Is Sonnet 5.5 cheaper on the list sticker than Sonnet 5?

No. Same $2 / $10 in/out. Cache and Batch match Sonnet 5. Any "up to 30% less per task" is a vendor efficiency claim until your eval proves it.

Which lines prove the bill fell ~30%?

Tokens/task (effort), cache write vs read, Batch vs real-time, serving model after fallback, and pre-output refusal category. Not the press headline.

What happens on the API if we never configure fallbacks?

HTTP 200 with stop_reason: "refusal". No automatic Sonnet 5 retry until you set fallbacks / SDK middleware / manual retry.

Which refusal categories fall back to Sonnet 5 vs block?

Default routing retries cyber and frontier_llm on Sonnet 5. bio, reasoning_extraction, and general_harms stand as refusals (no default Sonnet 5 retry).

Are pre-output refusals always free?

No. Docs bill pre-output refusals for bio, frontier_llm, and reasoning_extraction. cyber and general_harms pre-output are not billed. Mid-stream always bills streamed tokens.

What must be on the Order Form / runbook today?

Flat sticker lines; VENDOR/UNPROVEN tag on 30%; effort/cache/Batch/serving-model/refusal meters; apps + API fallback owners; accepted-outcome blank. Soft-open invented SKUs and cloud-hidden fallback fields.

← Reality Check

Claude Sonnet 5.5 Keeps the Same $2/$10 Sticker as Sonnet 5, Claims Up to 30% Lower Cost per Task, and Is the First Sonnet with Cyber/frontier_llm Fallback to Sonnet 5 (API Fallback Opt-In; Apps Default On)

Sonnet 5.5 list $/MTok is flat vs Sonnet 5 ($2/$10; cache read $0.20; cache write $2.50/5m and $4/1h; Batch $1/$5). Tag vendor "up to 30% less per task" UNPROVEN. Chargeback on tokens/task, cache, Batch, serving model after fallback, and refusal category. Apps auto-switch default ON; API fallback opt-in. Name owners before volume agents hit claude-sonnet-5-5.

Jonathan Kvarfordt · Published September 29, 2026 · 8 min read

Why trust this analysis?

The short answer

Sonnet 5.5 sticker is flat - which Order Form / chargeback lines prove the bill fell ~30%, and who owns cyber/frontier fallback (serving model + refusal category) before we point volume agents at claude-sonnet-5-5?

Sonnet 5.5 list is flat vs Sonnet 5: $2/MTok in, $10/MTok out; cache read $0.20; cache write $2.50 (5m) / $4 (1h); Batch $1/$5; model id claude-sonnet-5-5. Anthropic "up to 30% less per task" / "30%+ faster" = VENDOR/UNPROVEN until buyer eval - not OF receipt math. Prove bill move with: tokens/task (API effort default High; apps/Claude Code Medium; levels recalibrated - re-sweep), cache write vs read, Batch vs real-time, serving model after fallback (response model + usage.iterations fallback_message), pre-output refusal category (stop_details.category). First Sonnet with cyber safeguards: cyber + frontier_llm fall back to Sonnet 5; bio + reasoning_extraction (distillation) + general_harms block with no Sonnet 5 fallback. Apps auto-switch DEFAULT ON; API fallback NOT default (configure fallbacks:"default" + beta header, SDK middleware, or manual retry). Until configured, API returns 200 + stop_reason refusal. Fallback billed at Sonnet 5 rates + cache-miss credit; pre-output refusal billed for bio/frontier_llm/reasoning_extraction, not for cyber/general_harms. Chargeback = serving model + refusal category, not PRD string alone. Name apps auto-switch owner + API fallback-config owner before volume agents. Soft-open: vendor early-tester quotes; Bedrock/Azure/Foundry may hide fallback detail; no invented OF SKU.

Decision rule

Refuse unsigned fallback ownership. Confirm flat sticker on the paper, tag vendor 30% UNPROVEN, map effort/cache/Batch/serving-model/refusal meters, name apps auto-switch and API fallback owners, and fill the accepted-outcome blank or flag Proof Gap.

