For about a year my model-picking logic was lazy and it worked: if the task mattered, reach for Opus; if it did not, reach for Sonnet and accept slightly worse output. Claude Sonnet 5 broke that shortcut on June 30, 2026, and the replacement rule is genuinely better — but only if you understand one buried setting. Here is the answer up front: Sonnet 5 lands within a couple of points of Opus 4.8 on most benchmarks, beats it on knowledge work, and costs less than half as much per token. Yet the effort dial can quietly flip it from "bargain" to "more expensive than Opus for the same result." The whole decision lives in that detail.
I run these models daily across my own agent pipelines, where the invoice arrives whether the reasoning was worth it or not. This is the decision framework I actually use, kept current for the Opus 5 era.
Where Sonnet 5 Sits Now
Sonnet 5 is the mid-tier of the Claude family — above the entry models, below Opus and the Fable/Mythos ceiling. It became the default for Free and Pro plans at launch and is available everywhere else including the API. If you have used Claude on the web since the end of June without changing anything, you have probably been talking to it already.
The launch framing everywhere was "a cheaper way to run agents." True, but it undersells the structural shift: for the first time, Sonnet and Opus sit on a single cost-performance curve rather than two separate tiers. You stop asking "good model or great model?" and start asking "which point on the curve does this specific task need?" That reframing is the durable takeaway, and it survives every subsequent launch — including Opus 5, which arrived July 24 at the same $5/$25 sticker as Opus 4.8, leaving the math below intact whichever Opus you run. (I broke down what Opus 5 actually changed separately.)
The Benchmark Picture, Read Honestly
Column by column, not as a hype reel:
Hard agentic coding (SWE-bench Pro): Sonnet 5 hits 63.2%, up from Sonnet 4.6's 58.1%. Opus 4.8 leads at 69.2%. Six points — the widest clean gap on the board, and the one that should give you pause if your workload is autonomous, multi-file coding.
Terminal-driven coding: Sonnet 5 jumped to 80.4% from 67.0% in one generation. Where Opus lands on this harness is genuinely contested across reports, so I will not hang a conclusion on it. What is not contested: Sonnet 5 demolished its own predecessor.
Reasoning with tools: without tools Opus leads clearly; give both models search and function calling and the gap nearly evaporates. Most real agent workloads have tools. Keep that in your pocket.
Knowledge work (GDPval-AA v2): Sonnet 5 scored 1618 against Opus 4.8's 1615. The mid-tier model won the benchmark closest to "do useful professional work."
The shape is unmistakable: minimal drop-off almost everywhere, a win on knowledge work, and exactly one category — hard agentic coding — where Opus keeps real daylight. A narrow moat for a model costing more than twice as much. I went deeper on the coding side specifically in my hands-on Sonnet 5 agentic coding test.
The Pricing Math
Per million tokens at standard rates:
| Model | Input | Output |
|---|---|---|
| Fable 5 | $10 | $50 |
| Opus 5 / Opus 4.8 | $5 | $25 |
| Sonnet 5 | $3 | $15 |
Two footnotes that matter. First, Sonnet 5 launched with introductory pricing of $2/$10 that runs through August 31, 2026 — as I update this, that window is closing within weeks, so build your forecasts on $3/$15, not the teaser rate. Second, even at standard pricing, Sonnet 5's output tokens cost 40% of Opus's, and output is where agent bills live.
Scaled up, this is not a discount but a different budget category: an output-heavy agent workload costing $1,000 a day on Opus lands around $400-600 on Sonnet 5 depending on effort settings. For a solo builder, whatever. For a fleet of parallel agents, that is the line between a product with unit economics and a science project. It is the same class of decision I walk through in cutting Fable 5 usage costs — the premium tier exists to be invoked selectively, not to run the whole show.
But that savings figure is conditional on a setting most teams discover via the invoice.
