.50 per 1M, 1-hour cache writes $20.00 per 1M. US Tech Automations is the workflow step that keeps cache_control on the binder, opens the memo job, and routes a CPA review before the ledger write. It does not replace Fable. It does not replace Opus.

The rest of the path is boring on purpose. PBC request comes in. Tracker updates. Model drafts the explanation from the cached binder. Reviewer edits. Only then does the workpaper close. Onboarding still has to create the client ID that the memo hangs on; see CAS client onboarding in 30 and audit PBC request tracking. Knowledge management is the binder, not the chat log; see knowledge management for accounting firms.

according to IRS, 11 is the return count that pulls specified tax return preparers into the e-file mandate. That is the kind of hard number a knowledge memo must not invent. If the model cannot see the sourced rule, the memo is a draft, not a filing position.

Fable 5.1 thinking is adaptive and always on. Moving a conversation from Fable 5.1 back to Opus 5 drops Fable thinking blocks. A firm that "fails over to Opus when Fable is slow" must log that drop or the memo trail will look complete when the reasoning is gone.

Accounting memos fail in three predictable ways, and neither Claude SKU fixes them by itself. The first is a rule without a cite: the draft names a section of the Code or a state apportionment method and never attaches the sourced paragraph, so a reviewer has to rebuild the authority. The second is a client without an ID: the memo is titled with a trade name that does not match the engagement letter, the PBC tracker, or the QuickBooks customer, so the workpaper cannot be found in a week. The third is a conclusion that quietly becomes a posting: someone copies a suggested accrual into the ledger because the model sounded sure. Fable 5.1's extra analytical Elo helps the first failure if the binder is in context. It does not help the second or the third.

A CAS shop writing monthly variance memos has a different mix than a tax shop writing a one-time research position. Variance memos are short, repetitive, and cache-friendly: the chart of accounts and last year's memo template should sit behind cache_control all month. Research positions are long, unique, and uncached: you pay list input. Opus 5 is usually the variance SKU. Fable 5.1 is usually the research SKU. If you run both on one API key, tag the traces. Otherwise finance will see one Claude bill and argue about a model they are not actually using.

Staffing does not disappear because Elo moved nine points. The person who used to write the first draft still has to confirm the numbers against the trial balance, and the person who signs still has to own the conclusion. What changes is the queue: more first drafts in the same afternoon, more review, more risk that a fluent wrong answer ships. That is why the reviewer is in the recipe as a named step, not as a slogan about "human in the loop."

Keep the memo object boring. Title, client ID, period, question, sources, conclusion, open items. If Fable 5.1 writes a five-page narrative with no sources block, the extra Elo was spent on prose. If Opus 5 writes a one-page note with three cites and a number that ties to the TB, Opus won that file. Score 20 real memos the way you score 20 real returns: did it tie out, and how long did review take.

Benchmarks

according to Artificial Analysis, 1,694 Elo is Claude Fable 5.1 (max) on AA-Briefcase versus 1,685 for Claude Opus 5 (max), with GDPval-AA v2 at 1,853 versus 1,824 and overlapping confidence intervals. according to Artificial Analysis, $3.69 is Fable 5.1 Intelligence cost/task (max). Opus 5's cost/task is the reason many firms stay put: lower list, enough Elo for most memos.

Source line: Artificial Analysis 1 Sep 2026 Fable article; Anthropic pricing 1 Sep 2026. AA Fable cell includes ~4% Opus fallback tokens. Presentation Elo still favors Opus 5 on Briefcase.

Tool / build comparison

Fable 5.1 on AWS is a Covered Model: aws_review mode, up to 30-day retention plus AWS human review unless EFS-eligible ZDR through 31 Dec 2026. Client files in a tax binder are exactly the content that line is about. Opus 5 remains the less surprising cloud SKU for firms that have not had that conversation.

Cost and payback

Source line: arithmetic from Anthropic's 1 Sep 2026 price card, not a firm TCO study. Cache writes omitted; they depend on 5-minute versus 1-hour breakpoints. Batch is half on both cards when the memo can wait.

Opus list output: $25 / 1M is why CAS teams keep it as the default. Fable's cache cut is why research teams move the binder loop. Payback is not "Fable is cheaper." Payback is "does 9 extra Briefcase Elo and 45 extra analytical Elo justify ~2× uncached tokens?" For most write-ups, no. For multi-hour PBC narratives with a warm cache, maybe.

A partner who only sees the monthly Claude invoice will pick Opus 5 every time, and they will often be right. A tax group that rereads a 400-page binder for two weeks should model cache hits before they refuse Fable 5.1. Run one week's traces: share of tokens that were cache hits, share that were output, share that were uncached input. If cache hits are under a quarter of input, Fable's $0.25 column is a footnote and Opus wins on dollars. If cache hits are most of the binder loop, Fable's list I/O still costs more, but the gap shrinks enough to put Elo back on the table.

Do not mix Fable 5.1 into an Opus 5 conversation and expect a clean audit trail. Thinking blocks do not travel backward. If you need Opus as overflow, start a new thread with the memo object and the sources, not the hidden reasoning. That is operationally annoying and cheaper than explaining a missing thought process in a workpaper review.

according to Anthropic, $0.25 per million tokens is the Fable 5.1 cache-hit price, at 0.025× base input, while Opus 5 cache hits remain $0.50. Anthropic estimates typical token bills ~25% cheaper than Fable 5, up to ~45% on agent loops, because of that cache cut — list I/O is unchanged from Fable 5.

When NOT to use US Tech Automations: if partners only paste a PDF into Claude and email the memo, stay in Claude. If Karbon or TaxDome already files the only workpaper you need, keep that system. Zapier, Make, or n8n can move a finished memo into a tracker if you already own the recipe, the retries, and the access list; this page does not claim those tools cannot retry or cannot log. US Tech Automations is for the case where cache_control, the memo, and a JournalEntry draft must share a reviewer before anyone posts.

Pros and cons

Pros

• Claude Fable 5.1: AA-Briefcase 1,694 and GDPval-AA 1,853; cache reads $0.25; 1M context and 128k max output; live on paid Claude, API, AWS, GCP, and Foundry as of 1 Sep 2026; stronger analytical Elo on Briefcase.

• Claude Opus 5: list $5 / $25; cache hits $0.50; still Anthropic's default recommendation for most workloads; forced tool_choice still works; stronger Briefcase presentation Elo (1,572 vs 1,495); cheaper uncached memos.

Cons

• Claude Fable 5.1: 2× Opus list I/O; AA cost/task $3.69; ~4% Opus fallback in the AA cell; forced tool_choice any/tool returns 400; AWS Covered Model retention; thinking-block binding will surprise a failover router.

• Claude Opus 5: slightly lower Briefcase and GDPval Elo; cache hits 2× Fable 5.1's $0.25; not the SKU Anthropic points at for the longest agentic research jobs.

Key Takeaways

• Claude Fable 5.1 and Claude Opus 5 are effectively tied on AA-Briefcase (1,694 vs 1,685); Fable wins analysis, Opus wins presentation.

• List prices: Fable $10 / $50 with $0.25 cache reads; Opus $5 / $25 with $0.50 cache hits.

• Start on Opus 5 unless your evals at higher effort still fail; that is Anthropic's own fork, not a slogan.

• Knowledge memos do not post the ledger. A reviewer still owns JournalEntry.

• Native Claude is enough for one paste; add a workflow hold when the binder, the tracker, and the books must share a client ID.

The workflow layer's homepage is US Tech Automations. Use it when the memo has to wait on a CPA. See agentic workflows after the memo-to-ledger sequence is written down.