Evidence

anthropic.com/claude-sonnet-5-5 (28 Sep 2026): flat $2/$10 vs Sonnet 5; up to 30% less per task / 30%+ faster testing claims; first Sonnet with cyber safeguards; claude-sonnet-5-5 on API + AWS + Google Cloud + Azure
platform.claude.com/docs/en/about-claude/pricing: Sonnet 5.5 $2/$10; cache read $0.20; cache write $2.50 (5m) / $4 (1h); Batch $1/$5
platform.claude.com/docs/en/models/sonnet-5-5/whats-new-sonnet-5-5: effort default High on Claude API; levels recalibrated; cyber+frontier_llm default-route to Sonnet 5; bio/reasoning_extraction/general_harms no default retry; availability IDs
support.claude.com/en/articles/17161993-why-claude-switched-models-in-your-conversation-with-sonnet-5-5: apps auto-switch default ON; API not default; fallback billed at Sonnet 5 + cache-miss credit; pre-output billing by classifier
platform.claude.com/docs/en/build-with-claude/refusals-and-fallback: stop_reason refusal; category billing table; fallbacks default beta; iterations receipt; Bedrock/Foundry client-side only

Operator action

Today: lock flat sticker lines on the OF/runbook, tag vendor 30% UNPROVEN, name apps auto-switch + API fallback owners, wire chargeback on serving model + refusal category, re-sweep effort after recalibration, and fill the accepted-outcome blank before pointing volume agents at claude-sonnet-5-5.

Last reviewed September 29, 2026

Claude Sonnet 5.5 keeps the same $2/$10 sticker as Sonnet 5, claims up to 30% lower cost per task, and is the first Sonnet with cyber/frontier_llm fallback to Sonnet 5 (API fallback opt-in; apps default on)

Claude Sonnet 5.5 list dollars are flat vs Sonnet 5: $2 / MTok input, $10 / MTok output. Cache read $0.20 / MTok. Cache write $2.50 / MTok (5-minute) / $4 / MTok (1-hour). Batch $1 / $5. Model id: `claude-sonnet-5-5` (Anthropic Sonnet 5.5, 28 Sep 2026; Claude Platform pricing).

The argument

How this reality check breaks down

A map of the sections ahead, in the order the case is made. The sourced charts are in the research library.

Anthropic tags "up to 30% less per task" and "30%+ faster" as testing claims. Tag those VENDOR / UNPROVEN until your buyer eval reproduces them. Flat sticker means a bill drop is an efficiency claim, not a lower list rate. Do not put "30%" on the Order Form as receipt math. Sibling decoder on unit vs bill after a token cut: GPT-6 / Opus 5.5 token-cut Order Form.

Three-column Order Form strip - flat Sonnet 5.5 list $/MTok vs VENDOR/UNPROVEN up-to-30%-less-per-task claim vs chargeback keys (serving model, effort, cache, refusal category) — Flat list $/MTok. Vendor "up to 30%" is not an Order Form receipt. Chargeback keys prove the bill moved.

Chargeback lines that prove the bill moved

Reality Check

Claude Sonnet 5.5 Keeps the Same $2/$10 Sticker as Sonnet 5, Claims Up to 30% Lower Cost per Task, and Is the First Sonnet with Cyber/frontier_llm Fallback to Sonnet 5 (API Fallback Opt-In; Apps Default On)

Source-cited charts live in the research library.

Pull these meters. They move the invoice when the sticker does not:

Tokens / task. Effort default on the Claude API is High. Claude apps / Claude Code default Medium. Effort levels are recalibrated vs Sonnet 5. Re-sweep; do not carry the old setting over (What's new in Sonnet 5.5).
Cache write vs cache read. Write is 1.25x / 2x base input (5m / 1h). Read is 0.1x. Miss ratio on fallback paths matters.
Batch vs real-time. Batch is 50% of list ($1 / $5). Real-time stays at $2 / $10.
Serving model after fallback. Response `model` + `usage.iterations` (incl. `fallback_message`) name who actually ran.
Pre-output refusal category. `stop_details.category` drives whether a pre-output refusal is billed (Refusals and fallback).

