Claude Sonnet 5 Is Here, and Its New Tokenizer Quietly Raises the Price You Actually Pay

TL;DR

Anthropic just released Claude Sonnet 5, the new default model on Free and Pro plans and the API identifier claude-sonnet-5. It lands performance "close to Opus 4.8, but at lower prices," ships a 1 million token context window with up to 128K output tokens, and turns adaptive thinking on by default. The headline rate card is the same as Sonnet 4.6: $3 per million input tokens and $15 per million output, with an introductory discount to $2 / $10 through August 31, 2026. So far, so good. The catch is buried in the release notes: Sonnet 5 uses a new tokenizer that counts the same text as roughly 1.0x to 1.4x more tokens. Same words, more billable units. Your per-token price did not move, but your per-task bill did.

What Actually Shipped

The model itself is a straight upgrade over Sonnet 4.6 on reasoning, tool use, coding, and general knowledge work, and Anthropic positions it as the value pick that gets you most of the way to Opus 4.8 for a fraction of the cost. On agentic search (BrowseComp) and computer use (OSWorld-Verified), Anthropic says Sonnet 5 opens up a wider cost-performance band than 4.6 and reaches into territory previously reserved for Opus. On coding benchmarks it sits just behind Opus 4.8 while costing far less, which is exactly the slot Sonnet has always been built to fill.

It is live across the Claude apps, Claude Code, and the API today, and it is the new default for the free tier, so most casual users are already talking to it whether they noticed or not. Safety-wise, Anthropic reports lower misaligned-behavior rates than Sonnet 4.6 and keeps cyber safeguards on by default. All of that is the normal, expected shape of a point release. The interesting part is the meter.

The Tokenizer Is the Story

A tokenizer is the thing that chops your text into the units a model actually charges for and reasons over. Swap it, and the exact same prompt can resolve into a different number of tokens. Sonnet 5 ships a new one, and independent testing by Simon Willison put real numbers on the inflation: English text came out about 1.4x more tokens, Spanish 1.33x, Python 1.28x, and Mandarin essentially flat at 1.01x. Anthropic's own framing is gentler, "approximately 1.0 to 1.35x depending on content," but the direction is the same: the meter spins faster now.

Here is the analogy that makes it click. Imagine your taxi keeps the exact same fare per mile, then quietly swaps in an odometer that ticks 1.4 miles for every actual mile you drive. The rate card on the window is honest. The ride still costs more. That is a tokenizer change in one sentence.

Why do it at all? A tokenizer that splits text more finely can give the model cleaner units to reason over, which helps quality. Anthropic explicitly frames the change as a performance improvement, and that is plausible. But it also means the listed price and the felt price have quietly decoupled, and the gap is biggest for exactly the English prose and code that most of this blog's readers are pushing through the API all day.

So Is It Cheaper? Do the Math

This is where you have to be careful, because two true statements are pulling in opposite directions. Statement one: the rate card is unchanged from Sonnet 4.6 (and the intro discount makes it look cheaper still). Statement two: the same job now consumes around 30% more tokens on typical English and code workloads. Net those out at standard pricing and a task that cost you a dollar on Sonnet 4.6 lands closer to $1.30 on Sonnet 5, for identical input, before you have improved a single thing.

The honest counterweight: this is a better model, so the right comparison is not "same task, more tokens" but "what would matching Sonnet 5's quality have cost you before?" If you were reaching for Opus 4.8 to get this level of coding or agentic performance, Sonnet 5 is a real discount even after the token inflation, because Opus list pricing is several times higher. The verdict flips entirely on whether you are downgrading from Opus (Sonnet 5 saves you money) or upgrading from Sonnet 4.6 (Sonnet 5 costs you more for the same work). Know which move you are making before you swap the model string.

What You Should Actually Do

If you run anything at volume on the API, do not trust your old token math. Re-measure. Pull a representative sample of your real traffic, run it through Sonnet 5, and look at the actual token counts rather than assuming parity with 4.6. Your cost dashboards, budget alerts, and any per-request token caps were calibrated on the old tokenizer, and the worst time to discover a 30% drift is on the invoice.

• Context-window planning: the same 1M window now holds proportionally less of your English and code, because each chunk eats more tokens. If you were running near the edge, you are closer to it than the number suggests.
• Budgets and rate limits: anything expressed in tokens (spend caps, max-output limits, alert thresholds) silently got tighter. Recheck them.
• Multilingual workloads: if your traffic is heavy in Mandarin or similar, the tokenizer change barely touches you, and Sonnet 5 is close to a free upgrade. The pain is concentrated in English and code.
• Grab the intro pricing: the $2 / $10 rate runs through August 31, 2026. After that it is $3 / $15, so any cost modeling you do should assume the standard rate, not the honeymoon one.

None of this makes Sonnet 5 a bad deal. It is a genuinely stronger model at a price that is still aggressive against Opus. It just means the "same price as Sonnet 4.6" line you will see repeated everywhere is, politely, doing some load-bearing work it has not earned.

Key Takeaways

• The release: Claude Sonnet 5, the new default model, API id claude-sonnet-5, with a 1M context window and adaptive thinking on by default.
• The quality: a clear step up from Sonnet 4.6, positioned as close to Opus 4.8 on coding, agentic search, and computer use at much lower cost.
• The catch: a new tokenizer counts the same text as roughly 1.0x to 1.4x more tokens (about 1.4x English, 1.28x Python, 1.01x Mandarin).
• The pricing: rate card is unchanged at $3 / $15, with $2 / $10 introductory pricing through August 31, 2026, but the effective per-task cost rises with token count.
• The verdict: a discount if you are stepping down from Opus 4.8, a stealth price increase if you are upgrading from Sonnet 4.6.
• The action: re-measure your real token usage and recalibrate budgets, alerts, and context limits before the next invoice does it for you.

Sources: Anthropic: Claude Sonnet 5, Simon Willison: What's new in Claude Sonnet 5, The Next Web