The Vibe Father
Roundup
The single biggest mistake we see in agent setups is running one model for everything. It is either too expensive (a flagship doing grunt work) or too weak (a budget model attempting architecture), and usually both at different moments of the same session. The fix is the same one engineering teams figured out a century ago: roles. A planner, builders, scouts, and a reviewer — each seat staffed by the model whose published numbers actually fit the job.
This is the multi-model thesis we laid out in why the harness matters, made concrete. Every score below is from its public source, live at /benchmarks.
The lineup card
| Role | Our pick | Backup | Why |
|---|---|---|---|
| Planner |  Claude Fable 5 | Claude Opus 4.8 | Deepest reasoning on the board; plans are cheap in tokens |
| Builder | Claude Sonnet 5 |  GPT-5.3 Codex /  Kimi K2.7 | Price-performance for the highest-volume seat |
| Scout | Claude Haiku 4.5 |  Gemini 3.5 Flash | Fast, cheap reads; speed matters more than depth |
| Reviewer | GPT-5.5 |  Grok 4.5 | A different lab than the builder — no loyalty to the mistake |
Planner: Claude Fable 5 (or Opus 4.8)
The planner decomposes the goal, decides the architecture, and hands out tasks. It is the seat where capability matters most and cost matters least — a plan is a few thousand output tokens, while a bad plan poisons every downstream token in the session. That asymmetry is why we staff it with the strongest reasoner available regardless of price. Fable 5's 95.0 SWE-bench Verified is the best evidence anywhere of a model that understands real codebases, and its 1M context means it can hold the whole repo in view while it thinks. Even at $10/$50, a plan costs pocket change.
If Fable's pricing offends you, Opus 4.8 at 88.6 SWE-bench and half the price is a completely respectable planner. What does not belong in this seat is anything below the flagship tier — saving two dollars on planning is the most expensive economy in agent work.
Builder: Sonnet 5, GPT-5.3 Codex, or Kimi K2.7
Builders are the volume seat. They take a well-specified task and implement it, and they burn the overwhelming majority of your tokens — so this is where price-performance decides your monthly bill. Sonnet 5 is our default: 85.2 SWE-bench Verified at $3/$15, 89 tokens per second, the strongest published repo score at a builder price. GPT-5.3 Codex ($1.75/$14) trades SWE points (74.8) for the best budget-tier Terminal-Bench at 78.4 — better when your builders need to drive a shell hard. Kimi K2.7 Code at $0.95/$4 is the true economy option; its SWE and Terminal-Bench scores are not yet published, so audition it on your workload rather than assuming, but the price makes the audition nearly free.
The key discipline: builders should receive specific tasks with clear success criteria. A great builder with a vague brief becomes an expensive random-code generator.
Scout: Haiku 4.5 or Gemini 3.5 Flash
Scouts do the reading: mapping unfamiliar code, summarizing files, checking whether a symbol is used anywhere, drafting throwaway experiments. The work is high-frequency and low-stakes, which flips the selection criteria — speed and price dominate. Haiku 4.5 ($1/$5, 96 tokens per second) is built for exactly this, with one loud caveat from its own scorecard: a 35.5 Terminal-Bench means it must never be promoted to autonomous multi-step work. Keep it on reads. Gemini 3.5 Flash is the deluxe scout — 167 tokens per second with a real 79.3 SWE score, so it can safely absorb small build tasks when the scouting is done.
Reviewer: GPT-5.5 or Grok 4.5 — and always a different lab
Here is the rule we consider non-negotiable: the reviewer comes from a different lab than the builder. Models from the same family share training lineage, and in practice they share blind spots — a Claude reviewing Claude-written code tends to nod along at exactly the assumptions the builder made, because it would have made them too. A cross-lab reviewer has no loyalty to the mistake.
Since our default builders are Claude models, our default reviewer is GPT-5.5 — and its 83.4 Terminal-Bench, the top score on our entire board, means it can actually run the code it is reviewing rather than just reading it. Grok 4.5 is our second seat: its 87.4 LiveCodeBench shows sharp algorithmic judgment, and it arrived with a notably blunt critical register that suits review work (SWE and Terminal-Bench scores pending — it shipped July 9). If your builders are OpenAI models, invert the rule and review with Fable 5 or Opus.
Why the economics work
The role split is not just a quality argument; it is an arithmetic one. Token consumption across the seats is wildly lopsided — in a typical session, builders and scouts generate the overwhelming majority of tokens, while the planner's output is a few thousand tokens of decisions and the reviewer's is a few thousand tokens of findings. That means the expensive models sit precisely where volume is lowest and leverage is highest, and the cheap models sit where volume is highest and tasks are most forgiving. Run the same session all-Fable and you pay flagship prices for scaffolding; run it all-budget and you pay in failed plans and unlaunched retries. The split gets you flagship judgment at workhorse prices.
It also degrades gracefully. If a builder stalls on a task, escalation is built into the structure — the planner re-scopes it or a stronger model takes one pass at it, and the session continues. In a single-model setup, a stall means either grinding retries on a model that has already demonstrated it cannot do the task, or manually swapping models mid-flow. Roles turn "which model should I use?" from a daily agonizing decision into a routing rule you set once and tune quarterly.
Putting it together
A concrete stack we actually run: Fable 5 writes the plan, two Sonnet 5 builders implement in parallel, Haiku scouts ahead of them, and GPT-5.5 reviews every diff before merge. The token bill lands dramatically lower than all-Fable and the output quality lands dramatically higher than all-Sonnet, because each model spends its tokens where its scores say it should. The orchestration mechanics — how the seats hand work to each other without stepping on it — are in our multi-agent field guide.
Rosters change as scores change. Grok 4.5's missing benchmarks will publish, Kimi K2.7 will show its slate, and some seat above will flip. The current numbers, cited to their sources and refreshed nightly, are at /benchmarks.