First Sonnet with cyber safeguards

Sonnet 5.5 is the first Sonnet shipped with cyber safeguards and fallbacks like Anthropic's most capable models (launch; Help Center).

Classifier path (cite-only):

Request hits `claude-sonnet-5-5`, then classifier.
Serve Sonnet 5.5 on routine / allowed.
Fall back to Sonnet 5 when category is `cyber` or `frontier_llm` (recommended default routing on API with `fallbacks: "default"`).
Block / refuse (no Sonnet 5 fallback) when category is `bio`, `reasoning_extraction` (distillation), or `general_harms`.

Fallback flow from claude-sonnet-5-5 through classifier to serve, cyber/frontier fall back to Sonnet 5, or block - with apps vs API ownership and refusal billing callouts — Cyber / frontier_llm → Sonnet 5. Bio / distillation / general_harms → block. Apps auto-switch default ON. API fallback opt-in.

Apps vs API ownership (name it before volume)

Claude apps (web, Mobile, Desktop, Claude Code, and related surfaces): automatic model switching is DEFAULT ON. User / admin can turn it off in Settings → Capabilities (Claude Code: Config → MODEL & OUTPUT). After a switch, the chat stays on Sonnet 5 until someone picks Sonnet 5.5 again (Help Center).

Claude API: fallback is NOT default. You must configure `fallbacks: "default"` (or a named list) plus the `server-side-fallback-2026-07-01` beta header, or use SDK middleware / manual retry with fallback credit. Until configured, the API returns HTTP 200 with `stop_reason: "refusal"` and empty content. Server-side `fallbacks` is Claude API beta only; Bedrock / Google Cloud / Microsoft Foundry need client-side fallback (SDK middleware or your retry) (Refusals and fallback).

Name the owner of (1) apps auto-switch policy and (2) API fallback config before pointing volume agents at `claude-sonnet-5-5`.

Billing: serving model + refusal category, not the PRD string

Each attempt bills at the rates of the model that ran it. Fallback to Sonnet 5 bills at Sonnet 5 rates. Anthropic provides a cache-miss credit on the fallback path so you do not pay to re-cache the conversation twice.
Pre-output refusal billing by category (as of Sep 2026 docs): billed for `bio`, `frontier_llm`, `reasoning_extraction`. Not billed for `cyber`, `general_harms` (or null category). Mid-stream refusals bill input + streamed output.
Chargeback receipt = serving model (`model` / `fallback_message` in `usage.iterations`) + refusal category (`stop_details.category`). The PRD string `claude-sonnet-5-5` alone is not the bill line.

Soft opens (do not invent)

Slack / Zendesk / Atlassian early-tester quotes on the launch page = vendor only. Not your Order Form proof.
Bedrock / Azure / Foundry may hide fallback detail relative to first-party Claude API. Soft-open exact cloud invoice fields for serving model / refusal category.
No invented Order Form SKU / pack / commitment for "30% task savings."

Availability

`claude-sonnet-5-5` on Claude API; `anthropic.claude-sonnet-5-5` on Amazon Bedrock; `claude-sonnet-5-5` on Google Cloud and Microsoft Foundry (What's new).

Order Form / chargeback checklist

List sticker: $2 / $10; cache read $0.20; cache write $2.50 (5m) / $4 (1h); Batch $1 / $5. Flat vs Sonnet 5.
Vendor "up to 30%" tagged VENDOR / UNPROVEN. Not OF math.
Effort default map: API High vs apps Medium; re-sweep after recalibration.
Cache write vs read and Batch vs real-time on the chargeback sheet.
Serving model after fallback (iterations / response model) named as a meter.
Refusal category billing rules on the paper (bio / frontier_llm / reasoning_extraction billed pre-output; cyber / general_harms not).
Apps auto-switch owner (default ON) and API fallback config owner (not default).
Accepted-outcome blank. If missing, flag Proof Gap.

