Anthropic's Claude model family has a clear tier structure in 2026. **Sonnet 4.6** is the mid-tier model. It is cheaper, lower latency, and good enough for most production workloads. **Opus 4.7** is the top tier. It is more expensive per token, has higher time-to-first-token, and wins on the hardest reasoning tasks.
The right way to pick between them is to evaluate both on your own task. Write a test set that looks like real production traffic. Measure quality, latency, and cost for both. If Opus does not produce a quality lift that justifies the price and latency, ship Sonnet.
Option B captures this empirical framing. Cost and latency matter, and Sonnet is the right pick when your evaluation does not show Opus pulling ahead on the work you actually do.
The other options fail simple sanity checks. Option A says cost is marginal, which is false. Option C reverses the tier ordering. Option D invents an output-window limit that does not exist. The actual decision rule is empirical: pick the smaller, cheaper model unless your evaluation proves you need the bigger one.
**The decision is empirical, not theoretical.** Sonnet 4.6 is the cheaper, faster mid-tier; Opus 4.7 is the frontier tier with multi-x price and noticeably higher time-to-first-token. The right move is to evaluate both on YOUR workload and pick the cheaper model unless the eval shows Opus buying meaningful quality lift.
Option B is the only one that names this framing. The mistake in the other options is treating the choice as a property of the models in isolation rather than a property of the model crossed with your task.
Option A claims cost is marginal. In production, Opus typically runs at several times Sonnet's per-token price, and that premium compounds across millions of calls. It also adds noticeable TTFT, which matters for interactive UX.
Option C inverts the tier structure. Sonnet is the smaller, faster, cheaper model. Opus is the larger, slower, more expensive model. There is no regime where Opus is the latency-friendly option.
Option D invents an output-window ceiling. Both models share comparable output limits in the 2026 release line; the difference is reasoning depth, cost, and latency, not output length.
The practical rule is: ship Sonnet by default, run evals on your hardest task slices, and route only those slices to Opus where the lift justifies the premium. Most production fleets end up with a mixed-tier router rather than a single model.
Quick reference
TL;DR
Pick Sonnet 4.6 when evaluation on YOUR task does not show Opus 4.7 buying meaningful quality lift that is worth the cost and latency premium.
Think about renting a vehicle for a delivery route. A pickup truck and a heavy semi-truck can both haul the boxes, but the semi costs more per mile, takes longer to start up, and only pays off when you have to haul something the pickup actually cannot carry. The pickup is Sonnet 4.6: cheaper, faster, and good enough for most loads. The semi is Opus 4.7: stronger on the hardest jobs, but you pay more per call and wait longer to start. The right choice is not to always rent the semi out of caution, and not to refuse the semi on principle. You run a few real deliveries with each, look at which one actually moved your boxes well, and pick the cheaper option unless the bigger one solved a problem the pickup could not.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Workloads where a small absolute quality lift translates to large business value If a few percentage points of correctness on a high-stakes decision saves real money, the frontier tier premium can be cheap, and the empirical comparison may favor Opus even at higher cost.
Latency budgets that already exclude the larger tier If your TTFT budget cannot accommodate Opus, the choice is forced, and the empirical evaluation only confirms Sonnet.
Single-shot research or prototyping where cost per call is not the binding constraint Running Opus end-to-end is simpler than building a router, and at low volume the cost gap is negligible.
Quick scan · 2–4 min
Everything important, quickly.
3 min: tier structure of Sonnet 4.6 versus Opus 4.7 + empirical evaluation discipline + routing by difficulty + cost and TTFT realities + why the wrong distractors fail.
Real-world examples
Real products, models, and research that use this idea.
1Anthropic's pricing page lists Opus 4.7 at multi-x Sonnet 4.6 on both input and output token rates, making the cost gap concrete for any team running production traffic.
2Coding agents typically run a difficulty-aware router that sends multi-file refactor tasks to Opus and single-file edits to Sonnet, capturing the lift only where it exists.
3Customer-support automation routinely ships Sonnet for the long tail of routine intents and routes only escalated reasoning cases to Opus.
4Retrieval-augmented generation systems pair a cheap embedding stage with Sonnet for synthesis and reserve Opus for cases where the retrieved context is ambiguous and needs heavier reasoning.
5Internal evals at most enterprise deployments compare Sonnet 4.6 and Opus 4.7 on a held-out slice of real traffic before any tier change ships to production.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you design the difficulty classifier that routes between Sonnet and Opus?
Sign in or go Pro to see the approach for this follow-up.
QWhat metrics beyond aggregate win rate matter when comparing tiers?
Sign in or go Pro to see the approach for this follow-up.
QWhen would you accept Opus's higher TTFT despite an interactive UX?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Picking Opus everywhere as a safety default. The cost and latency premium is real and often unjustified once a task evaluation runs on the workload you actually serve.