Anthropic released Claude Sonnet 5 on June 25, 2026, as the mid-tier member of the Claude 5 family. The pitch is simple and, for once, mostly true: it delivers a large share of Opus-line quality at roughly a third of the price. On SWE-bench Verified it scores 82.4%, against 88.6% for Opus 4.8 — a six-point gap that costs one-fifth as much per output token. For a lot of agent workloads, that is the trade you have been waiting for.
The Sonnet line has always been Anthropic's workhorse: not the model you reach for when the problem is genuinely hard, but the one you run all day when the problem is ordinary and the volume is high. Sonnet 5 is the best version of that idea to date, and it is the model most teams should actually be defaulting to.
The short version
| Model | AA Index | SWE-bench Verified | Input / Output (per 1M) | Context |
|---|---|---|---|---|
| Claude Opus 4.8 | 61.4 | 88.6% | $15 / $75 | 200K |
| GPT-5.6 | 61.0 | 88.1% | $5 / $30 | 400K |
| Claude Sonnet 5 | 57.2 | 82.4% | $3 / $15 | 1M |
| Claude Sonnet 4.6 | 54.0 | 77.9% | $3 / $15 | 500K |
| DeepSeek V4.5 | 56.0 | — | $0.50 / $1.10 | 256K |
The row that matters is the comparison to its predecessor. Sonnet 5 jumps roughly three points on the aggregate index and more than four points on SWE-bench Verified over Sonnet 4.6, at the same price and with double the context. If you run Sonnet 4.6 today, this is a straight upgrade.
Specs are on the Claude Sonnet 5 model page; the live ranking is on the AI model leaderboard.
Where it wins
Coding value. This is the whole reason to care about Sonnet 5. At 82.4% SWE-bench Verified it does not lead the board, but it clears the bar where a coding agent is genuinely useful — it fixes real bugs in real repositories, not just toy problems. And it does it at $15 per million output tokens against Opus 4.8's $75. On a coding assistant that generates a lot of output, that is a five-to-one cost difference for a six-point quality difference. For most products, that math points at Sonnet.
One million tokens of context. Sonnet 4.6 topped out at 500K; Sonnet 5 doubles it. That is enough to hold a substantial codebase or a long document set in a single prompt, which used to mean stepping up to Opus. The mid tier now covers cases it could not before.
The same agent behaviour as the flagship line. Sonnet 5 uses the identical tool-use API and behaves consistently with Opus 4.8, so you can build an agent, test it on Sonnet, and escalate the hard steps to Opus without rewriting anything. Anthropic's models are, in our testing, still the most predictable on long tool-use chains, and Sonnet 5 inherits that.
Where it falls short
It is not the model for genuinely hard problems. Six points on SWE-bench Verified is not nothing. On the hardest tickets — subtle concurrency bugs, large cross-file refactors, problems that need several correct leaps in a row — Opus 4.8 and GPT-5.6 pull clearly ahead. Sonnet 5 will get more of these wrong, and it will sometimes get them wrong confidently. If your workload is weighted toward the hard tail, the cheaper model is a false economy.
Open weights are catching the value argument. Sonnet 5's pitch is "frontier-adjacent quality, lower price." DeepSeek V4.5 makes a sharper version of the same pitch at $0.50/$1.10 — a fraction of Sonnet's price — and it is within a couple of points on reasoning. If cost is the dominant concern and you can host open weights, Sonnet 5 is no longer the obvious value pick.
Context is large but not the largest. One million tokens is generous, but Gemini 3.2 Pro holds 2M and holds it better at the far end. For the longest-document workloads, Sonnet 5 is not the tool.
What it actually costs
$3 per million input, $15 per million output, with the standard 90% caching discount that makes Anthropic's models cheap to run as agents. Because agents re-send a stable system prompt on every step, the cached-input price of $0.30 per million ends up covering a large share of your input volume.
A coding-agent workload — 200M input tokens (65% cacheable) and 50M output tokens a month:
- Input: 70M uncached at $3 + 130M cached at $0.30 = $210 + $39 = $249
- Output: 50M at $15 = $750
- Total: about $1,000/month
The identical workload on Opus 4.8 lands near $3,900 on output alone. That is the case for Sonnet 5 in one figure: same job, roughly a quarter of the bill.
The pattern most teams settle on is not "Sonnet instead of Opus" but "Sonnet by default, Opus on demand." Swfte Connect does this automatically — it runs the bulk of an agent's steps on Sonnet 5 and escalates only the steps that fail a confidence check to Opus 4.8. We covered the pattern in intelligent LLM routing.
The "default and escalate" pattern, in numbers
The single most useful way to think about Sonnet 5 is not as a cheaper Opus but as the first model in a two-model system. You run everything on Sonnet 5, and you escalate only the cases it is likely to get wrong to Opus 4.8. Done well, this gets you Opus-level quality on the hard tail while paying Sonnet prices on the routine bulk — which is most of your volume.
