Sonnet 5 vs Opus 4.8: Which Model Should You Use?
Sonnet 5 vs Opus 4.8: Which Model Should You Use?
Anthropic just shipped Sonnet 5, and the headline is that it competes with Opus 4.8 at less than half the price. Sonnet 5 costs $2 per million input tokens and $10 per million output — versus Opus at $5 and $25 — yet it lands within a couple of percentage points of Opus on most benchmarks. If you're running routine AI tasks and don't need bleeding-edge performance, Anthropic finally has a middle-tier option worth paying attention to.
Here's the thing: the interesting question isn't Sonnet 5 versus the old Sonnet 4.6 — you're basically always going to reach for 5 now. The real decision is Sonnet 5 versus Opus 4.8, and it's more nuanced than "cheaper model wins." Let me break down the numbers and where each model actually makes sense.
How Much Better Is Sonnet 5 Than Sonnet 4.6?
We hadn't had a Sonnet upgrade since 4.6, so this was overdue. Comparing Sonnet 5 to 4.6, it's a straight upgrade across the board — a big leap in agentic coding, multi-disciplinary reasoning, computer use, and knowledge work. There's no category where 4.6 is the smarter pick on raw capability.
That's why the 5-versus-4.6 debate is mostly settled. You're going to want Sonnet 5 in almost every case where you'd previously have used 4.6. The only wrinkle is effort levels and cost, which I'll get to — because "better" and "cheaper for a given result" aren't the same thing.
How Does Sonnet 5 Compare to Opus 4.8?
This is the comparison that actually matters. The question isn't whether Sonnet 5 is close to Opus 4.8 — it's how much of a fall-off you get. And honestly, it's not much.
Here's how the benchmarks shake out:
- Knowledge work — Sonnet 5 actually posts better numbers than Opus 4.8. That's a little crazy for the cheaper model.
- Computer use — only 2% behind Opus.
- Agentic coding — only 2% behind.
- Multi-disciplinary reasoning — a handful of percentage points behind.
- SWE-bench Pro — the biggest gap: 63% versus 69%.
- Terminal-Bench 2.1 — 80% versus 82%. Basically a tie.
So across most of what people actually use these models for, the gap between Sonnet 5 and Opus 4.8 is small — and on knowledge work it flips in Sonnet's favor. The performance conversation only makes sense in the context of pricing, which is where this gets interesting.
What Does Sonnet 5 Cost Compared to Opus and Fable?
Here's the full pricing picture across the current Anthropic lineup:
- Fable 5 (Mythos 5) — $10 per million input tokens, $25 output. That's double Opus on input.
- Opus 4.8 — $5 input, $25 output.
- Sonnet 5 — $2 input, $10 output.
That puts Sonnet 5 at just 40% of Opus's input cost and less than half its output cost. You're getting numbers that sit within a couple points of Opus on most tasks, for well under half the price.
This is the whole pitch. If you're not doing bleeding-edge work — if you're running more routine tasks and still want real power — Sonnet 5 plugs a gap in Anthropic's lineup that's been open for a while.
When Is Opus 4.8 Actually Cheaper Than Sonnet 5?
This is the part most people miss, and it's why the price-per-token comparison can be misleading. Cheaper per token does not always mean cheaper per task.
Look at agentic search performance broken down by effort level (Opus 4.8 vs Sonnet 5 vs Sonnet 4.6):
- Low effort — Sonnet 5 hits about 55% pass rate. But low effort on Sonnet 4.6 actually performs better. Low on Sonnet 5 isn't about performance — it just means "this is going to be super cheap."
- Medium effort — Sonnet 5 gets essentially the same performance as Sonnet 4.6, but at a significant discount.
- High effort — this is where Sonnet 5 finally surpasses 4.6, and it does it at roughly what low effort used to cost on 4.6.
But here's the catch: once you push Sonnet 5 to high effort, you're paying about what Opus 4.8 costs on medium and high — and Opus posts a better pass rate at that price. So if you need high-effort quality on a hard problem, the cost advantage of Sonnet can evaporate.
The deeper reason is token efficiency. Opus 4.8 is so token-efficient that on complex problems it can end up cheaper overall, even at a higher per-token rate, because it solves the task in fewer tokens and fewer steps. Sonnet just isn't always the right tool for a genuinely hard job.
