Anthropic shipped Claude Sonnet 5 on June 30, positioning it as the go-to model for agentic work at a price that doesn't require a flagship budget. At $3 per million input tokens and $15 per million output tokens (after introductory pricing ends August 31), it's priced at roughly 60% of what Opus 4.8 costs. That sounds like an easy call. It's more complicated.
Where Sonnet 5 Actually Earns the "Agentic" Label
The benchmark numbers that matter for builders aren't the ones that get the most press. On SWE-bench Pro, Sonnet 5 scores 63.2% compared to Opus 4.8's 69.2%. That 6-point gap is real. For a coding agent doing open-ended software engineering, it matters.
But Terminal-Bench 2.1 tells a different story. Sonnet 5 scores 80.4%. Opus 4.8 scores 74.6%. That's the first time a mid-tier Sonnet has beaten its flagship sibling on a major coding benchmark, and the margin isn't narrow. For agents that work in the terminal, run shell pipelines, orchestrate Docker environments, or interact with databases, Sonnet 5 is the better model. And it costs less.
BrowseComp comes in at 84.7% as a single agent and 86.6% in multi-agent setup. OSWorld lands at 81.2%, just behind Opus 4.8's 83.4%.
The performance picture is nuanced: Sonnet 5 is weaker on general software engineering, roughly on par for computer-use, and genuinely stronger for terminal-based agents. If I were routing traffic today, I'd send CLI-heavy runs to Sonnet 5 and open-ended coding to Opus. The benchmarks support that split.
The Tokenizer Problem
Here's where the cost story falls apart.
Sonnet 5 uses a newer tokenizer that produces roughly 30% more tokens for the same text compared to older Claude models. The per-token rate is lower, but you're buying more tokens to do the same work. For one-shot tasks, this mostly cancels out. For agentic workflows, it compounds.
Think about what happens in a multi-step agent run. Step 1 generates output. That output goes into the context for step 2. More verbose output at step 1 means more input tokens at step 2. Step 2's output, also more verbose, feeds into step 3. The tokenizer inflation doesn't add cost locally and stop. It multiplies across every subsequent step in the chain.
Throw extended thinking into the mix and it gets worse. Each planning step can add thousands of thinking tokens on top of the extra output tokens. A workflow that's making tool calls, reasoning about results, and looping over a problem will consume far more than the per-token rate card implies.
The Actual Numbers
Artificial Analysis measured cost per task on their Intelligence Index. Sonnet 5 came in at approximately $2.29 per task. Opus 4.8 was around $2.00. Not a 40% discount. A 15% premium, for a model that's marketed as the affordable option.
On output tokens specifically, Sonnet 5 uses roughly 40% more than Sonnet 4.6 for equivalent tasks. On knowledge-work evaluations, it takes around 3x more agentic turns to complete the same work. The per-token rate looks attractive. The per-task cost doesn't.
This isn't a reason to avoid Sonnet 5. It's a reason to measure the right thing. Cost per token is a red herring. Cost per completed task, per agent, per feature is what you actually pay.
What to Actually Do
A few things I'd change in how I build given these numbers.
Route terminal-heavy agents to Sonnet 5. The Terminal-Bench 2.1 lead is real, and CLI tasks are less likely to trigger deep extended thinking chains, so per-task cost for those workloads stays reasonable.
For complex reasoning or open-ended software engineering, Opus 4.8's higher per-task completion rate may make it cheaper once you account for reduced retries and failed tool calls. I haven't verified this at scale, but the per-task numbers suggest it's worth measuring before you commit to a routing strategy.
Disable extended thinking on Sonnet 5 by default. That's where the token cost runs away. A verbose baseline model plus thinking tokens is a fast path to expensive. Only turn it on where the reasoning quality genuinely changes the outcome.
Set explicit token budgets per step. This matters for any model, but it matters more with Sonnet 5. An agent without per-turn token limits can get expensive silently, and the verbosity here means the budget burns faster than your old cost models expect.
One last thing: the introductory rate of $2/$10 ends August 31. If you're benchmarking your agent costs now, run the calculation against $3/$15, not the current rate. The model you evaluate this month is the one you'll pay more for in September.
Comments
Post a Comment