When to Escalate from Composer 2.5 to Fable 5: A Decision Tree

Cluster: CursorBench 3.2 analysis · Fable 5 tiers · Composer 2.5 baseline

When should you escalate from Composer 2.5 to Fable 5?

Escalate when failure cost on a Cursor agent task exceeds the premium Fable charges in dollars, tokens, and steps on CursorBench 3.2. Stay on Composer 2.5 for routine program work where 56.1% score at $0.44 per task is enough.

Who it is for: Anyone paying for premium Cursor models who wants escalation rules before expensive tiers become habit — founders on a budget, students on a thesis, operators governing a shared picker.

What you will learn: a gated decision tree for escalation, tier pick rules (Low through Max), and anti-patterns that turn premium models into unbudgeted default.

Composer 2.5 and Fable 5 answer different questions on CursorBench 3.2. Composer wins score per dollar and keeps 33 steps per task on the public table. Fable 5 Max wins raw score (70.5%) at $17.32 and 72 steps. Grok 4.5 High sits between them on score (66.7%, $1.51, 33 steps) with a Cursor-flagged training-data caveat.

I use Composer 2.5 as my program default because predictable rule compliance and a tight bootstrap matter for the delivery work I run daily (baseline model policy). CursorBench gives that habit a cost line. Fable is an escalation lane, not a replacement.

Why escalation needs gates

Without gates, every hard prompt drifts to the top model. At program scale that produces:

Budget bleed (~17 USD benchmark tasks on routine work)
Latency (72-step runs across a delivery team)
Review fatigue (long agent traces managers must audit)

Gates force you to name what failure costs the program before you pay for peak score.

Benchmarked anchors

Benchmarked numbers (from CursorBench 3.2):

Model	Score	Cost / task	Steps / task
Composer 2.5	56.1%	$0.44	33
Grok 4.5 High*	66.7%	$1.51	33
Fable 5 Low	62.1%	$4.46	31
Fable 5 Medium	65.2%	$6.80	41
Fable 5 High	66.5%	$8.77	48
Fable 5 Max	70.5%	$17.32	72

* Cursor notes Grok 4.5 may have an advantage from Cursor codebase in training data; impact unclear (evals disclaimer).

Gap to remember: Max buys ~14.4 points over Composer for ~39× benchmark cost.

Decision tree

Answer in order. Stop at the first yes.

#	Question	If yes →
1	Is the outcome easily reversible (rollback, discard draft, re-run)?	Stay on Composer 2.5
2	Is failure client-facing, production-facing, or compliance-sensitive?	Escalate to Fable 5 High or Max
3	Did Composer fail twice on the same task type?	Escalate one tier (Low → Medium)
4	Does the task need multi-file planning across unfamiliar material?	Fable 5 Medium
5	Is the task exploratory with a strict cost cap?	Grok 4.5 Low or GPT-5.5 Medium (work-mode map)
6	Otherwise	Composer 2.5

Tier selection after you decide to escalate

Escalation level	Pick	Benchmark cue
Light	Fable 5 Low	+6 points over Composer, 31 steps, $4.46
Standard	Fable 5 Medium	65.2%, 41 steps, under $7
Heavy	Fable 5 High	66.5%, still below Max cost
Critical	Fable 5 Max	70.5%, accept ~17 USD and 72 steps
Budget score bump	Grok 4.5 High*	66.7% near Fable High cost profile; read training caveat

Full tier ladder: Fable 5 pricing explained.

Anti-patterns

Anti-pattern	Fix
Escalate because the task "feels hard"	Require two Composer failures or a named risk class
Stay on Max after the incident clears	Drop back to Composer for follow-up commits
Escalate for greenfield scaffolding without fixing context	Fix bootstrap and memory policy first (baseline post)
Compare only SWE-bench rank	Use CursorBench for session cost (benchmark lenses)

Example implementation (how I run it)

Example implementation — my stack:

Default model: Composer 2.5 in Cursor for routine agent sessions.
Escalation list in a short policy note (docs/ai-model-escalation.md or equivalent): client-facing deliverables, production, security, regulated data.
After a Fable run, log task type, tier, and outcome in the session footer or CSV if you measure harness ROI (harness measurement).

Path A (any chat tool): Write three escalation triggers on a sticky note. Only change models when a trigger matches.

Limitations

Triggers are operational policy, not a guarantee of task success.
CursorBench costs are modeled; your subscription and usage caps differ.
Grok 4.5 rows carry Cursor's training-data caveat; do not treat as a clean apples-to-apples row.
Open models may fit budget experiments better than Fable Low (open models).

Reader action

Copy the six-row decision table into your program's AI governance docs.
Run ten tasks on Composer 2.5 without opening the Fable picker.
Log failures by type; escalate only on rule 2 or 3.
Review bills weekly; if Max usage exceeds 5% of tasks, tighten gates.