Claude Fable 5.1 vs Claude Opus 5: Chart Packets (2026)

Chart-summary packets for rounds are a retrieval problem wearing a writing problem’s clothes. The overnight census, last labs, meds, and prior notes already live in the EHR. Claude Fable 5.1 and Claude Opus 5 can both turn that pile into a one-pager. The buying question is whether you pay Fable’s $10/$50 list for long-horizon packet work, or start on Opus 5 at $5/$25 as Anthropic’s own model guide still recommends for most workloads.

This page compares those two Claude IDs for pre-rounds packets in a US medical practice or hospitalist group. It is not an EHR replacement, not a diagnosis engine, and not a HIPAA designation. The EHR remains the system of record. The model drafts. A clinician signs.

TL;DR

• Start with Claude Opus 5 when the packet is a short, structured recap (problem list, overnight events, today’s labs) and you want half the list price ($5 input / $25 output per 1M versus $10 / $50).

• Move to Claude Fable 5.1 when the packet is a long multi-note synthesis, a multi-step research pass across attached PDFs, or when Opus 5 at high effort still misses the overnight story.

• OpenAI’s 3 September provider table, used here only for the two Claude rows, shows Fable 5.1 ahead of Opus 5 on HealthBench Professional (56.6% vs 54.5%) and on Humanity’s Last Exam with tools (65.0% vs 63.6%). Independent AA knowledge-work Elos also sit slightly with Fable, with overlapping confidence on GDPval-AA.

• Orchestrate EHR → packet → model → clinician hold. Do not let a model write back to the chart unattended.

Who this is for

This comparison is for a CMIO, practice administrator, or hospitalist operations lead at a primary-care group, multi-site clinic, or hospitalist service that already runs an EHR (Epic, DrChrono, athenahealth, or similar) and still builds morning packets in Word, a shared drive, or a resident’s notebook. Typical stack: EHR + e-fax + a team inbox + a printer. Firm size is a named clinician owner plus a documentation lead, not a solo chat experiment on a personal Claude login.

Red flags: skip both models if the EHR’s native snapshot (problem list + last note + last labs) already is the packet. Skip Fable 5.1 if the only job is a 200-word overnight recap and Opus 5 already passes your rubric. Skip any cloud ID if you cannot name a BAA, a retention path, and a place the draft lives outside the model vendor’s default logs.

When NOT to use US Tech Automations: leave it out when the EHR already generates the rounds list, when a documentation vendor already files the overnight summary, or when a single Zapier, Make, or n8n scenario already drops a FHIR read into Slack for the charge nurse. Zapier, Make, or n8n can retry a failed EHR read and keep a run log if you design observability, idempotency, access, and retention. That is a fair DIY choice for one stable recipe. A proposed agent design would add a patient-encounter ledger and a clinician hold before anything resembling a note — not a claim that no-code cannot retry.

Rounds exist inside a large spend envelope. According to CMS National Health Expenditure data, 7.2% was 2024 national health spending growth, to $5.3 trillion, or $15,474 per person. Physician and clinical services are the packet’s home: according to CMS, $1,109.7 billion was 2024 physician and clinical services spend (8.1% growth). A missing overnight event is not a typing inconvenience; it is time stolen from that clinical hour.

Related reading: primary-care documentation backlog, intake form software for medical practices, and prior-authorization status updates.

How we evaluated

We scored Claude Fable 5.1 versus Claude Opus 5 as packet workers, not as chat products. Weights assume a morning rounds list with a clinician signer. A coding-agent team should not use this rubric.

Health and professional scores used OpenAI’s 3 September launch table for the two Claude rows only, labeled provider-run. Knowledge-work Elos used Artificial Analysis’s Fable 5.1 article. List prices used Claude API pricing counted 3 September 2026. We did not invent METR horizon hours and we did not write diagnostic or exploit how-tos.