Should You Use Sonnet 5 or Opus 4.8 for Agentic Computer Use?
Computer use tells a different story than search — and it's more favorable to Sonnet 5. Across the board, Sonnet 5 beats Sonnet 4.6 on agentic computer use, and it does it at a low cost. Unlike agentic search, there's no effort level where 4.6 wins, so here you'd reach for Sonnet 5 every time over the old model.
But Opus still sets the ceiling. Opus High outperforms Max Sonnet 5 on computer use — and it's cheaper doing it. That's the recurring theme: just because Opus's per-token price is higher doesn't mean it's the more expensive choice. For the hardest tasks, Opus reaches performance levels Sonnet simply can't, at a cost that's competitive because of how efficiently it gets there.
Is Sonnet 5 Safe? What About Misaligned Behavior?
It wouldn't be an Anthropic release without a look at misaligned behavior. Sonnet 5 is an improvement over prior Sonnet models, but it still sits below Opus 4.8 and the Mythos preview on alignment. So on the safety scorecard, it's better than what came before in its class but not best-in-lineup.
On the security side, there's good news: the cybersecurity exploit concerns that got the Mythos preview effectively locked up are not something you need to worry about with the Sonnet class. That whole category of risk isn't in play here.
That said, I wouldn't over-index on these safety charts either. For most builders, the practical decision is capability-per-dollar, not the alignment delta between model tiers.
Who Is Sonnet 5 Actually For?
The clearest win is for people building on the API rather than living inside the Claude Max plan. For anyone paying per token for their applications, Anthropic has historically been brutally expensive — expensive enough that I've told people for a long time to go look at OpenAI's mini models, because those are often more than enough for a lot of jobs.
Sonnet 5 changes that math. It gives Anthropic a genuine middle-tier option — real power, but not Opus pricing — for the routine, lower-complexity workloads that make up most production usage. If you were skipping Anthropic on cost alone, this is the release that brings you back into the conversation.
My honest take: it's going to be a lot of trial and error figuring out where Sonnet 5 makes the most sense versus Opus. What I'd love to see from Anthropic is a clearer line of demarcation — here's where Sonnet 5 does the job cheaper, and here's where the task gets complex enough that Opus is actually the more efficient pick. Until then, treat it as a case-by-case call.
Frequently Asked Questions
Is Sonnet 5 better than Opus 4.8?
Not overall — but it's close. Sonnet 5 lands within about 2% of Opus 4.8 on computer use and agentic coding, and actually beats it on knowledge work. The biggest gap is SWE-bench Pro at 63% versus 69%. For most routine tasks the difference is small enough that Sonnet 5's price advantage wins; for genuinely hard problems, Opus is still the stronger and often more efficient choice.
How much does Sonnet 5 cost?
Sonnet 5 is $2 per million input tokens and $10 per million output tokens. That's 40% of Opus 4.8's input cost ($5) and less than half its output cost ($25). Compared to Fable 5 (also called Mythos 5) at $10 input and $25 output, Sonnet 5 is dramatically cheaper.
Should I always use Sonnet 5 over Sonnet 4.6?
For raw capability, yes — Sonnet 5 is a straight upgrade over 4.6 in agentic coding, reasoning, computer use, and knowledge work. The only nuance is effort levels: on low effort, Sonnet 4.6 can still edge out Sonnet 5 on agentic search, but at that point you're using Sonnet 5 low effort mainly because it's cheap, not because it's the top performer.
When is Opus 4.8 cheaper than Sonnet 5?
On complex tasks. Opus 4.8 is token-efficient enough that it can solve a hard problem in fewer tokens and steps, making it cheaper overall despite the higher per-token price. Push Sonnet 5 to high effort on a difficult task and you end up paying roughly Opus-level costs anyway — but with a lower pass rate.
Is Sonnet 5 available now?
Yes. Sonnet 5 is available everywhere across Anthropic's platforms and API. Expect a period of trial and error as the community figures out exactly which workloads it handles best relative to Opus 4.8.
If you want to go deeper on picking the right Claude model for your builds — and getting the most out of Claude Code — join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance and my Claude Code Masterclass.
