Anthropic just shipped Claude Sonnet 5, and the pitch is straightforward: the most agentic Sonnet model yet, performance close to Opus 4.8, and a price tag that looks like a bargain at $2 per million input tokens. But there's a catch buried in the tokenizer that nobody's talking about.

Claude Sonnet 5 announcement hero image

The model dropped June 30 as the new default for all Claude free and Pro plans. On paper, it's a clear step up from Sonnet 4.6. Agentic coding jumps from 58.1% to 63.2% on SWE-bench Pro. Terminal-Bench 2.1 goes from 67.0% to 80.4%, a 13-point leap that's hard to ignore. And on GDPval-AA, the knowledge work benchmark, Sonnet 5 scores 1618, edging past Opus 4.8's 1615. That last number is the one Anthropic keeps repeating. A mid-tier model beating the flagship on real-world professional output.

Here's where the numbers get interesting. The introductory pricing runs through August 31: $2 per million input tokens, $10 per million output. After that, it jumps to $3/$15, identical to Sonnet 4.6. So the "cheaper than Opus" claim has an expiration date.

But there's a bigger problem. Sonnet 5 ships with a new tokenizer. The same English text now produces roughly 30% more tokens than on Sonnet 4.6. Simon Willison ran the numbers: UDHR in English goes from 2,356 tokens to 3,341. Python code inflates from 44,014 to 56,113 tokens for the same 4,000 lines. Simplified Mandarin barely changes. The tokenizer seems optimized for CJK at the expense of Latin scripts.

The result is a cost inversion. Artificial Analysis measured $2.29 per task on their Intelligence Index using standard ($3/$15 pricing, about 15% more than Opus 4.8 at $5/$25. The promotional pricing masks this, but the math doesn't lie: Sonnet 5 burns more tokens per task, and the per-token price is the same as its predecessor. You're paying more to get the same work done.

Anthropic's framing is clever though. They're positioning agentic capability as the new baseline, not a premium feature. Sonnet 5 can plan, browse, use terminals, and run autonomously on tasks that "just a few months ago required larger and more expensive models." The effort slider adds another dimension: five levels: low, medium, high, xhigh, and max, with xhigh as the new addition. At max effort, Sonnet 5 uses about 40% more output tokens than Sonnet 4.6 and takes three times as many agentic turns on knowledge work evaluations.

The Zapier quote in the launch blog is telling. Daniel Shepard, a senior engineer, said they gave Sonnet 5 a two-part job: update Salesforce tiers, send a launch announcement, and it finished end to end. "That used to stall halfway." If that's reproducible, it's a real capability jump. The model doesn't just think harder; it follows through.

Community reaction on Reddit has been mixed. Some users report Sonnet 5 as less robotic and noticeably better than 4.6. Others are posting side-by-side comparisons showing it's slower, uses more session usage, and delivers worse results on certain tasks. The HN thread landed at 155 points with 63 comments, mostly focused on the cost math and how the model fits into existing workflows rather than the marketing language.

On the safety front, Anthropic says Sonnet 5 scores better on automated behavioral audits than Sonnet 4.6. Fewer hallucinations, less sycophancy, improved resistance to prompt injection. Real-time cyber safeguards are on by default. The trade-off: it's substantially less capable of dangerous cybersecurity tasks than Opus 4.8 or Mythos 5. That's deliberate. The model sits below the export control threshold, keeping it available without government restrictions.

The honest assessment: Sonnet 5 is the right default for most agentic workloads right now. It beats Opus on knowledge tasks, comes close on coding, and the introductory pricing makes it genuinely cheaper through August. But once that window closes, the tokenizer inflation means you're paying Opus-level prices for Sonnet-level capability on Latin-script tasks. The "cheapest agent model" framing is a time-limited truth.

For developers already running Claude Code, the switch is painless. Just change the model string to claude-sonnet-5. No SDK changes, no code modifications. The question isn't whether to try it. It's whether the August 31 pricing cliff changes your long-term architecture.

Sources