Breaking

Users Report Claude Sonnet 5 Consumes Tokens Faster Than Opus 4.8

June 30, 2026 at 17:21 EDT

Foundation Models
Software Dev & Coding
Sales & Business

Users are increasingly reporting that Anthropic's newest model, Claude Sonnet 5, released on June 30, 2026, burns through tokens faster in real use than the higher-tier Claude Opus 4.8. A new tokenizer is seen as the main cause, fueling the perception that Sonnet 5, despite its lower per-token price, can end up costing more in total.

June 30, 2026 · Anthropic

Claude Sonnet 5: cheaper to buy, but it can burn tokens faster

Anthropic's "most agentic Sonnet yet" delivers near-Opus 4.8 performance at a lower price tier — yet at high effort settings users report it can consume tokens quickly enough to cost more than Opus per task.

~1/3

intro price vs Opus 4.8 at the lower tier

1.0–1.35×

tokens used for the same text (new tokenizer)

1M / 128k

context / max output — same as Opus 4.8

Price per 1M tokens — Sonnet 5 vs Opus 4.8

Sonnet 5 shown at standard (post-intro) pricing · taller column = more expensive

OUTPUT $/M

$15

Sonnet 5

$25

Opus 4.8

INPUT $/M

Sonnet 5

Opus 4.8

Intro pricing (until Aug 31, 2026) drops Sonnet 5 to $2 / $10 — but a hungrier tokenizer means the per-token win can shrink on a cost-per-task basis.

Benchmarks — close, but Opus still leads

Terminal-Bench 2.1

Sonnet 5 · 80.4%

Opus 4.8 · 82.7%

SWE-bench Pro

Sonnet 5 · 63.2%

Opus 4.8 · 69.2%

OSWorld-Verified

Sonnet 5 · 81.2%

Opus 4.8 · 83.4%

The upside

Near-Opus 4.8 quality at ~one-third the price
Stronger long-running autonomous execution
Max Effort can surpass Opus on some tasks
An excellent everyday & enterprise workhorse

The catch

Consumes tokens rapidly at high effort
Can cost more than Opus depending on setting
More hallucinations reported
Opus often finishes complex tasks in one turn

The takeaway

A lower sticker price is not the same as a lower bill. Developers are increasingly separating "low price" from "actual consumption" — judging Sonnet 5 on cost-per-task, not cost-per-token.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.