Claude Sonnet 5 vs Opus 5: The Answer Changes on September 1

Claude Sonnet 5 is currently on introductory pricing, and that expires soon:

Most Sonnet-vs-Opus comparisons quote the promo rate without mentioning it's temporary. That matters, because the answer to "is Sonnet enough?" is partly an economic question, and the economics change by 50% at the end of this month.

We ran both through the same battery as the rest of this series.

Quality: no separation

Seventh consecutive post where the hard battery separates nothing on quality. Sonnet 5 found the interval bug that only surfaces on a fully-contained interval, counted letters correctly, and held two simultaneous constraints — same as the model costing 2.5× more.

Sonnet 5 is cheaper and faster

Cheaper, about 20% faster, and marginally more concise. On this battery there is no dimension where paying for Opus 5 bought anything measurable.

And it follows rules better

Ten runs each of the no-letter-"e" constraint test:

One Opus 5 run returned a provider error and is excluded rather than counted as a failure — hence 9 trials, not 10.

This continues the clearest pattern in the series. Across five separate runs of this identical prompt, Opus 5 has now scored 1/5, 3/5, 4/5, 0/5, and 5/9 — 13 clean runs out of 29, about 45%. That's a large enough sample to say something real: Opus 5 honours this constraint roughly half the time, and which half you get is a coin flip.

Sonnet 5 at 9/10 is meaningfully better, and both Gemini Flash models went 10/10 at the same sample size. The most expensive model in Anthropic's lineup is the least reliable at obeying a rule a script can check in one line.

The September 1 problem

Here's the part that isn't in other comparisons. Our measured costs, projected forward:

The same work, the same models, and the advantage collapses from 78% to 19% on a calendar date.

That's still cheaper, and given Sonnet 5 also won on latency and constraint adherence here, it remains our pick. But 19% is a margin where other factors legitimately dominate — and if you built a cost model on the promo rate, it's wrong from September 1. Anyone forecasting spend on Sonnet 5 right now should be forecasting at $3/$15.

It also reframes the alternatives. At standard pricing Sonnet 5 costs roughly what Kimi K3 does, and considerably more than Grok 4.5, both of which passed this same battery. The promo isn't just a discount on Sonnet — while it lasts, it's what keeps Sonnet competitive with the cheap tier.

What we'd do

Use Sonnet 5 over Opus 5 for this class of work. It matched on quality, beat it on speed, beat it clearly on constraint adherence, and costs less under either pricing regime.

Re-forecast at $3/$15 now. If your budget assumes $2/$10, it expires 2026-08-31.

Don't reach for Opus 5 to make output rules more reliable. Our 29-trial result says that's backwards. Validate the output instead — that works regardless of tier and costs nothing.

Re-benchmark in September. At standard pricing this becomes a genuinely open comparison against the open-weight and budget-frontier options, in a way it isn't today.

Run it yourself

No account, no API key:

• The constraint test — run it a few times and watch the variance
• The interval-bug task — the bug only fires on a contained interval

To compare both models on your own prompts with real token counts, that's Promptster.

Run 2026-08-09 against claude-sonnet-5 and claude-opus-5 via the Promptster API, max_tokens: 4000, default temperature (both models reject an explicit temperature). Costs computed from list pricing: Sonnet 5 at its introductory $2/$10, Opus 5 at $5/$25 per million input/output tokens; the September projection multiplies Sonnet's measured cost by the $3/$15 standard rate. The three graded tasks were run once per model — enough for pass/fail against a deterministic answer, not enough to characterise latency variance. The constraint task was run ten times per model; one Opus 5 run returned a provider error and is excluded from its denominator rather than scored as a failure.