Claude Sonnet 5: cheaper agents on paper, until you count the new tokenizer's tokens

Anthropic shipped Claude Sonnet 5 on June 30, and made it the default model for free users the next day. The pitch is blunt: near-Opus-4.8 quality for running agents, at roughly half the price. For anyone paying per token to run browser-and-terminal agents at scale, that’s the number that matters.

But there’s a catch buried in the fine print that Simon Willison caught within hours. Sonnet 5 ships with a new tokenizer, and it chops English into more pieces than the old one. Same prose, more tokens, higher bill, even though the per-token sticker price went down. So the honest version of the headline is “cheaper, mostly, if you measure it on your own prompts first.”

What Sonnet 5 actually is

Sonnet is Anthropic’s middle tier, the workhorse between the small Haiku models and the flagship Opus line. Sonnet 5 is the newest of those, and Anthropic calls it “the most agentic Sonnet model yet”: it can make plans, drive tools like browsers and terminals, and run autonomously for stretches that used to demand a pricier Opus-class model.

That “agent” framing is the whole story. An agent here means the model doesn’t just answer a prompt, it loops: read a task, call a tool, read the result, decide the next step, repeat. Each loop burns tokens, and a real task might run dozens of loops. So the cost of running agents is dominated by the price per token multiplied by how many tokens the model chews through. Drop the tier that can do the job and you drop the bill, sometimes by a lot.

Sonnet 5 is live now across free, Pro, Max, Team, and Enterprise plans, plus Claude Code and the API. TechCrunch reports it became the default model for free and Pro accounts starting July 1. That last part is the quiet flex: free-tier users now get Anthropic’s most capable agent model without paying anything.

The benchmark and pricing story

Start with price, because that’s the sell. Through August 31, Sonnet 5 runs at $2 per million input tokens and $10 per million output, according to Anthropic’s announcement. After the promo ends it settles at $3 input and $15 output. Compare that to Opus 4.8 at $5 and $25, per BleepingComputer, and the standard Sonnet rate is about 40% cheaper on both ends. TechCrunch notes it also undercuts GPT-5.5 and Gemini 3.1 Pro, while sitting above Gemini 3.5 Flash.

Now the quality. Anthropic’s framing is that Sonnet 5 lands “close to” Opus 4.8, and the benchmarks back a nuanced version of that. On agentic coding, TechCrunch cites Sonnet 5 at 63.2% against Opus 4.8’s 69.2%, up from 58.1% for Sonnet 4.6. So it closed most, not all, of the gap. On knowledge work it slightly beats Opus 4.8. On agentic search (BrowseComp) and computer use (OSWorld-Verified), Anthropic says Sonnet 5 matches Opus-4.8 capability at higher effort levels while costing far less to get there.

Read those numbers together and the tradeoff is clear. If your workload is pure hard coding, Opus 4.8 still wins outright, and 6 points is a real gap. For most agent work, retrieval, tool orchestration, day-to-day automation, Sonnet 5 gets you close enough that the price cut is the deciding factor. Daniel Shepard of Zapier, quoted by TechCrunch on a task that used to fail, put it plainly: “That used to stall halfway. For day-to-day automation, it’s a no-brainer.”

Sonnet 5 also picks up the usual safety improvements: Anthropic says it’s better at refusing malicious requests, more resistant to prompt injection, and lower on hallucination and sycophancy. For anyone wiring a model into MCP tools that can touch real systems, injection resistance isn’t a nice-to-have.

The tokenizer catch

Here’s where the clean pricing story gets muddy. A model doesn’t read characters, it reads tokens, and the tokenizer decides how text gets sliced into them. You pay per token. Change the tokenizer and you change how many tokens the same text costs, which changes your bill even if the price per token never moves.

Sonnet 5 has a new tokenizer, and it slices English finer. Anthropic’s own docs note the same input maps to roughly 1.0 to 1.35x more tokens depending on content type. Simon Willison ran the numbers against real documents and found the English end of that range is where most people live. Testing the Universal Declaration of Human Rights, he watched 2,356 tokens become 3,341. His summary: “The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6”, which he called “effectively a 30% price increase.”

Broken out by language, Willison found “the new token is roughly 1.4x times more expensive for English, 1.33x for Spanish, 1.28x for Python code and effectively the same cost for Simplified Mandarin.” So an English-heavy agent that reads and writes a lot of prose could see its input costs climb by 40% per equivalent chunk of text, exactly the window where the standard-rate price cut also sits at 40%. The two can cancel. On the promo rate through August, you’re still ahead; on the standard rate afterward, an English workload might land close to where Sonnet 4.6 was.

This isn’t a gotcha aimed at Anthropic so much as a warning about how model pricing gets compared. List prices per million tokens look like apples to apples across models and versions. They aren’t, because a token isn’t a fixed unit of text. A tokenizer that packs English more densely lowers your real cost at the same sticker price; one that splits it finer raises it. The only honest comparison runs your actual prompts through each tokenizer and counts the tokens you’ll actually be billed for. That’s a step almost nobody takes when a shiny new model drops with a lower headline number, and it’s exactly the step Sonnet 5 rewards.

Why you’re hearing about this now

The timing isn’t an accident. Anthropic filed confidential IPO paperwork with the SEC on June 1, with a public listing expected as early as October and bankers treating a trillion-dollar debut as the base case. The company’s revenue reportedly went from $1 billion annualized at the end of 2024 to a run rate in the tens of billions by mid-2026, and Claude Code alone is a large chunk of that. When you’re about to sell public investors on AI-boom growth, pushing your most agentic model to every free user and undercutting rivals on price is exactly the kind of adoption story that reads well in an S-1.

For developers, the IPO backdrop mostly means one thing: expect the aggressive pricing and the free-tier default to stick around, because Anthropic needs the usage numbers.

What this means for you

If you build agents, Sonnet 5 is worth testing this week, especially while the $2/$10 promo runs through August. For high-volume automation, retrieval, and tool orchestration, it gets close enough to Opus 4.8 that the cost gap is hard to argue with. For heavy code generation, keep Opus 4.8 in the mix; the 6-point coding gap is real. And before you commit, run your actual prompts through both tokenizers and compare token counts, not list prices. Willison’s finding means an English-heavy workload could quietly eat the entire discount once the promo ends. The model got cheaper. Whether your bill does depends on what you feed it.