Opus 5 vs Sonnet 5 for coding

On llmrelay, Claude Opus 5 is $2.50 input / $12.50 output per million tokens and Claude Sonnet 5 is $1.00 / $5.00 — a 2.5× rate gap, both exactly half of Anthropic list ($5/$25 and $2/$10). Spec for spec they are closer than the price suggests: both carry a 1M-token context window and 128k max output, both support adaptive thinking, and both are pinned snapshots on the same key. What you are actually buying with the premium is the May 2026 knowledge cutoff against Sonnet’s January 2026, and Anthropic’s own latency ratings put Sonnet 5 at Fast against Opus 5’s Moderate.

At a realistic heavy-agent volume of 50 million input and 5 million output tokens a month, Opus 5 bills $187.50 and Sonnet 5 bills $75.00 — a $112.50 monthly gap at our prices, $225 at list. The arithmetic is (input millions × input rate) + (output millions × output rate), so you can rerun it on the calculator with your own traffic before trusting either number. At double that volume the gap doubles; it scales linearly because there is no tiering or commit discount on either model.

The 2.5× rate gap is the floor, not the expectation. We sent one identical code-review task to both models on 31 July 2026 and read the bill out of our own logs: Opus 5 cost $0.0152, Sonnet 5 cost $0.0047 — 3.27×, because Opus 5 emitted 875 output tokens against Sonnet 5’s 592 for the same prompt. The full breakdown sits in the measured block above. The pattern to take from it: on generative work — writing modules, reviewing diffs, producing migrations — the expensive model’s verbosity compounds the rate gap, while on retrieval-shaped traffic (large prompt, short answer) the effective gap narrows back toward 2.5×.

Caching shifts the math again. Cached input reads bill at 0.10× the input rate and writes at 1.25×, so a repeated system prompt or codebase context costs $0.10 per million read tokens on Sonnet 5 and $0.25 on Opus 5. If most of your input is repeated context — the common shape for agents working inside one repo — the per-call difference between the two models shrinks toward the output side, which is where the measured gap actually came from.

The policy that falls out of this: default to Sonnet 5, and escalate to Opus 5 per task when the work depends on library behaviour newer than January 2026, when a long agent chain makes one bad step expensive to unwind, or when you want the breadth of review the measurement showed the premium buying. Escalating by exception is cheaper than defaulting to Opus 5 and trying to find downgrades, because the cases where the premium pays are specific and the rest is diffuse.

Opus 5’s reliable knowledge cutoff is May 2026; Sonnet 5’s is January 2026, per Anthropic’s published model documentation. Four months does not sound like a reason to pay 2.5× until a dependency you use shipped a breaking change inside that window. The failure mode is quiet: the model does not signal uncertainty, it confidently writes against an API signature that has since changed, everything reads fine, and it fails at runtime.

If you would rather stay on Sonnet 5, the mitigation is to convert the knowledge problem into a context problem: paste the current signatures, the relevant doc section, or the actual type definitions into the prompt. At $1.00 per million input tokens that context is cheap, and it is usually the correct fix even on Opus 5. What the cutoff difference really prices is the work where you did not know which dependency details mattered until the failure.

We publish no benchmark scores for either model here or anywhere on the site, because we cannot source them. The one measurement we do have is the single measured task above — an existence proof that the premium does not automatically buy correctness, not a quality ranking. Anthropic’s latency rating for Sonnet 5 (Fast) did not show up in our one 2000-token-capped run, where both models landed within a second of each other; read that as no meaningful difference observed at that size, not as evidence against the rating.

If you want the full task, both models’ complete answers, and the method for repeating the comparison on your own prompts, they are on the measured Opus-vs-Sonnet page at /compare/opus-4-7-vs-sonnet-4-6-for-coding/ — same two models, same data.