Running every task through the flagship model is still like hiring a senior architect to update a spreadsheet. The work gets done. The economics stay absurd. The only thing that changed in July 2026 is that the absurdity now has product names: Sol, Terra, and Luna.

I wrote the first Model Ladder post when the practical choice for most solo builders was still roughly "local, mid-tier cloud, or frontier." That ladder still holds. What broke is the fantasy that "frontier" is one button. Frontier labs are shipping families — three price points, three latency profiles, and a subagent mode that will happily burn tokens while looking industrious. If your routing rule is still "always use the best," you are funding someone else's margin with your operating cash.

What actually changed

The public story in early July was straightforward: GPT-5.6 arrived as three tiers. Sol is the expensive flagship with heavier reasoning and subagent behavior. Terra aims at near-prior-generation quality at a lower unit price. Luna is the fast, cheap rung for volume work. The exact list prices move; treat any dollar figure you see on a launch post as a dated snapshot, not a permanent law. What does not move is the structure: one brand, three economic objects.

That pattern is not unique to one lab. Mid-2026 is full of the same shape elsewhere — an "agentic" top tier, a workhorse middle, a fast bottom, plus a separate local or open-weight path that does not care about your cloud invoice. The model ladder is no longer a nice diagram for a blog post. It is the only way a solo builder keeps a multi-agent desk from becoming a random tax.

Two other facts matter more than leaderboard screenshots:

  1. System cards and third-party evals are not your product. Launch coverage openly discussed pre-deployment eval friction and unauthorized-action rates on agent tasks. That is useful as a warning label. It is useless as a routing rule until you measure on your tasks.
  2. Speed and price are different dimensions. A coding agent that streams at extreme tokens-per-second can still be expensive if every "success" requires three retries and a human rewrite. Dollars per accepted task beat dollars per million tokens.

The ladder in 2026 (four rungs, not three)

I still use four rungs. The names changed; the discipline did not.

Rung 0 — Local / open weights on hardware you own.
Classification, extraction, reformatting, offline work, anything that must not leave the machine. On a desk that already has a high-memory Mac and a small Blackwell box, this is not a hobby lane. It is the free-at-the-margin lane and the sovereignty lane. If the correct answer is highly constrained, local wins until it does not.

Rung 1 — Fast cloud (Luna-class, Flash-class, Haiku-class).
High-volume drafts, simple transforms, router decisions, "is this worth a bigger model?" triage. The job of this rung is to be cheap enough that you stop thinking about it — and honest enough that you promote failures upward.

Rung 2 — Workhorse cloud (Terra-class, Sonnet-class, "good enough with review").
Most real shipping work lives here: code that will be tested, docs that will be edited, research that will be checked, agent loops that touch tools under a permission boundary. This should be your default for interactive build sessions.

Rung 3 — Flagship / Ultra / multi-agent (Sol-class, Opus-class, heavy reasoning).
Architecture calls, novel debugging, high-stakes analysis, the rare task where a wrong answer costs a week. This is not a personality setting. It is a budget line. If more than about a fifth of your calls land here without a written reason, your router is broken or your task design is.

The important mental shift: Sol is not "me, but smarter." Sol is "me, with a more expensive error-correction budget and more room to thrash."

A routing table you can actually use

Use the task shape, not the vendor brochure.

Task shape Default rung Promote to flagship when… Demote when…
Classify / extract / reformat Local or Luna-class Ambiguous labels keep flipping Never — keep it boring
First draft of prose or boilerplate code Workhorse Stakeholder-visible final with legal/brand risk Draft will be rewritten anyway
Multi-file coding agent with tests Workhorse Design is greenfield or failure is expensive Tests are green and diffs are tiny
Long tool-using agent overnight Workhorse + hard caps Explicit human gate on irreversible steps Loop is thrashing — stop, don't "Ultra harder"
Architecture / strategy Flagship — Problem is actually a lookup
Voice / real-time product UX Separate lane Product requires barge-in / duplex Async text would be clearer

If you only remember one rule: default down, promote up on evidence. Defaulting up is how API bills become a personality trait.

The subagent tax

Ultra modes and subagent orchestration look like leverage. Sometimes they are. Often they are a way to spend flagship tokens on work a workhorse model could finish with a better checklist.

Before you enable a "smarter with helpers" switch, write down:

  • What does success look like in one sentence?
  • What tools can the agent touch?
  • What is the hard stop (time, dollars, steps)?
  • Who reviews irreversible actions?

If you cannot answer those four, you do not need a subagent mode. You need a spec. Solo builders who skip the spec do not look ambitious. They look like they are renting confusion by the minute.

How to measure without lying to yourself

I do not trust a model card the week it ships. I trust a five-task board that I re-run when the family changes.

Pick five tasks you actually do:

  1. A boring transform (JSON reshape, log slice, CSV cleanup).
  2. A mid-size code edit with a test you already have.
  3. A research-and-summarize pass over material you know well.
  4. A multi-step tool loop (search → read → produce a patch).
  5. One "this actually matters" design question.

For each tier you are considering, record:

  • Wall-clock time to an acceptable draft
  • Tokens in/out (or dollars if the dashboard only shows money)
  • Whether you accepted the result, edited lightly, or threw it away
  • One sentence on failure mode

Then compute the only metric that matters for a solo desk: dollars (or minutes) per accepted outcome. A cheap model that fails twice is not cheap. An expensive model that lands once can be.

When you publish those numbers — even only in your own notes — label them with the date and the model IDs. Last month's Terra is not this month's Terra. Snapshot, do not mythologize.

The catch

Three-tier branding makes people feel sophisticated while they keep routing with vibes.

The catch is not that Sol is "bad." The catch is that families create a false sense of precision. You see three names and assume you have a strategy. You do not. You have a menu. Strategy is the written map from task shape to rung, plus a measurement loop that survives marketing weeks.

The second catch is contamination of judgment. Eval drama, arena ranks, and launch-stream demos are optimized for narrative. Your product is optimized for accepted diffs, correct classifications, and bills you can stand to look at. If those two ever disagree, keep your desk and throw away the narrative.

The third catch is local amnesia. Every time a new flagship drops, builders abandon hardware they already paid for. Local is not anti-cloud. Local is the free rung and the private rung. Cloud families are for the rest. A ladder with a missing bottom rung is just a slide into spend.

What I am doing on my desk

My defaults after the July family launches:

  • Local first for anything constrained, bulk, or sensitive.
  • Workhorse cloud for normal build loops and drafting.
  • Flagship only with a written reason in the task itself — not a sticky note on my soul.
  • No Ultra/subagent on irreversible work without a human gate.
  • Re-run the five-task board when a vendor renames the menu again.

That is not ideology. It is how you stay solvent while the industry renames the same three economic objects every quarter.

Who this is for

If you are a team with a forward-deployed engineer living inside your repo, you can afford messier routing for a while. Someone else is paid to clean up the bill and the blast radius.

If you are solo, you are the deployment company. The ladder is your staffing plan. Sol, Terra, and Luna are just job titles with price tags. Promote carefully. Demote without ego. Measure on your work, not on theirs.

The menu will change again. The discipline should not.