Modeled snapshot · 2026-07-20

Stay, switch, or route the work?

Start with your current workflow. Leave with an action only when direct, dated evidence supports it—not when price, modeled headroom, or a model benchmark merely looks persuasive.

Evidence rule: native-tool quality, productivity, and governance remain unranked unless a same-scope receipt supports the comparison.

Starting state · 01

Terminal incumbent

You already work in a CLI agent and need to decide whether changing tools is justified.

Action
Evidence limited · unranked
Evidence receipt
OpenAI documents Codex subscription use with five-hour windows and weekly limits; that establishes a quota shape, not a cross-terminal winner.Unit: 5-hour window + weekly limit · Scope: Codex on ChatGPT subscriptions; native plan mechanics only · Source: OpenAI Codex pricing and credit rates ↗ · Checked 2026-07-20
Wrong if
Wrong if you treat a published quota window, modeled headroom, price, or AA Coding as proof that one terminal agent is better in your repository.
Reverse when
Re-open Stay versus Switch after both candidates have direct, same-task terminal evidence and the official-source audit is back inside the weekly SLA.

Starting state · 02

IDE incumbent

You already work in an editor assistant and need evidence that migration improves the native workflow.

Action
Evidence limited · unranked
Evidence receipt
Cursor documents separate Auto/Composer and selected-model usage pools; that establishes billing mechanics, not comparative editor productivity.Unit: separate monthly usage pools · Scope: Cursor individual plans; billing and routing mechanics only · Source: Cursor models and pricing ↗ · Checked 2026-07-20
Wrong if
Wrong if editor integration, repository indexing, extension support, or policy controls are decisive and have not been tested in both candidates.
Reverse when
Re-open Stay versus Switch after the same repository task and required editor controls are verified in both candidates and the source audit is current.

Starting state · 03

API/BYOK alternative

You are considering metered inference instead of relying only on an included subscription pool.

Action
Evidence limited · unranked
Evidence receipt
Official provider documentation separates direct API or credit billing from included subscription pools, making the meter—not a modeled subscription allowance—the proof route.Unit: USD per million input/output tokens · Scope: direct API, usage-credit, or BYOK spend route; no tool-quality claim · Source: Anthropic model pricing ↗ · Checked 2026-07-20
Wrong if
Wrong if the coding client cannot use your key, the required model or feature is subscription-only, or you cannot enforce a tested spend cap.
Reverse when
Route only after the client/key path, current API rate, and a hard spend cap are verified; until then this path remains unranked.