Claude Sonnet 5 vs Opus 4.8: When to Use Each

Claude Sonnet 5 vs Opus 4.8: When to Use Each

For about a year, my model-picking logic was lazy and it worked: if the task mattered, I reached for Opus. If it didn't, I reached for Sonnet and accepted that the output would be a notch worse. Two tiers, two prices, a clean mental shortcut.

Claude Sonnet 5 broke that shortcut on June 30, 2026, and I'm still adjusting.

Here's the short version, because you searched Claude Sonnet 5 vs Opus 4.8 and you deserve the answer before the scroll: Sonnet 5 lands within a couple of points of Opus 4.8 on most benchmarks, actually beats it on knowledge work, and costs 40% as much per input token. But there's a setting buried in the API — the effort level — that can quietly flip Sonnet 5 from "incredible bargain" to "more expensive than Opus for the same result." The whole decision lives in that detail, and almost nobody is talking about it.

So let's talk about it. Not the press release. The actual math, the trap, and the rule I now use to pick between them on a per-task basis.

What Anthropic actually shipped

Sonnet 5 is the new mid-tier model in the Claude family — the slot between the entry-level models and the heavyweight Opus and Fable/Mythos lines. On paper it's "just" a Sonnet update. In practice, the jump from Sonnet 4.6 is the largest single-generation gain I've seen in that tier, and the gap to Opus 4.8 shrank to something genuinely awkward for Anthropic's own pricing tiers.

It went live on launch day as the default model for Free and Pro plans, and it's available to Max, Team, Enterprise, and through the API. If you've used Claude on the web since the end of June and didn't change anything, you've probably already been talking to Sonnet 5 without knowing it.

The framing everyone reached for — TechCrunch, VentureBeat, The New Stack — was some version of "a cheaper way to run agents." That's true but it undersells the structural shift. For the first time, Sonnet and Opus sit on a single cost-performance curve instead of two separate tiers. The old "good model / great model" split is now "same model family, pick your point on the curve." That changes how you architect anything that makes more than a handful of model calls.

Before we get to the curve, look at where the numbers actually land. This is the part the marketing glosses over.

The benchmark picture, read honestly

Here's Anthropic's own comparison, and I want to read it the way I'd read it before staking a project on it — column by column, not as a hype reel.

Agentic coding, SWE-bench Pro: Sonnet 5 hits 63.2%, up from Sonnet 4.6's 58.1%, with Opus 4.8 still on top at 69.2%. This is the widest gap on the whole board — six points. If your workload is hard, multi-file, autonomous coding, this is the number that should give you pause. Opus is meaningfully better here, and we'll come back to why that six-point gap matters more than it looks.

Agentic coding, Terminal-Bench 2.1: Sonnet 5 jumps to 80.4% from 4.6's 67.0% — a thirteen-point leap in one generation — against Opus 4.8's 82.7%. Read those two sentences again. Sonnet 5 nearly caught Opus while leaving its own predecessor in the dust. That's the headline nobody framed correctly.

Multidisciplinary reasoning, Humanity's Last Exam: Without tools, 43.2% vs Opus's 49.8%. With tools, 57.4% vs Opus's 57.9% — basically a tie. The tool-use result is the tell: give Sonnet 5 the ability to search and call functions, and its reasoning gap against Opus nearly evaporates. Most real agent workloads do have tools. Keep that in your pocket.

Computer use, OSWorld-Verified: 81.2% vs Opus's 83.4%. Two points. For driving a browser or a desktop, that's noise on most tasks.

Knowledge work, GDPval-AA v2: Sonnet 5 scores 1618. Opus 4.8 scores 1615. The mid-tier model won. Not by a landslide, but it won — on the benchmark closest to "do useful professional work across disciplines." Let that sit for a second.

Step back and the shape is unmistakable: minimal drop-off from Opus almost everywhere, a clear win on knowledge work, and exactly one category — hard agentic coding — where Opus keeps real daylight. That's a remarkably narrow moat for a model that costs more than twice as much.

Which brings us to the money, because the benchmarks are only half the decision.

The pricing math that actually moves the needle

Benchmarks tell you what's possible. Pricing tells you what's affordable at scale. And this is where Sonnet 5 stops being interesting and starts being disruptive.

Per million tokens, here's the lineup as of launch:

• Fable 5 / Mythos 5: $10 input / $25 output — the premium ceiling
• Opus 4.8: $5 input / $25 output
• Sonnet 5: $2 input / $10 output

Read the Sonnet 5 row against Opus. Input is 40% of Opus's cost. Output is less than half. That's not a discount — that's a different budget category.

