AI Foundations

Think harder, consult, delegate—or switch?

An occasional hard decision may need a stronger advisor. A task full of hard decisions may need a stronger executor.

Reviewed 22 September 2026

AI Caramba · Editorial illustration

I. Start with the problem, not the model menu

Model choice and the cost of finishing · Part 3 of 3

An occasional hard decision may need a stronger advisor. A task full of hard decisions may need a stronger executor. Neither is a reason to make every step more expensive.

Match the intervention to what is missing

The hard part is over

→ Downgrade
Hand over the plan and acceptance tests.

One decision is blocking progress

→ Think harder or consult
Buy a bounded diagnosis, not a whole new session.

Several tasks are independent

→ Delegate
Separate the questions; count integration work.

Difficult decisions keep recurring

→ Upgrade
Avoid paying for repeated handoffs.

Evidence, access or context is wrong

→ Repair the setup
More intelligence does not grant missing permissions.

Already working well and nearly finished? Stay. These are decision rules to test, not benchmark findings.

More reasoning tackles a difficult line of thought. More agents divide independent work. A model upgrade changes who carries the task. These are different interventions, not steps on a mandatory ladder.

A finished plan can justify a downgrade, but implementation may uncover new design problems. Pass on the constraints, failed approaches and tests—and a condition for asking for help again.

II. Consult the expert. Keep the executor.

Instead of handing over the entire job, ask a stronger model for one decision: “Which explanation fits this failure, and what test would distinguish the alternatives?” The main agent then checks the advice and continues.

One focused consultation—not a full model switch

1 · Sol briefs
Frame the question
$0.0700

2 · Astra advises
Return a diagnosis
$0.3500

3 · Sol checks
Five further calls
$0.3752

Scenario total: $0.7952 · Parent cache hit assumed.

Assumes one useful answer from a 20k-token briefing and five successful follow-ups. This outcome has not been measured.

This is our compact-briefing calculation using GPT-5.6 Sol and GPT-6 Astra, not Claude Code's native advisor tariff. Any advantage can disappear if the briefing omits key evidence, the advice fails, or consultation becomes constant.

The feature itself is real: Claude Code's experimental /advisor opus enables an advisor, and the main model chooses when to consult it. It requires the Anthropic API route and supported pairings. Claude Code documentation.

III. Cache invalidation: keeping the model is not enough

Preserving the executor's context is useful, but the next request still needs a cache hit. Losing its 102,000-token prefix on return adds $0.4692, raising our consultation total from $0.7952 to $1.2644. That is a calculated downside scenario, not an observed failure rate.

Reasoning-setting changes are not universally complete cold starts. Supported update mechanisms can retain the prefix; replacing earlier instructions or tool definitions can prevent reuse. Verify the actual harness behavior. OpenAI reasoning, Anthropic effort.

A cache discount is not a reason to keep wrong or obsolete context. Compaction may rebuild a prefix yet lower subsequent costs by shortening it; compare the whole remaining trajectory. Claude Code caching.

IV. A useful rule for the next attempt

After repeated failures, require a new hypothesis or a bounded consultation—not another identical retry. Limit cost, time and permissions. If difficult decisions recur throughout the task, compare a full upgrade with continued consultation.

Part 1 prices the downgrade. Part 2 adds rework and failure costs. The same rule covers both: change only what the remaining work needs, then verify that it helped.

V. Details, evidence and implementation

How the $0.7952 consultation is calculated

Sol starts with 100k cached input tokens. Preparing the request writes 2k new tokens and produces 1k billed output: $0.0700. Astra writes 20k input and produces 2k billed output: $0.3500. Five further Sol calls cost $0.3752; the first adds 4k input, including the 2k advisor answer, and the rest add 2k each. Each produces 1k billed output. These are net input increments; do not add the carried responses again. Billed output includes reasoning.

The calculation uses Standard short-context prices per million tokens: Sol reads/writes/output $0.40/$5/$20; Astra $1/$12.50/$50. These are the same tariff snapshot as Part 2. Pricing.

On return, Sol's old prefix is 102k. A full read-to-write change costs 102,000 × ($5 − $0.40) / 1,000,000 = $0.4692 extra. Do not add ordinary input pricing again. Partial misses use only the lost portion; input that is not cache-written belongs to the ordinary-input category instead. This example assumes the parent prefix is re-written. Billing categories.

The resulting $1.2644 excludes external tools and unlisted human review. Both outcomes assume a sufficient briefing, useful advice and acceptable completion—not measured speed or reliability. Compare with the cost of the actual alternative, including any cache loss it would experience too.

Does consultation actually improve results?

Anthropic reports Sonnet 4.6 plus an Opus 4.6 advisor improving SWE-bench Multilingual by 2.7 percentage points at 11.9% lower cost per task. Thinking settings and the system prompt also changed: this supports the tested configuration, not an isolated advisor effect. Advisor study.

A separate Chartography comparison provides a counterexample: an executor/advisor pairing performed within reported noise of Fable 5.1 alone at medium effort, at about 2.6 times the cost. Frequent consultation can cost more than letting the stronger model work directly. Cost comparison.

Parallel agents are also conditional: a study of 180 configurations found benefits on some decomposable work and deterioration on sequential tasks. Distinct questions and isolated edits matter more than agent count. Agent scaling study.

Cache details that the UI does not show

For GPT-6 Astra, a supported appended configuration_update can change effort while request-level effort stays unchanged. The documented path is Standard single-agent mode and has compaction/truncation constraints; do not assume a UI effort toggle implements it. Reasoning guide.

Anthropic also documents prefix-preserving per-message effort for supported models, including Fable 5.1 and Opus 5. Top-level effort changes can affect message caches, with effects on earlier prompt components varying by model. Effort, cache invalidation.

Claude Code documents that toggling its advisor leaves the executor cache intact, while advisor calls process the full transcript without advisor-side reuse. Our calculation instead supplies a selected 20k briefing. Claude Code advisor.

The Advisor API separately offers optional advisor caching. Include the advisor_message entries in usage.iterations: top-level usage totals cover only the executor. Do not add executor iterations a second time after counting those top-level totals. Advisor API.

More capability does not require more write access. A pure advisor can return guidance without tools; a researching specialist may need read-only tools. Give either a clear stopping condition and test the proposed answer.

Sources checked September 22, 2026. No new model, provider-cache or harness experiments were run; numerical examples were independently recomputed.

Sources

Sources checked 2026-09-22.