ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMEARTICLESWHICH CLAUDE MODEL TO USE, AND WHAT EACH COSTS
ARTICLES · COMPARISON

Which Claude Model to Use, and What Each Costs

Anthropic names its own default. The four current models with prices, context windows, knowledge cutoffs — and the one that could retire first.

READ4 min
WORDS758
SECTIONS5
TYPECOMPARISON
CHECKED16 SEP 26

Four models, and Anthropic tells you which one to start with. Its own documentation says to begin with Claude Opus 5.5 for most work and move to Fable 5.1 only when your evaluations say Opus is not enough. That is an unusually direct answer from a vendor, and it is worth taking at face value.Anthropic, Models overview, read at source 23 Sep 2026: “start with Claude Opus 5.5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.” Opus 5.5 took Opus 5’s place in that advice when it was released on 22 September 2026; Claude Opus 5.5 vs Opus 5 covers what changed.Anthropic, Claude Opus 5.5, read at source 23 Sep 2026: “Released September 22, 2026.”First-hand: until 23 Sep 2026 this paragraph said to begin with Claude Opus 5, quoting the same documentation as read on 16 Sep 2026.

TL;DR — THE SHORT VERSION

Pick by the job, then check the retirement date before you build on it.

  • Opus 5.5 is the default — $4 in, $20 out per million tokens, since 22 Sep 2026.Anthropic, Models overview, read at source 23 Sep 2026: Claude Opus 5.5 “$4 / input MTok, $20 / output MTok”.First-hand: until 23 Sep 2026 this bullet read “Opus 5 is the default — $5 in, $25 out per million tokens.”
  • Fable 5.1 is the expensive one at $10 and $50, for long-horizon reasoning.Anthropic, Pricing and Models overview, read at source 16 Sep 2026: Opus 5 “$5 / MTok”, Fable 5.1 “$10 / MTok”, Sonnet 5 “$2 / MTok”.
  • Sonnet 5 is $2 and $10, and its introductory price became the permanent one.Anthropic, Pricing and Models overview, read at source 16 Sep 2026: Opus 5 “$5 / MTok”, Fable 5.1 “$10 / MTok”, Sonnet 5 “$2 / MTok”.
  • Haiku 4.5 has the earliest retirement floor — not sooner than 15 October 2026. It is still Active, and no retirement notice has been given.Anthropic, Model deprecations, read at source 22 Sep 2026: claude-haiku-4-5-20251001 “Active”, deprecated “N/A”, tentative retirement date “Not sooner than October 15, 2026”.
  • Three of the four hold a million tokens. Haiku holds 200,000.Anthropic, Pricing and Models overview, read at source 16 Sep 2026: Opus 5 “$5 / MTok” input; Fable 5.1 “$10 / MTok”; Sonnet 5 “$2 / MTok”; Haiku 4.5 context window “200K tokens”.
◈ PRICES ON THIS PAGE

Every price here was correct when this site last checked it, between 16 Sep and 23 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: Anthropic pricing.

The lineup

Anthropic’s published figures, read at source 16 Sep 2026; the Opus 5.5 column 23 Sep 2026. Prices are per million tokens
Fable 5.1Opus 5.5Sonnet 5Haiku 4.5
Input$10$4$2$1
Output$50$20$10$5
Context window1M1M1M200K
Max output128K128K128K64K
Reliable knowledge cutoffJun 2026Jun 2026Jan 2026Feb 2025
Training data cutoffJun 2026Jun 2026Jan 2026Jul 2025
Retires no sooner than1 Sep 202722 Sep 202730 Jun 202715 Oct 2026

Anthropic, Models overview and Pricing, both read at source 16 Sep 2026. The two cutoff rows re-read 22 Sep 2026: Anthropic defines the reliable cutoff as “The date through which the model’s knowledge is most extensive and reliable”, and the training data cutoff as “the broader range of data used”.