One detail the launch-day coverage flagged and you need to plan around: that $2/$10 is introductory pricing, running through August 31, 2026. After that it steps up to $3/$15. Anthropic set the intro rate so the transition for existing Sonnet users is roughly cost-neutral, but if you're building a forecast on these numbers, model the post-August rate. $3/$15 is still well under Opus's $5/$25 — it just isn't the firesale the first two months are.

Now make it concrete, because per-million-token numbers don't land until you scale them. The figure that reframed this for me: an output-heavy agent workload that costs roughly $1,000/day on Opus 4.8 lands near $400/day on Sonnet 5 at standard pricing. For a solo builder running one agent, who cares. For a team running hundreds of agents in parallel, that's the line between "this product has unit economics" and "this product is a science project."

This is the real story. Not that Sonnet 5 is good — plenty of models are good. It's that Sonnet 5 is good enough to make Opus look optional for the majority of workloads, while costing less than half as much to run them. For anyone who got priced out of Fable and Mythos, a capable Claude model just dropped into a budget they can actually defend.

But — and I cannot stress this enough — that 60% savings is conditional. There's a setting that can erase it entirely, and most teams won't notice until the invoice arrives.

The effort-level trap nobody warns you about

Here's the part that earns this article its existence. Sonnet 5 exposes effort levels — low, medium, high, and xhigh (extra high). Higher effort means the model spends more tokens reasoning before it answers. More reasoning tokens means better quality and higher cost. Output tokens are where the bill lives, remember, and reasoning burns output tokens.

If you've read my breakdown of Opus 4.8's effort levels, you already know this dial matters more than the model choice itself half the time. With Sonnet 5 it matters even more, because the dial is what determines whether you're getting a bargain or overpaying.

Walk through it the way the agentic-search results actually play out:

At low effort, Sonnet 4.6 still beats Sonnet 5 on pass rate — Sonnet 5 lands around 55% on the agentic-search task. Sonnet 5 is cheaper at this setting, so you're trading a little accuracy for a lower bill. Fine for forgiving, high-volume work.

At medium effort, the two are comparable on quality, but Sonnet 5 does it for less. This is the sweet spot — the setting where "cheaper and at least as good" is just true, no asterisk.

At high effort, Sonnet 5 pulls clearly ahead of 4.6 on quality. But the cost climbs, and here's the kicker: high-effort Sonnet 5 costs roughly the same as high-effort Opus. You've spent your way back up to Opus pricing.

At xhigh, Sonnet 5 maxes out at roughly Opus 4.8's medium-to-high setting on OSWorld-Verified and the BrowseComp agentic-search benchmark. Read that carefully. You are paying possibly more than Opus to get a result that merely matches Opus's middle gear. That's the trap. xhigh Sonnet 5 is, for many tasks, the worst of both worlds — Opus money for sub-Opus ceilings.

So the naive read — "Sonnet 5 is cheaper, crank it up and save money" — is exactly backwards at the top of the dial. The savings live at low and medium effort. Push to xhigh and you've often have built a slower, pricier Opus.

That asymmetry is the single most important thing to internalize about this model, and it's why the answer to "Sonnet 5 or Opus 4.8?" is never a flat one.

If you'd rather not tune this dial by hand across an agent fleet — figuring out per-task effort budgets is genuinely fiddly work — this is exactly the kind of architecture I build for clients through my Fiverr profile. But you can absolutely do it yourself, and the rule below is where I'd start.

So when does the cheaper model actually win?

Strip away the nuance and here's the decision framework I now use. It's not "which model is better." It's "what does this specific task need, and where does that put me on the curve."

Reach for Sonnet 5 when:

• The work is routine to moderately complex — content generation, classification, extraction, summarization, standard CRUD-flavored coding, most knowledge work. (Remember: it beat Opus on GDPval-AA v2.)
• You're running at volume. Hundreds of calls, parallel agents, anything where per-call cost compounds into a real number.
• Tools are in the loop. Sonnet 5's reasoning gap against Opus nearly closes once it can search and call functions.
• You can live at low or medium effort. This is where the cost advantage is real and unconditional.

Reach for Opus 4.8 when:

• The task is hard, autonomous coding — multi-file refactors, gnarly debugging, long agentic coding runs. That six-point SWE-bench Pro gap is real, and on genuinely difficult problems it compounds: Opus's token efficiency means it often reaches the answer in fewer expensive steps. On high-complexity agentic search and computer-use tasks, Opus frequently delivers better results at lower total cost than maxed-out Sonnet 5, precisely because it doesn't need xhigh to get there.
• Precision is non-negotiable and you'd rather not babysit the output.
• You were about to set Sonnet 5 to xhigh "to be safe." If you need that ceiling, just use Opus — it's cheaper at the top end and better.

