The weekend agent's model ladder: route jobs without overpaying

One Saturday morning, your side-project agent is summarizing a pile of notes, drafting a reply, and re-reading the same project context for the tenth time. The newest model is tempting, but the bill is the real test.

On September 22, 2026, Anthropic released Claude Opus 5.5, and OpenAI released GPT-6 Sol and GPT-6 Luna within roughly an hour of each other. The price spread matters more than the timing. OpenAI cut API prices for GPT-6 Sol and GPT-6 Luna by 50% relative to GPT-5.6 promotional pricing.

Ask which job can survive a cheaper model. Pick the cheapest lane that still works, then upgrade only when the output fails review.

The checklist is the whole trick

Route by context, volume, and cache behavior. Each rule is a one-minute test: look at the job, pick the lane, and only move it when the cheaper lane fails.

• Send high-volume routine work to GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens.
• Route long-input jobs that need a mid-tier answer to GPT-6 Sol.
• Put repeated-context work on Claude Opus 5.5 at $4 per million input tokens, with output tokens at $20 per million, down from $25 per million for the prior Opus model, and cache reads at $0.20 per million tokens, a 60% reduction from the Opus 5 cache-read price.

Routine work is where volume becomes a bill. Long inputs punish a model that is too small, but a top model is often overkill. Repeated context is where cache behavior changes the math.

The cheap lane is for volume, not heroics

Use the cheap lane for jobs that are frequent, short, and forgiving: status updates, label cleanup, simple extraction, boilerplate checks, and low-stakes rewrites. If the output can be reviewed in a minute, the model does not need to be the smartest one in the room.

A tag suggestion or a short summary can be checked quickly. If the answer is wrong, you fix it by hand. The cheap lane also suits work where a small mistake is cheap to fix; redoing it later is part of the plan. These are the jobs that pile up while you are debugging something else.

Long inputs are where the middle lane earns its keep

Long-input jobs are different. They may need enough reasoning to hold a thread, but they do not always need the top model. The middle lane fits drafts, summaries, and comparisons where the input is heavy and the answer can be checked.

A long-input job might be a stack of customer notes, a repo file, or a long conversation. The model needs to hold the thread, but it does not need to invent a new strategy. If the output is a draft, the human is the final reviewer.

If the job keeps failing because the model is too small, move it up. If the job is mostly repetition, move it down. The route is a setting, not a commitment.

Repeated context is where cache pricing changes the math

Agent work often re-sends the same project notes, code files, or conversation history. Cache reads matter there. A model that is expensive on fresh input can still be reasonable when most of the context is already cached.

Opus 5.5 was more than 30% faster at generating output than Opus 5. For a weekend builder, that speed changes the feel of the tool: an agent that waits makes you go make coffee, while an agent that moves keeps the weekend going.

The repeated-context lane fits agents that keep the same project open: the same rules, the same codebase, the same style guide. Fresh input is expensive. Cached context is not. The stable context is the part you do not want to pay fresh-input prices for every turn. Keep it in the lane where cache reads are cheap.