Anthropic, Claude Opus 5.5, read at source 23 Sep 2026: “Input pricing: $4 / MTok”, “Output pricing: $20 / MTok”, reliable knowledge and training data cutoff “Jun 2026”, retirement “Not sooner than September 22, 2027”. Claude Opus 5 page, same day: “Although Claude Opus 5 is still available, you should consider migrating to Claude Opus 5.5 for improved performance.” First-hand: until 23 Sep 2026 this column showed Claude Opus 5 — $5 in, $25 out, both cutoffs May 2026, retiring no sooner than 24 Jul 2027.

First-hand: until 22 Sep 2026 the table had one row, labelled “Knowledge cutoff”, showing Feb 2025 for Haiku 4.5. That is Anthropic’s reliable knowledge cutoff; Haiku 4.5’s training data runs to July 2025. Anthropic, How up-to-date is Claude’s training data?, read at source 22 Sep 2026: “Claude Haiku 4.5 was trained on data up until July 2025.”

Two things about Fable 5.1 the table cannot show. It is the same model as Claude Mythos 5.1; the difference is the safeguards, and Mythos is available only through Anthropic’s trusted access programs. And its effort setting changes by product: Fable 5.1 defaults to High effort in Claude Code and to Medium in Claude Cowork and on Claude.ai. Anthropic also estimates it costs about a quarter less than Fable 5 for typical work, because cache reads got cheaper.Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1, read at source 22 Sep 2026: “Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards.”; “Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads, wherever usage is billed by token. This is because we’re reducing our pricing on cache reads”; “Fable 5.1 defaults to High effort in Claude Code, and to Medium in Claude Cowork and on Claude.ai.”

TAKEAWAY

The retirement row is the one to read before the price row. A model that costs less but could leave within months is not cheaper — it is a migration you have not scheduled.

The date worth putting in a calendar

Haiku 4.5 is the cheapest model here and has the earliest retirement floor: Anthropic gives its retirement as not sooner than 15 October 2026, and no other model in the table carries a 2026 date. That is a floor, not a date. As of 22 September Haiku 4.5 is still listed as Active with no deprecation, and Anthropic promises “at least 60 days’ notice before model retirement”, so it cannot go on 15 October.Anthropic, Model deprecations, read at source 22 Sep 2026: “Anthropic notifies customers with active deployments for models with upcoming retirements, providing at least 60 days’ notice before model retirement for publicly released models.” claude-haiku-4-5-20251001: “Active”, deprecated “N/A”.Reasoning, 22 Sep 2026 — sixty days from a notice given today is 21 November 2026, so that is the earliest Haiku 4.5 could now retire, and later for every day without notice.First-hand: until 22 Sep 2026 this paragraph called Haiku 4.5 “the first to go” and 15 October “weeks rather than years”, and the takeaway above spoke of a model that “leaves in four weeks” — reading a floor as a date.Anthropic, Models overview, read at source 16 Sep 2026: Haiku 4.5 retirement “Not sooner than October 15, 2026”, against “Not sooner than June 30, 2027” for Sonnet 5 and later dates above it.

RETIRES NO SOONER THAN
The four models in the order Anthropic says they may leave, read at source 16 Sep 2026, Opus 5.5 on 23 Sep 2026. Each date is a floor, not a schedule.
OCT 2026Haiku 4.5Not sooner than 15 October 2026. The cheapest model, and the only 2026 date.
JUN 2027Sonnet 5Not sooner than 30 June 2027.
SEP 2027Fable 5.1Not sooner than 1 September 2027.
SEP 2027Opus 5.5Not sooner than 22 September 2027.
Anthropic, Models overview, read at source 16 Sep 2026: Haiku 4.5 retirement “Not sooner than October 15, 2026”, against “Not sooner than June 30, 2027” for Sonnet 5 and later dates above it. Anthropic, Claude Opus 5.5, read at source 23 Sep 2026: retirement “Not sooner than September 22, 2027”.

Its reliable knowledge cutoff is also the oldest by a wide margin — February 2025, against June 2026 for the top two — though its training data runs to July 2025. For anything that depends on recent facts, that gap matters more than the price does.First-hand: until 23 Sep 2026 this said “against May and June 2026 for the top two”, when Opus 5, with a May 2026 cutoff, was in the table.