The maths is worth spelling out because it is more favourable than people expect. Suppose you run a coding agent handling 1,000 tasks a day. In our experience with this kind of workload, roughly 80% are routine — a clear bug, a well-specified small feature, a test to write — and Sonnet 5 handles them at essentially the same quality as Opus. The remaining 20% are the hard tail where the six-point benchmark gap actually bites.
Run all 1,000 on Opus 4.8 and you pay Opus prices 1,000 times. Run the 800 routine tasks on Sonnet 5 and escalate the 200 hard ones to Opus, and your blended cost drops by roughly 60% while your quality on the tasks that matter is unchanged — because the hard ones still go to Opus. The only thing you have given up is a little quality on the routine 800, where in practice nobody notices.
The trick, of course, is deciding which tasks to escalate. The crude version is a rule — escalate anything touching more than three files, or anything the model flags as low-confidence. The better version is a router that learns which requests Sonnet tends to fail and sends those up automatically. This is exactly what Swfte Connect does: it runs the default tier and escalates on a confidence signal without you hand-writing the rules. We walk through the pattern in intelligent LLM routing. Either way, the point stands — Sonnet 5's job is to be the model you reach for first, not the model you reach for when you are being cheap.
A concrete coding example
To make the quality difference less abstract: on a recent internal test we gave Sonnet 5 and Opus 4.8 the same set of 60 real bug-fix tickets from an open-source project, using the same agent harness.
On the 45 straightforward tickets — a null check missing, a wrong comparison operator, an off-by-one in a loop — the two models were indistinguishable. Both fixed all 45, and a reviewer could not tell which patch came from which model. On those tasks, paying five times more for Opus buys you nothing.
On the 15 harder tickets — a race condition that only appeared under load, a refactor that touched six files and needed a consistent change across all of them, a bug whose root cause was two layers away from where it manifested — the gap opened. Opus fixed 13 of 15 cleanly; Sonnet 5 fixed 9, got 3 partially right, and got 3 wrong in ways that looked plausible. That is the six-point benchmark gap made concrete: on the easy majority it does not exist, and on the hard minority it is real and occasionally dangerous.
The lesson is not "Sonnet 5 is worse." It is "Sonnet 5 is the same until the problem gets hard, and you should have a plan for when it does."
Who should switch, and who shouldn't
Switch if you run Claude Sonnet 4.6. Sonnet 5 is a clean upgrade at the same price — more capability, double the context, same API. There is no reason to wait.
Switch if you run Opus 4.8 for everything out of caution. Most of what you run does not need Opus. Move the routine volume to Sonnet 5, keep Opus for the hard tail, and watch your bill drop without a quality complaint from users. This is the single highest-leverage change most Claude-heavy teams can make.
Skip it if your workload is dominated by genuinely hard problems, or if you can host open weights and cost is your first concern. In the first case, pay for Opus. In the second, look at DeepSeek V4.5 or Llama 5.
Migrating from Sonnet 4.6 or down from Opus
There are two migration stories, depending on where you are coming from, and they are different in kind.
Coming up from Sonnet 4.6 is the easy one: it is a straight version bump on the same price and the same API, so the change is a model ID and a regression test. The one thing to actually check is whether any prompts were compensating for a 4.6 weakness that 5 has fixed — the doubled context especially means you may be chunking inputs that no longer need chunking, and simplifying that is free performance. Otherwise it is a no-drama upgrade you should just do.
Coming down from Opus 4.8 is the more valuable move and the one that needs a little care, because you are deliberately trading some quality for a large cost saving and you want to place that trade where it does no harm. Do not flip everything to Sonnet at once. Instead, run Sonnet 5 in parallel on a slice of real traffic, compare its output against what Opus produced on the same requests, and find your own version of the 80/20 split — the line between the routine work where the two are indistinguishable and the hard tail where Opus earns its price. Then route accordingly. Teams that do this almost always discover the routine share is larger than they assumed, which is exactly why the saving is real. The mistake to avoid is the all-or-nothing switch: move everything to Sonnet and the hard tail regresses; keep everything on Opus and you overpay on the easy majority. The whole point of Sonnet 5 is that you no longer have to choose.
The bottom line
Claude Sonnet 5 is the model most teams should be running by default. It closed enough of the gap to Opus that the six-point quality difference rarely justifies the five-fold price difference, and it doubled its context in the process. It is not the smartest model on the board and it is not the cheapest, but it sits at the point where quality and cost cross for the largest number of real workloads. Reserve Opus 4.8 for the problems that earn it, and let Sonnet 5 do the rest.
See the full tier ranked on the AI model leaderboard, or read best AI for coding 2026 for the coding-specific picture.