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

1. Classify the task. Bucket it: mechanical (rename, format), standard coding, hard reasoning (architecture, security review), long-context (>200K tokens input), or web research.
2. Estimate volume. Rough input/output token estimate from the described scope (files touched × average size; state the assumption).
3. 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 (4/20 USD), Claude Opus 5 (5/25 USD), Claude Sonnet 5 (2/10 USD), Fable 5.1 (10/50 USD, 1M context, explicit-only), Codex GPT-5.6 Sol (4/20 USD), Terra (2/12 USD), Luna (0.20/1.20 USD), GPT-6 Astra (10/50 USD, explicit-only), Perplexity Sonar Pro (3/15 USD), and the included-cost seats (agy, copilot, ollama, cursor-agent) at 0 USD. For Astra requests above 272K input tokens, apply 2x input and 1.5x output pricing to the full request.
4. 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 the current Opus default at high effort; only a bounded judgment-class call (ambiguous architecture, API design, product tradeoffs) justifies Fable 5.1 at 10/50 USD per MTok.
5. Check risk surfaces. Regardless of the classification, escalate specifically to the current Opus default—Opus 5.5 on Claude Code v2.1.280+, otherwise 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.
6. 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 USD, say so explicitly before any dispatch happens.