The hidden cost of manual chart packets

Manual packets look free because the resident “just prints the list.” The cost is after-hours chart clicking, missed overnight events, and a rounds start that waits on a printer. Time-motion work on ambulatory clinicians still frames the desk load: according to the American Medical Association, 2 hours of desk and EHR work per 1 hour of direct patient time is the widely cited allocation from that literature, which is why a 20-patient rounds list rebuilt by hand is a staffing problem before it is a model problem.

Illustrative internal cost for one morning list. $95/hr is a blended documentation hour, not a physician wage quote. Native EHR snapshots can zero several rows without a new model.

If those 220 minutes already shrank to an EHR rounds report, stop. Claude will not beat a working snapshot. If the 220 minutes are real, pick a model ID and a hold, not a chat tab.

How the automation actually works

The durable design is census → FHIR read → packet → model → clinician hold → optional file back as a draft. The EHR remains source of truth. The model never “admits” or “orders.” It proposes a summary. A workflow layer stores the encounter ID so a second overnight event updates the draft instead of creating a second packet.

Anthropic’s own model overview still says start with Claude Opus 5 for most workloads and use Claude Fable 5.1 for demanding long-horizon agentic work, or when Opus 5 at higher effort still falls short. That sentence is the product split. Opus 5 is the default packet worker. Fable 5.1 is the exception path for the 8-note, 3-PDF chart.

Fable 5.1 thinking is always on; forced tool_choice any/tool returns 400. Opus 5 still fits older tool-forcing clients more easily. If your integration still forces a get_labs tool, test Opus 5 first. If your integration sends a 200-page overnight dump and asks for a causal recap, test Fable 5.1.

Worked example

A configurable US Tech Automations workflow can read a FHIR DocumentReference for the last discharge summary and overnight notes (HL7 FHIR R4 DocumentReference), require DocumentReference.status, type, and subject on every pull, skip the row when status is entered-in-error, and hold the packet until a named clinician accepts the draft. On a 20-chart morning list with 3 control figures — 20 encounters, 8 attached PDF pages as the median packet, and 1 unsigned draft allowed per chart — the orchestrator writes one summary file, one exception list, and zero EHR writes. Prerequisites: EHR FHIR credentials, a BAA, a uniqueness key on encounter ID, and a signer. Nothing here is a live customer result. Nothing here is medical advice.

PHI belongs in the EHR and in a covered workflow, not in a personal claude.ai thread. If you cannot name the retention mode (including Bedrock Covered Model review on Fable 5.1), do not put overnight notes in that ID.

Benchmarks: before vs after

Provider-run health scores and list prices are different instruments. Do not average them.

Fable 5.1 HealthBench Professional is 56.6% according to OpenAI’s GPT-6 Astra launch table, 56.6% length-adjusted versus 54.5% for Claude Opus 5 on that same provider-run row (footnote 11 in the table). Use it as a labeled health-communication proxy, not as a clinical trial.

Fable 5.1 HLE with tools is 65.0% according to OpenAI’s GPT-6 Astra launch table, 65.0% versus 63.6% for Claude Opus 5. That is an academic-with-tools row, useful as a “can it keep facts straight while using tools” signal for packet research, not as a rounds quality guarantee.

Opus 5 list input is $5 according to Claude API pricing, $5 per 1M input and $25 per 1M output versus $10 / $50 for Claude Fable 5.1, with Opus cache hits at $0.50 per 1M versus $0.25 on Fable 5.1.

HealthBench and HLE rows are OpenAI-run (3 Sep 2026). GDPval-AA and Briefcase are Artificial Analysis (Fable 5.1 article, 1 Sep 2026); the GDPval lead is within AA’s confidence interval. Anthropic still tells most teams to start on Opus 5.

Before versus after for the 20-chart list is a process change. Before: 220 minutes of chart clicking. After a FHIR pull plus a hold: the model is called once per encounter, the clinician reviews exceptions, the EHR still owns the note. After a chat window with no encounter ID: you still have 220 minutes, now with PHI in the wrong place.

Build vs buy vs orchestrate

Build the script when one informatics engineer owns one list. Use no-code when the only hop is a census file to Slack. Orchestrate when the FHIR read, the model, and the signer must share an encounter ID.