Proof Gap

Flat sticker ≠ lower bill by itself.
Vendor efficiency ("up to 30%") ≠ accepted outcome.
Meter moves with effort, cache, Batch, serving model, and refusal category - not the press "30%."
Unit ≠ bill without those chargeback lines and named fallback owners.

Who reviews which line (Eight-Seat lite)

RevOps / Finance: list vs task-cost claim; chargeback keys; Batch vs real-time; accepted-outcome blank.
GTM eng / platform: API `fallbacks` / SDK middleware; effort defaults; cache policy; Bedrock / Foundry client-side path.
Enablement: train apps auto-switch notices; gate volume agents until API fallback is configured.
Security / compliance: cyber + frontier_llm fallback to Sonnet 5; bio / distillation / general_harms block; named policy owner for auto-switch OFF cases.

Sources

Anthropic, Introducing Claude Sonnet 5.5 (28 Sep 2026): anthropic.com/claude-sonnet-5-5
Claude Platform docs, Pricing: platform.claude.com/docs/en/about-claude/pricing
Claude Platform docs, What's new in Claude Sonnet 5.5: platform.claude.com/docs/en/models/sonnet-5-5/whats-new-sonnet-5-5
Claude Help Center, Why Claude switched models in your conversation with Sonnet 5.5: support.claude.com/en/articles/17161993-why-claude-switched-models-in-your-conversation-with-sonnet-5-5
Claude Platform docs, Refusals and fallback: platform.claude.com/docs/en/build-with-claude/refusals-and-fallback
Anthropic pricing hub: anthropic.com/pricing

Soft CTA

If you want the next Order Form decoder before the next flat-sticker efficiency claim, subscribe

Take it to the room

The short list from this piece

Pulled from the argument above, written so you can read it out in a pipeline or board review.

Frequently asked questions

Is Sonnet 5.5 cheaper on the list sticker than Sonnet 5?: No. Same $2 / $10 in/out. Cache and Batch match Sonnet 5. Any "up to 30% less per task" is a vendor efficiency claim until your eval proves it.
Which lines prove the bill fell ~30%?: Tokens/task (effort), cache write vs read, Batch vs real-time, serving model after fallback, and pre-output refusal category. Not the press headline.
Who owns cyber / frontier fallback?: Name the apps auto-switch owner (default ON) and the API fallback-config owner (opt-in). Chargeback on serving model + refusal category.
What happens on the API if we never configure fallbacks?: HTTP 200 with stop_reason: "refusal". No automatic Sonnet 5 retry until you set fallbacks / SDK middleware / manual retry.
Which refusal categories fall back to Sonnet 5 vs block?: Default routing retries cyber and frontier_llm on Sonnet 5. bio, reasoning_extraction, and general_harms stand as refusals (no default Sonnet 5 retry).
Are pre-output refusals always free?: No. Docs bill pre-output refusals for bio, frontier_llm, and reasoning_extraction. cyber and general_harms pre-output are not billed. Mid-stream always bills streamed tokens.
What must be on the Order Form / runbook today?: Flat sticker lines; VENDOR/UNPROVEN tag on 30%; effort/cache/Batch/serving-model/refusal meters; apps + API fallback owners; accepted-outcome blank. Soft-open invented SKUs and cloud-hidden fallback fields.

Share this page

Posting to Instagram or TikTok? Copy the link, it carries the title, summary and share image.

Get the next Reality Check before you sign the order form.

Keep reading

Reality Check

AI Review Time Is a Line Item. Price It and Name the Owner.

On TextQL's Argo-Bench the best AI model fully solved 34.8% of 210 data tasks. Put review time and a named owner into your AI business case.

Reality Check

Call Center Headcount Is Falling. Can You Prove AI Did It?

Revelio Labs says global call-center headcount is 5.3% below its Dec 2023 peak. Revelio: "The decline is surely multifaceted." How to test your own org.