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.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Frontier model selection looks like a single choice and is actually a routing decision. In the Claude 2026 lineup, Sonnet 4.6 is the mid-tier model and Opus 4.7 is the top tier. The price gap is several times per token. The latency gap is meaningful on time-to-first-token. The quality gap is real but concentrated on the hardest slice of any given workload.
This question tests whether you treat model selection as an empirical, workload-conditioned decision or as a global preference. The correct framing rejects both lazy defaults (always Opus for safety, always Sonnet for cost) and replaces them with an evaluation discipline that measures lift on the task you actually serve.
The deep dive walks through the tier structure, the eval methodology, the routing architecture that follows from the eval, and the specific reasons each distractor fails. By the end you should be able to design a mixed-tier serving system and defend each routing decision with an eval result.
The Claude 2026 tier structure
Anthropic's lineup follows a tier-by-capability pattern. Sonnet 4.6 is the mid-tier general-purpose model: priced for high-volume production, latency tuned for interactive UX, quality strong on the bulk of real workloads. Opus 4.7 is the frontier tier: heavier active parameter count, deeper reasoning behavior, multi-x per-token price, and noticeably higher time-to-first-token because the larger model is more expensive to bring up to the first emitted token.
The price gap matters at scale. At a hundred million tokens per month, the multi-x premium is the difference between viable unit economics and a write-off. The latency gap matters in interactive surfaces: a chat UI that crosses the perceptual TTFT threshold feels sluggish even when total generation time is comparable.
The quality gap is real but concentrated. On easy and median traffic, the two tiers are often statistically indistinguishable on a representative eval. The lift, when it exists, shows up on the hardest slice: long-context reasoning, multi-step planning, ambiguous retrieval contexts, code refactors that span many files. Treat the lift as a property of the task slice, not a uniform property of the model.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic'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.
- Coding 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.
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?
Start with cheap signals available before the LLM call: prompt length, presence of code blocks, retrieved-context ambiguity, intent classification from a small fast model. Tune thresholds against the eval set where Opus actually wins. Be conservative; misrouting a hard case to Sonnet is more visible than misrouting an easy case to Opus.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
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.
60 second bullets to scan on the way to the call.
The Claude 2026 tier ordering by price, latency, and reasoning depth
Why model selection is an empirical question, not a theoretical one
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.