Claude Fable 5.1 vs Opus 5: Claim Memo Work (2026)

A claim memo is the narrative an examiner, a reinsurer, or a later adjuster has to trust: what happened, what the policy might do, what is still unknown, and which documents were in the file when the note was signed. Claude Fable 5.1 and Claude Opus 5 both write that narrative. Anthropic’s own docs tell you to start with Opus 5 for most workloads and to reach for Fable 5.1 when long-horizon reasoning still falls short. That is the opposite of “always buy the expensive one.”

This page is a memo picker, not a CAT model, not an AMS replacement, and not a license to auto-send reservation-of-rights letters.

TL;DR

• Start with Claude Opus 5 when the memo is a structured file note, a status letter, or a coverage outline that a senior examiner will still rewrite. List price is $5 / $25 per 1M tokens.

• Move named files to Claude Fable 5.1 when the packet is long, multi-document, and your evals on Opus 5 at high effort still miss the loss chronology. List price is $10 / $50; cache reads are $0.25.

• Independent AA Intelligence still prints 66 for Fable 5.1 at max, with about 4% of that eval’s output tokens routed to Opus through Anthropic’s default safety fallback. Do not treat the 66 as a pure-Fable-only run.

• Leave a second platform out when the AMS note field is already the memo. Add a hold when the narrative must wait on a claim number, a reviewer, and an archive write.

Key Takeaways

• Anthropic’s default for most work is Claude Opus 5 at $5 / $25; Claude Fable 5.1 is the demanding-reasoning SKU at $10 / $50 with $0.25 cache reads.

• Fable 5.1 leads independent Intelligence (66) for long knowledge work; Opus 5 is the cheaper daily driver for claim file notes.

• Forced tool_choice of type any or tool returns HTTP 400 on Fable 5.1. Keep auto and put schema in structured outputs.

• Claim memos are supervised communications. A fluent draft is not a sent notice.

• Mythos 5.1 is invite-only. It is not a public third option on this vs page.

How we evaluated

Weights assume a P&C claims unit that already has an AMS and a written authority ladder. We scored token economics and Anthropic’s own “start with Opus 5” guidance first, then independent knowledge-work rank, then breaking API changes. We did not invent METR hour numbers. We did not put a third lab into the vs table.

Adjuster jobs: 389,700 in 2025 according to BLS (Occupational Outlook Handbook), 389,700 claims adjusters, appraisers, examiners, and investigators, with 2025 median pay of $78,020. That wage is the rewrite cost. It is not a model score.

US P&C net premiums written print at $918.6 billion according to III (2025), $918.6B. Direct written premium is the bigger top line. Your claim memo should say which numbers in the file are paid, reserved, or still unknown.

The step-by-step build

Step 1 is naming the memo type. File note, status letter to the insured, internal large-loss chronology, and reservation-of-rights draft are four jobs. Only the last two are Fable-shaped. Opus 5 is the default for the first two unless your evals say otherwise. Write the four types on one page and date it. If marketing can add a fifth type without that date changing, you do not have a memo workflow.

Those token dollars are list math at $5 / $25 (Opus 5) and $10 / $50 (Fable 5.1), not a promise that either model will match your file. Use the table to stop putting Fable 5.1 on eight-thousand-token notes.

Step 2 is the model call. Anthropic’s models overview lists Claude API id claude-opus-5 at $5 / $25 per million tokens and claude-fable-5-1 at $10 / $50, both with 1M-token context and 128k max output on the current card. Default effort is high on both. Thinking is adaptive; on Fable 5.1 it is always on.

Step 3 is the worked example. A 14-adjuster unit running 120 large-loss chronologies a quarter can call claude-fable-5-1 with output_config.effort set to high, keep tool_choice at {"type": "auto"}, and refuse any JSON write that does not include a claim number. Anthropic’s Fable 5.1 what’s-new page documents output_config.effort, the HTTP 400 on forced tool_choice type any or tool, and the 1M / 128k ceilings. The 120, $10 / $50, and 128,000 figures are the sample, the list price, and the output cap — not a promised reserve accuracy. The same packet on Opus 5 uses claude-opus-5 at $5 / $25 when the chronology is short and a senior examiner will still rewrite the ending.

