Guide

Start at the cheapest model that meets the contract, then prove you need more

The matcher below records RossLabs routing heuristics rather than universal rankings. Context length, tools, modality, privacy, availability, latency, and budget can each override the route it suggests, so the deciding evidence is a task-specific evaluation on your own workload.

The short version

Start with the lowest-cost model that meets the task contract. Use task-specific evaluations to decide whether a higher tier improves the result enough to justify its cost. The matcher below records RossLabs routing heuristics, not universal rankings. Context length, tools, modality, privacy, availability, latency, and budget can override its starting route.

Find a model for your task

Pick the task and how open-ended it is. The result is a starting route with a required check, not a benchmark verdict.

Starting route. Task and complexity narrow the choice. Context length, tools, modality, privacy, availability, latency, and budget can override it. Validate the route on a labeled evaluation set.
1 What are you doing?

build from a clear spec

2 How hard is the work?
Start withSonnet

A lower-cost coding model is a reasonable starting point when the specification and tests are explicit.

The checkGive it a plan that has been pressure-tested. It builds whatever the plan says, gaps and all.

Why models differ

Start with measured capability and labeled RossLabs observations, then apply hard constraints such as tools, context, privacy, latency, availability, and cost.

1. The ceiling

The hardest evaluated task it can pass. Use your acceptance tests to find the lowest-cost model that clears the required bar.

2. The style

How it works. What it pays attention to. What it reaches for. This is what counts once the work gets hard.

Treat behavior descriptions as routing hypotheses. A model update, prompt, tool set, or runtime can change the result. Match the model to the task, then test the failure mode that matters most.

The five models

Open a card for the full picture. Behavior statements are RossLabs observations, not vendor guarantees or benchmark results.

Fable RossLabs observation: favors grounded, inspectable answers
What it doesIn RossLabs tasks, it tended to inspect the available system evidence, choose a relatively simple path, and review its output before handoff.
Best forJudgment calls. Code review, planning, audits, and design direction.
WhyUse it as a candidate when a wrong call is expensive, then verify its claims against the system and acceptance tests.
Watch forOn things it cannot inspect, like live runtime behavior, it can build a clean story that is wrong. Make it test those claims.
Opus 5 RossLabs observation: maps structure and failure paths
What it doesIn RossLabs tasks, it tended to map ordering, dependencies, and load-bearing parts before proposing a complete solution.
Best forCoordination and hard reasoning. Orchestrating work, breaking down big problems, fixing tangled async bugs, writing plans. Long autonomous runs where you give the whole task up front and let it go.
WhyIn RossLabs tasks, it has been a useful starting point for coordination and reasoning across many moving parts. Validate that pattern on your own workload.
Watch forIt writes longer than you expect — both its replies and the files it leaves on disk — and asking it to double-check its work now makes that worse rather than better, since it already verifies by default. It can also widen the task beyond what you asked. Tell it to keep the change minimal and say so about length.
Sonnet Builds exactly what the plan says
What it doesFollows the spec and the local conventions. Ships fast and cheap.
Best forDay-to-day building. Carrying out a clear plan, bulk work, and scanning a codebase for issues.
WhyIts current list price is lower than Fable and Opus. Measure whether it meets the same acceptance tests on your well-specified work.
Watch forIt inherits the gaps in whatever plan you give it. Hand it a plan that has been pressure-tested first.
Haiku Fast, cheap, and literal
What it doesApplies clear patterns and conventions quickly. Follows an explicit checklist well.
Best forRote work. Mechanical edits, repetitive refactors, and a fast first pass.
WhyIts lower list price makes it a candidate for mechanical work with explicit acceptance tests.
Watch forDo not infer broad issue recall from a successful mechanical edit. Evaluate scans against a known issue set.
Codex (GPT-5.6) Designed for long coding-agent runs with active steering
What it doesThe Codex family is designed for extended coding tasks. It can iterate through failures and work across a large codebase, but completion and scope control still require tests and review. Three tiers provide different price and capability points.
Best forLong coding-agent runs in a large codebase, multi-file refactors, extended debugging, structured output, and independent review when a separate model family adds useful diversity.
WhyIt is designed for long agent runs and large code contexts. Validate task completion, side effects, and cost against the other models you can use.
Watch forLong prompts with many constraints can still produce scope drift or missed requirements. Keep the task contract explicit, verify front-end and UI work visually, and confirm the active tier and long-context pricing before budgeting.

Where each family stands as of August 11, 2026 List price per 1M tokens, standard context · all three Codex tiers share 1.05M context and 128K max output

  • Fable claude-fable-5 $10 in / $50 out Anthropic's most capable widely released model.
  • Opus claude-opus-5 $5 in / $25 out Thinking is on by default. 1M context, 128K max output.
  • Sonnet claude-sonnet-5 $2 in / $10 out Near-Opus quality on coding and agentic work. Introductory rate through 31 Aug 2026; $3 / $15 from 1 Sep.
  • Haiku claude-haiku-4-5 $1 in / $5 out 200K context. The cheap tier.
  • Codex — Sol gpt-5.6-sol $5 in / $30 out Top tier. The one to reach for on the hardest runs.
  • Codex — Terra gpt-5.6-terra $2 in / $12 out Balanced agentic coding tier. Long-context requests use higher rates.
  • Codex — Luna gpt-5.6-luna $0.20 in / $1.20 out Cost-sensitive agentic coding tier. Long-context requests use higher rates.

Per-token rates are not directly comparable across the Claude rows: Haiku 4.5 uses the older tokenizer, while the 5-series produces roughly 30% more tokens for the same text. The cards above describe observations that may change with a version bump. This table also changes — check the source before you budget against it: Anthropic pricing · OpenAI pricing.

Should you use one model for everything?

No. Even if the best model were also the cheapest. Three reasons.

  1. 1
    Style does not rank cleanly

    Model behavior varies by task, prompt, tools, and version. No single style should be treated as a universal winner without comparative evaluation.

  2. 2
    Most work does not need a top model

    When a task has a clear contract and strong checks, a lower-cost model may meet the requirement. Measure the pass rate before paying for a higher tier.

  3. 3
    An outside opinion is worth keeping

    A model from a different provider can supply an independent implementation and training lineage. Disagreement is a reason to inspect the evidence, not proof that either model is correct.

A way to think about the next model

When a new model lands, you do not have to ask if it is the best. Ask four things.

1Where is its ceiling?

What is the hardest task it gets right? Everything below that line is a job for the cheapest model that also clears it.

2What is its style?

What does it pay attention to, and where does that style slip? The strength and the weakness are the same habit.

3What does it cost for your common work?

Most of your work is everyday building. Price it on the common job, since that is where the volume is.

4Is it from a different company?

A different source buys you an independent check the others cannot give.

Think in roles. The reviewer role. The builder role. The outside-opinion role. Fill each one with whatever model fits it best today. The roles outlast the models.