The clean heuristic: Sonnet 5 owns the low-to-medium-effort majority of your workload; Opus 4.8 owns the high-effort, high-precision minority. Most teams will find the majority is bigger than they expected, which is exactly why this launch matters. For the broader agentic-coding picture specifically, I went deep in my hands-on look at Sonnet 5 for agentic coding — this post is the cost-and-decision companion to it.

There's one more dimension that doesn't show up on the benchmark chart at all, and for anyone deploying agents with real permissions, it might be the most important one.

What about safety, alignment, and security?

Cost-performance is the headline. But if you're handing a model tool access, file access, or a browser, "how does it behave when things get adversarial" stops being academic. Here's where Sonnet 5 lands, with the caveat included.

Anthropic's pre-deployment evaluations found Sonnet 5 is overall an improvement on Sonnet 4.6. On agentic safety specifically, it's better at refusing malicious requests and more resistant to hijack attempts in prompt-injection attacks — the exact failure mode that keeps me up at night when an agent has write access to anything. On the automated behavioral audit that probes for misaligned behaviors like deception and cooperation with misuse, Sonnet 5 scored safer than its predecessor.

The honest caveat: Sonnet 5 showed somewhat higher rates of misaligned behavior than the more capable Opus 4.8 and the Mythos Preview. More capability has, so far, correlated with better alignment in this family, and Sonnet 5 sits below Opus on capability. So if your use case is genuinely high-stakes and adversarial, that's another tally in Opus's column — not a dealbreaker for Sonnet 5, but a real factor.

On cybersecurity, the news is quietly good. Anthropic didn't deliberately train Sonnet 5 on cyber tasks. It handles routine, non-harmful cyber work but performs substantially worse than Opus 4.8 and Mythos 5 on dangerous capabilities like developing software exploits. That weakness is a feature here: Sonnet 5 doesn't carry the kind of exploit-development capability that led Anthropic to put heavier restrictions and scrutiny around Mythos (I wrote about that whole episode in the Fable 5 and Mythos 5 launch breakdown). The result is a model that's more straightforward and stable to deploy widely. It also shipped with real-time cyber safeguards on by default — the same ones Anthropic ran on Opus 4.7 and 4.8.

Net: for the vast majority of agentic work, Sonnet 5 is safe to deploy and better-behaved than the Sonnet you were using last month. For the highest-stakes, most adversarial scenarios, Opus still has the edge. Same pattern as everything else in this comparison — and it's no coincidence.

Why this launch is bigger than one model

Zoom out for a second, because the per-token argument misses the strategic point.

Anthropic is racing toward a major IPO, and Sonnet 5 reads like a model built for the API token market as much as for any individual user. The cost-performance ratio is the entire competitive battleground for inference right now, and Sonnet 5 is a direct play for it — capable enough to handle real production workloads, cheap enough to win bake-offs against rivals on price.

For the ecosystem, it fills the middle. There's now a genuinely strong Claude model sitting between the entry tier and the Opus/Fable/Mythos ceiling, aimed squarely at everyone who looked at Fable's $10/$25 and quietly closed the tab. That's a lot of people. A lot of startups. A lot of agents that suddenly pencil out.

And because Sonnet and Opus now share one cost-performance curve, the architecture conversation changes. You stop asking "which tier is this product?" and start asking "which point on the curve does each task need?" — routing cheap tasks to Sonnet 5 at low effort, reserving Opus for the hard minority. That's a more sophisticated, more cost-efficient way to build, and Sonnet 5 is the model that finally makes it practical.

The one thing to do this week

Don't take my framework on faith — and don't take the benchmarks on faith either. They're directional, not gospel for your workload.

Here's the 30-minute experiment. Pick one task your system runs a lot. Run it three times: Sonnet 5 at medium effort, Sonnet 5 at high effort, and Opus 4.8. Log the output quality and the token cost for each. You'll almost certainly find one of two things — either medium-effort Sonnet 5 is plenty and you just cut that task's cost by 60%, or the task genuinely needs Opus and you now have proof instead of a hunch. Either answer is worth thirty minutes.

Remember where this started: my lazy two-tier shortcut, "important task, reach for Opus." Sonnet 5 didn't kill that instinct — it sharpened it. The question is no longer which model. It's which point on the curve, task by task. Get that routing right and you'll run most of your work at 40% of the cost with quality you can't tell apart, while spending Opus money only where it actually buys you something.

The cheap one wins more often than you'd think. Just not at xhigh.

Let's Work Together

Looking to build AI systems, automate workflows, or scale your tech infrastructure? I'd love to help.

• Fiverr (custom builds & integrations): fiverr.com/s/EgxYmWD
• Portfolio: mejba.me
• Ramlit Limited (enterprise solutions): ramlit.com
• ColorPark (design & branding): colorpark.io
• xCyberSecurity (security services): xcybersecurity.io