Caching, and the multiplier that surprises people

Anthropic publishes caching as multipliers on the base input price rather than as separate numbers, which makes the arithmetic easy to check.Anthropic, Pricing, read at source 16 Sep 2026: “Prompt caching uses the following pricing multipliers relative to base input token rates”.

Prompt caching, as multipliers of the base input price
OperationMultiplier
Cache read (hit)0.1×
Cache write, 5 minutes1.25×
Cache write, 1 hour
Batch API50% off both directions

Anthropic, Pricing, read at source 16 Sep 2026: “A cache hit costs 10% of the standard input price”, and the Batch API offers “a 50% discount on both input and output tokens.” Anthropic also notes these “stack with other pricing modifiers”.

Two exceptions are documented and easy to miss: on Fable 5.1 a cache hit is 0.025× the base price, and on Opus 5.5, released 22 Sep 2026, it is 0.05×, rather than 0.1×. In dollars that is $0.25 and $0.20 per million tokens, against $0.50 on Opus 5 and $0.10 on Haiku 4.5: the most expensive model has the lowest multiplier, not the cheapest cache read.Anthropic, Pricing, read at source 23 Sep 2026: “On Claude Fable 5.1 and Claude Mythos 5.1, a cache hit costs 2.5% of the standard input price ($0.25 USD per million tokens). On Claude Opus 5.5, a cache hit costs 5% of the standard input price ($0.20 USD per million tokens).” Cache hits: Claude Opus 5 “$0.50 / MTok”, Claude Haiku 4.5 “$0.10 / MTok”.First-hand: until 23 Sep 2026 this said Fable 5.1’s rate “makes the most expensive model the cheapest to re-read from cache”. In dollars it did not: Haiku 4.5’s cache read was $0.10 against Fable 5.1’s $0.25 then too.

PROMPT CACHING · MULTIPLE OF BASE INPUT PRICE
Anthropic’s multipliers drawn to scale, read at source 16 Sep 2026, the Opus 5.5 rate on 23 Sep 2026. The Fable 5.1 and Opus 5.5 cache hits are the exceptions the table leaves out.
Cache write, 1 hour
Cache write, 5 minutes1.25×
Base input price
Cache read (hit)0.1×
Cache read (hit) on Opus 5.50.05×
Cache read (hit) on Fable 5.10.025×
Anthropic, Pricing, read at source 16 Sep 2026: “Prompt caching uses the following pricing multipliers relative to base input token rates”, and “A cache hit costs 10% of the standard input price”. Read at source 23 Sep 2026: “On Claude Opus 5.5, a cache hit costs 5% of the standard input price ($0.20 USD per million tokens).”

The same lever exists on OpenAI, at a similar ratio — what the OpenAI API costs covers it.

A price rise that did not happen

Sonnet 5 launched at $2 and $10 as introductory pricing due to end on 31 August 2026, with a rise to $3 and $15 scheduled behind it. Anthropic’s pricing page now records that the introductory rate became the standard one and the increase was dropped.Anthropic, Pricing, read at source 16 Sep 2026: the $2/$10 pricing “announced at launch as introductory pricing through August 31, 2026, is now the standard price.”

Worth noting because the opposite is more common. Google has a doubling scheduled for 1 January 2027 on its cheapest model; the comparison is on Gemini vs Grok prices.

What this page could not verify

  • Whether any model is better at your task. Nothing here was tested; the guidance is Anthropic’s own.
  • Real cost per job. A cheaper model that needs more attempts is not cheaper.
  • Actual retirement dates. “Not sooner than” is a floor, not a schedule.
  • Models outside this table. Anthropic’s pricing page references others; only the four current ones are covered here.
SOURCES

Anthropic, Models overview and Pricing, both read at source on 16 September 2026 and again on 23 September 2026 for Claude Opus 5.5, with the Claude Opus 5.5 and Claude Opus 5 model pages.

Model lineups change without notice. Re-read before committing.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE