Skip to content

Model Routing

Pay for work you can accept.

Effective routing chooses the exact model and effort level that produce acceptable work at the lowest total cost. That cost includes failed attempts, review and rework.

In The Token Department, I separate buying model access from allocating work. This companion explains the allocation. My own routing table is a starting preference, revised through experience and checked against the task in front of me.

A telephone operator connecting calls at a switchboard.
The four factors depend on one another. This is a conceptual framework with no numerical scale.
Chapters
  1. Exact routes
  2. Accepted cost
  3. Task frontier
  4. My routing table
  5. Orchestration
  6. Feasibility
  7. Review and escalation
  8. Key takeaways
  9. Sources and notes

Specify the model and effort together

Model selection means the token plan or access arrangement you buy. Routing means choosing a model and effort for a particular task. Buying access to a capable model leaves that second decision open.

A route such as GPT 6.1 Sol at high is specific enough to compare with GPT 6.1 Sol at max. Record the model name, the effort setting and the application you ran it in. Evidence at one effort level measures that setting alone. Treat other settings as separate candidates until you test them.

Keep the surrounding system in view. The harness supplies instructions, context, tools and checks. A benchmark run in one harness gives you a candidate to test in yours. The Harness guide explains that part of the system.

Define acceptable work before dispatch. A financial analysis needs calculations that tie to its sources. A code change needs the required behaviour and passing checks. A presentation also needs readable charts and the agreed design. Classification needs labels checked against examples. The right route can differ across all four.

Count the cost of getting to acceptance

The cheapest first attempt can become an expensive completed task. A stronger route may justify its premium if it avoids repeated correction or reduces the time I spend checking the result.

Compare model and tool usage charges, allocated fixed costs and human time across a batch of comparable tasks. Include briefing, coordination, every attempt, review and rework. Divide the total cost by the number accepted. Keep what failed attempts cost in the total. A batch with no accepted work has no finite cost per accepted result.

Total cost per accepted result

Usage charges + fixed costs
+ valued human time
Accepted results
Costs cover successful and unsuccessful work. Value human time at a stated rate and count each charge once. Divide by accepted results.

Keep cash, allowance use and time visible. Where you put a money value on human time, state the rate. Spread fixed subscription fees across the work they cover, and do not count them twice. A route can consume less allowance while leaving this month's bill unchanged. Its benefit may be more usable capacity.

For a deadline, check elapsed time separately. Work that arrives too late fails that task's standard. I want the lowest total cost that reliably produces acceptable work. That still leaves room to pay for better output when the improvement matters.

Find the frontier for your task

A Pareto frontier removes routes for which another delivers at least as much quality at no greater cost, with an improvement in one of them. The remaining routes involve a trade-off. The frontier narrows the choice; your required standard determines which option to use.

The original essay preserved an Arena Overall snapshot captured on 11 October 2026. Within its six selected configurations, Opus 5.5 high had a higher net improvement score and a lower median task cost than Astra max or Fable 5.1 max. Sol 6.1 max cost less again, with a lower score.

Opus dominates Astra and Fable in this subset

Historical capture · 11 October 2026 · Overall category
Historical Arena cost and quality comparisonSix configurations captured on 11 October 2026. Logarithmic cost axis. Luna, DeepSeek, Sol and Opus form the frontier within this subset. Opus has a higher score and lower cost than Astra and Fable. Exact values follow in the table.Net improvement (%)0510150.1110Median cost / task (US$, log scale)Luna maxDeepSeek maxSol 6.1 maxOpus 5.5 highAstra maxFable 5.1 maxHistorical Arena cost and quality comparisonSix configurations captured on 11 October 2026. Logarithmic cost axis. Luna, DeepSeek, Sol and Opus form the frontier within this subset. Opus has a higher score and lower cost than Astra and Fable. Exact values follow in the table.Net improvement (%)0510150.1110Median cost / task (US$, log scale)Luna maxDeepSeek maxSol 6.1 maxOpus 5.5 highAstra maxFable 5.1 max
A historical frontier within six selected configurations. Cost and quality refer to Arena's benchmark tasks. GPT 6 Luna in this capture differs from GPT 5.6 Luna in the routing table below.
Original six-configuration capture
Configuration Median cost per task Net improvement
GPT 6 Luna · max US$0.07 +1.93%
DeepSeek V4.1 Flash · max US$0.12 +3.80%
GPT 6.1 Sol · max US$0.59 +11.72%
Claude Opus 5.5 · high US$1.80 +14.33%
GPT 6 Astra · max US$3.45 +13.09%
Claude Fable 5.1 · max US$6.52 +12.66%

That is a historical comparison on one measure. Net improvement is a benchmark score. These costs do not measure my review time, the cost of an accepted deliverable or subscription capacity. Arena's cost method samples at most the first three tasks in a session.

For my work, I compare routes on a representative set of tasks under the same acceptance rules. Retain failures and review effort. A route that saves time on coding may require more correction on financial analysis. Visual judgement needs its own test. Close benchmark scores also need care because rankings carry uncertainty.

Once routes meet the standard, choose the cheapest reliable one. If maintaining separate routes costs more than it saves, use a capable default. A detailed routing table has to earn its maintenance cost.

My routing table is a dated preference

This is a snapshot of my main routing preferences observed on 15 November 2026. Each entry names a model and effort. These are starting judgements, conditional on access, capacity and tools.

Barnaby's preferences, observed 15 November 2026
Work Starting route Backup route
Finance and strategy GPT 6.1 Sol · xhigh Muse Spark 1.3 · xhigh
Legal GPT 6.1 Sol · high Muse Spark 1.3 · high
Writing GPT 6.1 Sol · high Muse Spark 1.3 · high
Visual presentation GPT 6.1 Sol · high Muse Spark 1.3 · xhigh
Coding GPT 6.1 Sol · high Muse Wonder 1.0 · max
Systems debugging GPT 6 Astra · xhigh Muse Wonder 1.0 · max
Classification GPT 6.1 Sol · medium Muse Spark 1.3 · medium
Listed mechanical edits GPT 5.6 Luna · xhigh Muse Spark 1.3 · low
Orchestration GPT 6.1 Sol · high Muse Spark 1.3 · medium

My rules call for a final pass at max effort on finance, strategy, legal and writing work. Effort labels such as high belong to the application where you run the model. Independent review goes to another provider, routed as that kind of work. The table combines experience, benchmark evidence and explicit personal rulings. It carries no measured claim that every row is optimal or that a listed backup is currently available.

Give the orchestrator a bounded job

I tend to use a highly capable orchestrator with enough usable context to direct the work. It needs the objective, sources, constraints and acceptance conditions. It then assigns bounded tasks using the routing table and checks the results. I delegate heavy reading and building to fresh worker contexts and hand off orchestration at around 30% context usage, before compaction can degrade judgement.

Follow one company briefing from supplied sources to reviewed output. This is an illustrative dispatch using my dated routing table.

Follow one company briefing

1 / 9

The backup case assumes verified access, tools and allowance.

Start with a bounded brief

Produce a short company briefing from supplied sources. The orchestrator defines the outputs and acceptance checks, then builds the dependencies shown below.

Choose a step or play the flow.

A company briefing task graphThe orchestrator assigns source classification and figure checks in parallel. Writing needs both outputs. Page layout follows writing, then final passes, prose review and owner acceptance.OrchestratorGPT 6.1 SolhighPlanningClassify sourcesGPT 6.1 SolmediumWaitingCheck figuresGPT 6.1 SolxhighWaitingWrite the briefGPT 6.1 SolhighWaiting for inputsLay out the pageGPT 6.1 SolhighWaiting for textReview the proseMuse Spark 1.3highWaiting for final passesOwner acceptanceHuman ownerAgainst the briefWaiting for reviewA company briefing task graphThe orchestrator assigns source classification and figure checks in parallel. Writing needs both outputs. Page layout follows writing, then final passes, prose review and owner acceptance.OrchestratorGPT 6.1 SolhighPlanningClassify sourcesGPT 6.1 SolmediumWaitingCheck figuresGPT 6.1 SolxhighWaitingWrite the briefGPT 6.1 SolhighWaiting for inputsLay out the pageGPT 6.1 SolhighWaiting for textReview the proseMuse Spark 1.3highWaiting for final passesOwner acceptanceHuman ownerAgainst the briefWaiting for review

Why this route?

Orchestrator · GPT 6.1 Sol · high

Keeps the objective, sources and acceptance conditions in view. Assigns bounded jobs and checks the evidence returned.

Select a worker to inspect its route. Red packets carry a brief or returned output. The receiving task changes state on arrival.

Illustrative task graph using the dated routing table. Playback timing is for explanation. Each worker has a separate brief and output. The backup case assumes its tools and allowance have been checked.
  1. The orchestrator defines the briefing and checks access. Source classification uses Sol medium; figure checks use Sol xhigh. These tasks can run in parallel.
  2. It waits for both returned outputs and checks their evidence before releasing the writing task on Sol high.
  3. A visual worker uses Sol high. Finance and writing receive final passes at max effort. A separate prose review uses Muse Spark 1.3 high when that provider's tools and allowance are available.
  4. The orchestrator checks the deliverable and review findings against the brief. The owner accepts the result. Publication requires separate permission.

Every worker brief needs an output, source paths and a test for completion. State which files it may change. Ask workers to return evidence and unresolved questions. A confident summary alone cannot establish that a file was read or a check passed.

Parallel work helps when the jobs are independent. Source checks can run alongside a layout task after the text is fixed. Two agents changing the same draft create coordination work. Keep dependent jobs in order and give each file a clear owner. Sometimes one capable agent is the cheapest complete workflow.

Check feasibility before preference

An attractive route can be unusable. Check the exact model and effort in the receiving application, the required tools and each limit that applies, by hour, week or month. A model that accepts images may still lack an image-generation tool. Access in one application establishes nothing about another application's tool connections.

Capacity observations also have a time. Another task can consume the remaining allowance after the check. Treat remaining subscription headroom as an estimate, and re-check after a limit, reset or access failure. A historical dollar cost per benchmark task cannot establish the debit against today's subscription window.

Prepare a fallback with the sources, working state and instructions needed to resume. Test it on representative work before depending on it. A backup row names a candidate; a checked alternative gives you continuity.

Review the failure before escalating

When a first pass fails, inspect the brief, missing context and tools before changing the route. More reasoning cannot repair an inaccessible source. If the problem is reasoning depth, try a higher effort. If the model remains a poor fit, change model or category. Record the failed attempt in the total cost.

Independent verification should examine the deliverable against its sources and acceptance rules. Another provider can supply a separate judgement, but agreement between models still needs evidence. Recalculate figures, run the required checks and inspect the rendered result. Keep self-check and independent review distinct.

Escalate to the person who owns the work when a missing ruling changes the argument or an unavailable source prevents a factual claim. A fallback must preserve the required standard. Quietly weakening that standard makes the routing comparison meaningless.

The Token Department's continuing job is to revisit these choices as tasks, models, prices and the harness change. Keep a record of accepted work, retries, review time and access failures. Use it to update the table and the access you buy. The Model Selector helps compare purchases and allocations; actual work tests whether those choices hold.

Key takeaways

  • Route to an exact model and effort against a stated acceptance standard.
  • Choose on total cost per accepted result, including failed attempts and review.
  • Build task-specific evidence and keep benchmark dates visible.
  • Give the orchestrator bounded briefs, feasible routes and checks on the delivered work.
  • Keep a tested fallback and revise the table when the evidence changes.

Sources and notes


  1. The original founding essay and its preserved review contain these six configurations. Arena Overall was captured on 11 October 2026 at 22:34 HKT, with rankings dated 8 October. Cost is the 50th-percentile median per task over the preceding 14 days. The frontier is within this selected subset and uses point estimates. Arena Agent Pareto is a live page whose current values differ from this capture. ↩

  2. Arena's categories and task-cost method describes task segmentation and the first-three-task sampling limit. Its evaluation methodology explains the benchmark setting. These links were checked on 15 November 2026. Neither source measures the cost of accepted work in my own workflow. ↩

  3. Barnaby Robson's maintained orchestration rules and model-identity registry, read on 15 November 2026. This table transcribes their current preferences. The published orchestrator skill in public-skills provides a reusable provider-neutral slot template. The table here shows my concrete dated bindings. Older benchmark and experiment notes have their own dates and do not measure every current route. The mechanical-edit preference names GPT 5.6 Luna; the historical Arena figure names GPT 6 Luna. They are different configurations. ↩

Relevant posts & resources

Next in the series

Building AI CapabilityEssay · Read

Develop people through bounded work, shared practice and responsibility for the result.

Say hello

Draft a message about the work you need to complete.

Local draft preview. No message is sent. Review the draft, then copy it for use elsewhere.