Sonnet 5 Is Here — And It Comes With a Hidden Billing Trap 💸

Claude's latest model just dropped overnight. Is your API budget still intact? 💀

Anthropic quietly shipped a new model. Fable — the one everyone's been waiting for — is still nowhere to be seen. Sonnet, however, leapfrogged straight to version 5, pitched at the old Opus performance tier and now set as the default on the web app.

For context, here are the same-bench Opus 4.6 / 4.7 numbers side by side 👇

Quick Take

Performance ranking: Opus 4.6 < Sonnet 5 < Opus 4.8.

After running several long-context tasks, it holds up well. That said, toggling the thinking budget on vs. off produces a significant IQ swing — worth keeping in mind.

Pricing: Same as last-gen Sonnet — roughly 40% of Opus cost.

…except it may still end up costing you more than Opus in practice. 🪤 (More on that below.)

Benchmark Comparison vs. Opus

Here's how Sonnet 5 stacks up against Opus based on official stats:

If you recall, Sonnet 4.6 was a standout release when it first launched. On official benchmarks, Sonnet 5 is a clean sweep upgrade over 4.6 — sitting just below Opus 4.8 overall, but actually pulling ahead on knowledge tasks.

According to Artificial Analysis, IQ landed at 53, ranking #4 overall — essentially neck-and-neck with Zhipu's GLM-5.2.

Availability & Pricing

Available now. Sonnet 5 is live across all tiers — free and Pro users get it as the default, and in Claude Code / API it's accessible as claude-sonnet-5. The context window remains at 1M tokens. Pricing matches last-gen Sonnet: $3 / $15 per MTok (input / output) after August 31.

⚠️ The Hidden Cost Trap — Read This Before You Scale

Sonnet 5 ships with a new tokenizer, meaning the same block of text gets split into more tokens — officially 1.0× to 1.35× depending on content type. This is why the "promotional" price sits exactly where it does. Anthropic's own language describes the switch as "roughly cost-neutral" — which is essentially an admission that the new model tokenizes more finely, and you can't rely on the per-token sticker price alone.

Here's what that actually means for your bill:

• New tokenizer → same text = 1.0–1.35× the token count
• After August 31, pricing increases 50% → $3 / $15 per MTok
• Artificial Analysis testing at full effort found Sonnet 5 burns ~40% more output tokens per task compared to the previous Sonnet

Bottom line: a single Sonnet 5 task can end up costing more than Opus 4.8. Factor this into your usage before scaling any automated workflows.