Fable 5 Pricing on Cursor: Every Tier Explained (Max to Low)

Cluster: CursorBench 3.2 cost analysis · When to escalate Composer → Fable · Composer 2.5 baseline

What is Fable 5 pricing on Cursor?

Fable 5 is Anthropic's agentic model family in Cursor's model picker, offered at five effort tiers (Max, Extra High, High, Medium, Low). Each tier trades task success rate against cost per agent task, tokens burned, and agent steps until the task closes.

Who it is for: Anyone comparing Fable 5 tiers in Cursor who sees a leaderboard headline and needs unit economics — a founder watching burn rate, a student on a tight budget, an operator setting team defaults.

What you will learn: the CursorBench 3.2 ladder from Max to Low, what each tier buys in score and steps, and a default rule: specialty tier for must-not-fail work, budget tier for volume.

Vendor launches sell peak score. People paying the bill need cost per agent task, tokens, and steps — especially when many sessions share one picker default.

Anthropic returned Fable 5 to public Cursor model pickers in 2026. Cursor publishes side-by-side numbers on CursorBench 3.2. The table below is benchmarked numbers (from CursorBench 3.2). They price tasks from published per-million-token rates on ambiguous, multi-file agent work. Your real bill will differ by prompt, context size, and retries.

Why tier choice matters more than peak score

A model picker that only shows rank hides four costs:

Cost type	What breaks if you ignore it
Dollars per task	You burn budget on tasks that a cheaper tier would pass
Tokens per task	Context windows fill faster; long sessions stall
Steps per task	Latency and review fatigue rise even when the task eventually passes
Variance	Cursor notes small score gaps may not be statistically meaningful

Fable 5 Max can be rational for one class of work. It is a poor default if you run dozens of agent tasks per week.

The five tiers on CursorBench 3.2

Benchmarked numbers (from CursorBench 3.2):

Tier	Score	Cost / task	Tokens / task	Steps / task
Fable 5 Max	70.5%	$17.32	103,525	72
Fable 5 Extra High	68.4%	$11.73	64,971	56
Fable 5 High	66.5%	$8.77	43,747	48
Fable 5 Medium	65.2%	$6.80	30,366	41
Fable 5 Low	62.1%	$4.46	18,182	31

Score spread: Max to Low is 8.4 percentage points (70.5% vs 62.1%).

Cost spread: Max is roughly 3.9× Low on cost per task ($17.32 vs $4.46) for that score gap.

Steps spread: Max takes 72 steps; Low takes 31. Fewer agent turns can mean shorter wall-clock time on step-heavy tasks.

Tier-by-tier read

Fable 5 Max

Best for: one-off tasks where failure is expensive (production incident, security-sensitive remediation, executive-visible deliverable).

Tradeoff: $17.32 per benchmark task and 72 steps. Score per dollar on this table is among the worst rows (~4.1 score per USD when derived from public numbers).

Fable 5 Extra High and High

Best for: high-stakes work with a cost ceiling below Max. Extra High lands 68.4% at $11.73; High lands 66.5% at $8.77.

Tradeoff: You give up 2.1 to 4.0 points versus Max. You still pay premium prices versus Composer 2.5 on the same benchmark.

Fable 5 Medium

Best for: "better than budget tier" without Max bills. 65.2% score at $6.80 and 41 steps.

Tradeoff: Still ~15× the benchmark cost of Composer 2.5 ($0.44 per task on the same table) for roughly 9.1 points more score.

Fable 5 Low

Best for: trying Fable economics without Max pricing. 62.1% at $4.46 with the lowest step count in the Fable family (31).

Tradeoff: Lowest Fable score on the table. On derived score per step, Low is competitive within the Fable line (~2.0 score per step on public rows).

How Fable tiers compare to Composer 2.5

On CursorBench 3.2, Composer 2.5 posts 56.1% at $0.44, 14,286 tokens, 33 steps.

Comparison	Takeaway
Fable 5 Low vs Composer 2.5	+6.0 points score for ~10× cost in the benchmark
Fable 5 Max vs Composer 2.5	+14.4 points for ~39× cost
Steps	Composer 2.5 (33) matches Fable Low (31); Max more than doubles step count

Note: Grok 4.5 High ranks 66.7% at $1.51 and 33 steps on the same table — between Fable High and Medium on score at a fraction of Fable cost. Cursor flags a training-data caveat on Grok rows (see open models).

For routine program volume, Composer 2.5 remains the Pareto corner on this benchmark. Fable tiers are escalation tools, not replacements for a cost-efficient default. See when to escalate.

Where this sits in the benchmark landscape

CursorBench measures realistic Cursor agent sessions (ambiguous, multi-file). Vendor launch posts and SWE-bench rows measure different harnesses. A tier that wins on CursorBench may not rank the same on SWE-bench or HumanEval-style sets.

Use Fable tier tables for Cursor session economics. Use SWE-bench for repository patch success comparisons. Do not merge the two without labeling the source.

Limitations

CursorBench rows have variance; treat small score gaps as directional.
Cost per task uses Cursor's pricing model at publish time; API list prices change.
Tiers reflect effort settings, not separate model weights in every case. Read the picker labels when Cursor updates naming.

Reader action

Open CursorBench 3.2 and confirm the five Fable rows still match.
Set Composer 2.5 (or your budget default) for routine agent work.
Pre-pick one escalation tier (often Medium or High) for tasks with a defined cost cap.
Reserve Max for a written trigger list (incident, security, irreversible migration).
Read CursorBench cost analysis for score-per-dollar tables across all models.