US Tech Automations belongs at the end of step 3: a configurable workflow can require the claim number, the memo type, and a reviewer id before it writes the AMS note, then flag any run that used Fable 5.1 on a two-paragraph status letter that should have stayed on Opus 5. That is a route, not a third model.

Step 4 is the hold. No insured-facing letter leaves without a human. Adjacent AMS hygiene — data entry, reporting, certificates — stays on CRM data entry, reporting software, and slow COI delivery. The memo model does not replace those tools.

Step 5 is the cache and fallback watch. Fable 5.1 cache reads list at $0.25. Anthropic estimates typical token bills about 25% lower than Fable 5, up to about 45% on agent loops. The AA Intelligence eval used Anthropic’s default safety fallback, with about 4% of output tokens routed to Opus. If your production path also falls back, log it. Silent fallback is how a “Fable-only” SOW becomes a mixed invoice.

Tooling landscape

Fable 5.1 AA Intelligence: 66 according to Artificial Analysis (1 September 2026), 66 on Index v4.1.1 at max with the fallback note, and about 4% of output tokens routed to Opus. That 4% is why this vs page exists: the eval already mixes the two products at the margin. US auto insurance written premiums sit near $340 billion according to NAIC (2024), $340B, which is why a claim memo that only lives in a personal chat history does not finish the file.

Fable 5.1 list input and output stay at $10 / $50 per 1M tokens according to Anthropic (checked 2026-09-03), $10 / $50 for Claude Fable 5.1 and $5 / $25 for Claude Opus 5 on that card, with Fable cache hits at $0.25. The Verge’s 1 September 2026 product note is dated access context, not a claims-handling benchmark.

OpenAI’s provider-run AutomationBench sheet prints 31.4% for Fable 5.1 and 26.9% for Claude Opus 5. That is a third-party lab’s competitor table, not Anthropic’s. It is directional for multi-app work and still secondary on a memo page.

The ROI math

A 120-memo quarter at 1.8 hours of assemble-and-rewrite per chronology is 216 hours. At $78,020 / 2,080 hours the wage floor is about $37.50 per hour, or about $8,100 of adjuster time before benefits. Token stickers at $5 / $25 or $10 / $50 will not be the bill that matters until the hold is real.

Those token bills assume about 80,000 input tokens and 16,000 output tokens per chronology. If your packets are shorter, both invoices shrink and the wage line still dominates. If your packets include full medical files, watch Fable 5.1’s $10 / $50 list and keep Opus 5 as the default for the two-paragraph note.

Do not tell a cost owner that Fable 5.1 is cheaper than Opus 5 on list I/O. It is not. Tell them cache at $0.25 can narrow an agent-loop invoice, and that Anthropic still wants Opus 5 as the daily driver.

Pitfalls and red flags

The first pitfall is running Fable 5.1 on every file note because the name is newer. That doubles list I/O for work Opus 5 was priced to do. The second is copying a tool_choice of type any from an older Claude integration and shipping HTTP 400s during a large-loss week. The third is treating AA Intelligence 66 as a pure-Fable number; about 4% of that eval’s output tokens went to Opus.

Red flags: auto-sending insured letters from either model; Mythos 5.1 in the picker; editing earlier turns on Fable 5.1 and invalidating thinking blocks; moving a Fable 5.1 thread onto Opus 5 and expecting those thinking blocks to survive; skipping the reviewer because the chronology “sounds right”; putting US Tech Automations in the vs table.

AWS Fable 5.1 is a Covered Model: up to 30-day retention plus AWS human review unless EFS-eligible ZDR through 31 December 2026. If your claim files cannot live under that retention story, do not pick Bedrock as the path without the ZDR paperwork.

Who this is for

This page is for a claims manager, large-loss examiner, or operations lead choosing a default Claude SKU for narratives. It assumes an AMS already holds the claim. Commercial AMS comparisons stay on Applied Epic vs HawkSoft.

Red flags: no claim number in the template; a plan to auto-issue coverage denials; no senior examiner who will read the last page; a cutover during a CAT spike.

Zapier, Make, or n8n can drop a finished memo PDF into the claim file, retry a failed write, and keep a run log if you design observability, idempotency, access, and retention. That is a fair DIY path for one stable recipe. It does not pick Fable 5.1 versus Opus 5.

When NOT to use US Tech Automations: skip it when the AMS note field is already the memo and a person already pastes one PDF into Claude, when a no-code scenario already files that PDF with a retry you trust, or when there is no second system to write.

A claims unit that still stores chronologies in personal Word folders will not get a cleaner exam trail by switching SKUs. The model can draft faster. It cannot invent the claim number, the authority ladder, or the person who will read page four. Do that staffing work first. Then pick Opus 5 or Fable 5.1 against the memo-type table, not against a keynote.

Pros and cons

Claude Fable 5.1

Pros

• Independent AA Intelligence 66 for long chronologies and multi-file packets.

• Cache reads $0.25; Anthropic’s typical-bill estimate is about 25% lower than Fable 5.

• 1M-token window, 128k max output, June 2026 cutoff, thinking always on.

• Live on paid Claude, API, and major clouds on 1 September 2026.

Cons

• List I/O $10 / $50 — double Opus 5 on the sticker.

• Forced tool_choice any/tool returns 400; earlier models cannot read Fable 5.1 thinking blocks.

• AA eval mixed in about 4% Opus tokens via default safety fallback.

• AWS Covered Model retention unless EFS/ZDR through 31 December 2026.

Claude Opus 5

Pros

• Anthropic’s documented starting point for most workloads.

• List I/O $5 / $25 — half of Fable 5.1 on the sticker.

• Same 1M-token window and 128k max output on the current overview card.

• Default effort high; adaptive thinking; May 2026 cutoff.

Cons

• Not the independent AA Intelligence leader on this page’s Fable cell.

• Cache reads $0.50 — higher than Fable 5.1’s $0.25.

• OpenAI’s AutomationBench sheet prints 26.9% versus 31.4% for Fable 5.1; secondary on a memo page.

• Will still miss long, messy chronologies that your evals already failed on Opus.

FAQs

Should a claims unit default to Claude Fable 5.1 or Claude Opus 5?

Default to Claude Opus 5 for structured file notes and status letters; move named long chronologies to Claude Fable 5.1 when Opus 5 evals at high effort still miss the loss order.

Does AA Intelligence 66 mean Fable 5.1 ran alone?

No. Artificial Analysis notes Anthropic’s default safety fallback, with about 4% of output tokens routed to Opus on that eval.

Can we force a tool call on Fable 5.1?

No. tool_choice type any or tool returns HTTP 400. Keep auto and use structured outputs or prompt the tool by name.

How should we price cache on claim-memo loops?

Fable 5.1 cache reads list at $0.25 per 1M tokens; Opus 5 cache reads list at $0.50. List I/O still favors Opus 5 at $5 / $25 versus $10 / $50.

When NOT to use US Tech Automations for claim memos?

Skip it when the AMS note is already the memo, when Zapier, Make, or n8n already files the PDF with a retry you trust, or when no second system needs the claim number.

Is Mythos 5.1 a third memo SKU?

No. Mythos 5.1 is the same weights with looser cyber and life-science gates, trusted-access only. It is not a public picker option.

The team at US Tech Automations can map a configurable claim-number hold around whichever Claude SKU you already picked. Review workflow pricing after you have named the memo types, the model ids, and the examiner who signs the last page.