The Effort-Level Trap
This section is why the article exists. Sonnet 5 exposes effort levels — low, medium, high, xhigh, and max at the top of the dial. Higher effort means more reasoning tokens before the answer; reasoning burns output tokens; output tokens are the bill. The dial often matters more than the model choice, and with Sonnet 5 it determines whether you got a bargain at all:
- Low: cheapest, and quality can dip below Sonnet 4.6 on demanding agentic tasks. Fine for forgiving, high-volume work.
- Medium: the sweet spot. Comparable quality to the old tier at meaningfully lower cost — "cheaper and at least as good" is simply true here, no asterisk.
- High: clearly better output, but the cost climbs to roughly what high-effort Opus costs. You have spent your way back to Opus pricing.
- Xhigh and max: firmly in Opus money, for results that land around Opus's middle gear on agentic benchmarks. You are paying as much as — sometimes more than — Opus to approximate its medium setting.
So the naive read — "Sonnet is cheaper, crank it up to be safe" — is exactly backwards at the top of the dial. Maxed-out Sonnet 5 is, for many tasks, the worst of both worlds: Opus money for a sub-Opus ceiling. The savings live at low and medium effort, full stop. If you catch yourself reaching for xhigh "to be safe," that is not a settings decision; that is the signal to switch models. I first mapped this dial behavior on Opus 4.8's effort levels, and Sonnet 5 makes the same trap cheaper to fall into and more expensive to stay in.
The Decision Framework
Reach for Sonnet 5 when:
- The work is routine to moderately complex — content generation, extraction, classification, standard CRUD-flavored coding, most knowledge work (it beat Opus there).
- You are running at volume, where per-call cost compounds.
- Tools are in the loop, which is where its reasoning gap nearly closes.
- You can live at low or medium effort — the only settings where the cost advantage is unconditional.
Reach for Opus when:
- The task is hard autonomous coding: multi-file refactors, gnarly debugging, long agentic runs. The six-point gap compounds, and Opus's token efficiency means it often reaches the answer in fewer expensive steps — frequently at lower total cost than maxed-out Sonnet 5, because it does not need the top of its dial to get there.
- Precision is non-negotiable and you will not babysit output.
- You were about to set Sonnet 5 to xhigh. If you need that ceiling, buy the real one.
The clean heuristic: Sonnet 5 owns the low-to-medium-effort majority of your workload; Opus owns the high-effort, high-precision minority. Most teams find the majority is bigger than they expected. And if you are eyeing the tier above both, my Fable 5 launch breakdown explains why that model belongs in an escalation seat, not the executor seat.
One non-benchmark note for anyone deploying agents with real permissions: Anthropic's pre-deployment evaluations found Sonnet 5 more resistant to prompt-injection hijacks and better at refusals than Sonnet 4.6, while showing somewhat higher misaligned-behavior rates than Opus 4.8. Same pattern as everything else — for the highest-stakes adversarial work, Opus keeps the edge; for the broad middle, Sonnet 5 is the better-behaved Sonnet you were already using.
The Thirty-Minute Experiment
Do not take my framework on faith, and do not take the benchmarks on faith either — they are directional, not gospel for your workload. Pick one task your system runs constantly. Run it three ways: Sonnet 5 at medium effort, Sonnet 5 at high, and Opus. Log output quality and token cost. You will find one of two things: medium-effort Sonnet 5 was already enough and you just cut that task's cost roughly in half, or the task genuinely needs Opus and you now have proof instead of a hunch. Either answer is worth thirty minutes.
The question is no longer which model. It is which point on the curve, task by task. Route that correctly and you run most of your work at a fraction of the cost with quality you cannot tell apart — while spending Opus money only where it buys you something. The cheap one wins more often than you think. Just never at the top of the dial.
If you are wiring multiple Claude tiers into one production system and want the per-task effort budgets dialed in before the invoice teaches you the hard way, that routing design is work I do for client teams — get in touch and bring the results of your thirty-minute experiment.
