Starter: compare model costs for a described task — maps task shape to the cheapest adequate seat and shows the price spread
SkillScore95Maintained
On-demandOn-demand: invoked explicitly — never occupies your agent's context until called.
◆AI Engineer pack
Last verified 9/9/2026
What it changes
Passed 21 of 22 test cases with the skill, and 6 of 22 without it.
Without skill27%
With skill95%
Checks it passed that the model alone missed
Recommends Claude Opus 5 for the security review
Does not recommend Fable 5.1 for the security audit
Recommends Claude Sonnet 5 for routine coding tasks
and 17 more in the full report
15 fixed6 passed either way1 still failing0 broke
Pass-rate lift
+68 pp
Tokens
+14%
Agent turns
0%
Response time
-17%
Model gemini-3.6-flashMeasured 9/9/2026v221 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 15 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 15
Prompt
We need an in-depth security analysis and vulnerability review of our authentication module code before deployment. Evaluate the workload and suggest the appropriate model seat and estimated cost.
What was checked, without and with the skill
Checked
Without skill
With skill
Recommends Claude Opus 5 for the security review
Missed
Passed
Does not recommend Fable 5.1 for the security audit
Missed
Passed
case-04 · gemini-3.6-flash · 9/9/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md22 linessingle file
Model Cost Comparison (Starter Pack)
Answer "which model should I use for this, and what will it cost?" with numbers instead of vibes.
When to use
The user describes a task (bulk refactor, deep review, quick lookup, long-context analysis) and wants the cheapest seat that is still adequate.
Steps
Classify the task. Bucket it: mechanical (rename, format), standard coding, hard reasoning (architecture, security review), long-context (>200K tokens input), or web research.
Estimate volume. Rough input/output token estimate from the described scope (files touched × average size; state the assumption).
Price the roster. Using the cost table in CLAUDE.md ($/MTok input/output), compute the estimated cost for each plausible seat: Claude Opus 5 ($5/$25), Claude Sonnet 5 ($2/$10), Fable 5.1 ($10/$50, 1M context, explicit-only), Codex GPT-5.6 Sol ($4/$20), Terra ($2/$12), Luna ($0.20/$1.20), GPT-6 Astra ($10/$50, explicit-only), Perplexity Sonar Pro ($3/$15), and the included-cost seats (agy, copilot, ollama, cursor-agent) at $0. For Astra requests above 272K input tokens, apply 2x input and 1.5x output pricing to the full request.
Recommend one seat. Pick the cheapest adequate option and defend it in two sentences. Mechanical work goes to included or budget seats; hard reasoning justifies Opus 5 at high effort; only a bounded judgment-class call (ambiguous architecture, API design, product tradeoffs) justifies Fable 5.1 at twice the Opus price.
Check risk surfaces. Regardless of the classification, escalate specifically to Opus 5 when the task touches API or schema contracts, security-sensitive code or CI configuration, release artifacts, user-facing UI, a new module, or a breaking change. Fable 5.1 remains limited to bounded judgment-class calls and is never the security-audit seat. Astra is also explicit-only and does not provide independence from GPT-5.6. Cheap-seat agreement never settles a judgment-class decision.
Show the spread. A three-row table: recommended seat, one cheaper-but-riskier option, one premium option, each with estimated dollars for this task.
Guardrails
Never recommend Fable 5.1, Astra, or fast-mode Opus by default; these premium seats are opt-in only.
Never recommend Fable 5.1 for security audits; its safety classifiers can refuse offensive-security phrasing. Security review goes to Opus 5 (see skills/blocks/fable5-prompting.md).
If the estimate exceeds $1, say so explicitly before any dispatch happens.