Pros and cons

Claude Fable 5.1

Pros

• Stronger on the labeled HealthBench Professional and HLE-with-tools Claude rows in OpenAI’s 3 September table.

• Cache reads at $0.25 per 1M make a reused service-line prompt cheap across a 20-chart list.

• Document, spreadsheet, and long-horizon agentic work is the job Anthropic assigns this ID; that matches a multi-note packet.

• 1M context / 128k max output, generally available on paid Claude, API, and clouds.

Cons

• Double Opus 5’s list I/O ($10/$50 vs $5/$25), so a short recap is an expensive way to get a small HealthBench bump.

• Forced tool_choice any/tool returns 400; thinking blocks bind to this model; editing earlier turns invalidates them.

• AWS Covered Model retention (up to 30-day review) unless EFS-eligible ZDR through 31 December 2026 — a problem for overnight notes.

• AA’s GDPval lead over Opus 5 is within confidence; do not sell it as a blowout.

Claude Opus 5

Pros

• Anthropic’s default recommendation for most workloads, including structured enterprise recaps.

• Half the list I/O ($5/$25) and a $0.50 cache read that is still cheaper than Fable on a one-shot packet with little reused prefix.

• Older tool-forcing clients are less likely to 400 than on Fable 5.1.

• 1M context / 128k max output, generally available, with a May 2026 knowledge cutoff versus Fable’s June 2026.

Cons

• Trails Fable 5.1 on the labeled HealthBench Professional (54.5% vs 56.6%) and HLE-with-tools (63.6% vs 65.0%) Claude rows.

• Cache reads at $0.50 versus $0.25, so a cache-heavy 20-chart prefix is the one place Opus 5 is the more expensive read.

• Knowledge-work Elos sit a hair behind Fable 5.1 on AA GDPval-AA and Briefcase.

• If your rubric is “multi-PDF causal recap,” starting here and never testing Fable 5.1 leaves the long-horizon ID unused.

FAQs

Should a hospitalist group pick Claude Fable 5.1 or Claude Opus 5?

Pick Opus 5 for short structured packets at half the list price. Pick Fable 5.1 when the packet is a long synthesis or when Opus 5 at high effort still misses overnight events on your own rubric.

Can the model write the progress note back to the EHR?

Not in any design this page will bless. Draft, hold, clinician sign, then the EHR’s own note workflow. Auto-write is the disqualifier in the scoring table.

Do we need Fable 5.1 if Opus 5 already drafts a usable recap?

No. Anthropic’s own guide says stay on Opus 5 until evals fall short. A 2-point HealthBench gap does not justify doubling I/O on a 200-word snapshot.

When is cache more important than list I/O?

When the same service-line prompt is reread for every patient on the list. Fable 5.1’s $0.25 cache read can beat Opus 5’s $0.50 on that shape even though Fable’s uncached I/O is higher.

What FHIR object should the packet hang off?

A DocumentReference (and the documents it points at), plus an encounter ID you control. A generic object.status token is not a substitute. Confirm the resource in your EHR’s FHIR spec before you subscribe.

How should we pilot without touching the legal chart?

Run 15 mornings on 20 charts in hold-only mode: pull, draft, require a signer, write nothing back. Expand only when missed overnight events on the exception list trend down and PHI stays out of personal chat logs.

Key Takeaways

• Opus 5 is the default packet worker at $5/$25; Fable 5.1 is the long-horizon exception at $10/$50 with cheaper cache reads.

• Labeled HealthBench Professional and HLE-with-tools Claude rows sit slightly with Fable 5.1; they are not a license to auto-write the chart.

• Hang the workflow on FHIR DocumentReference plus an encounter ID and a clinician hold.

• CMS’s $1,109.7 billion physician-and-clinical-services line is the economic backdrop; 220 minutes of pre-rounds clicking is the operational one.

• Skip a new orchestrator when the EHR snapshot already is the packet.

The team at US Tech Automations can map a configurable census-to-packet trail with a clinician hold. Review agentic workflows after you have named the EHR FHIR resource, the Claude ID you will actually call, and the signer who owns the draft.