Breaking

Anthropic Unveils Claude Sonnet 5 as Users Flag Higher Real Costs at High Effort

July 1, 2026 at 09:31 EDT

Foundation Models
AI Agents
Software Dev & Coding

Anthropic launched Claude Sonnet 5, a model with strengthened agentic capabilities, on June 30, 2026, making it immediately available across its app, API and Amazon Bedrock. Yet within 24 hours of release, users began noting that, depending on settings, the higher-tier Opus 4.8 can be cheaper to run.

Jun 30, 2026 · Anthropic

Claude Sonnet 5: Cheaper Per Token — But Not Always Cheaper

A new mid-tier model promises near-flagship agentic performance at a lower rate. Yet a fatter tokenizer means that in high-load workloads it can cost more than Opus 4.8.

~30%

More tokens for the same text (1.0–1.35× via new tokenizer)

≈ Opus

Scores near flagship on BrowseComp & OSWorld even at medium effort

Default

Now standard for Free & Pro across Platform, Chat, Cowork & Code

Output price per million tokens (USD)

Lower unit rate — but higher token counts can erase the gap.

$10

Sonnet 5
intro→Aug 31

$15

Sonnet 5
std, from Sep 1

$25

Opus 4.8

Input rates: Sonnet 5 $2/$3 · Opus 4.8 $5 per million tokens.

The Catch

Cheaper per token ≠ cheaper per job.

Intro pricing was tuned to be roughly cost-neutral. But in high-effort, high-token scenarios, the ~30% extra tokens can outweigh the lower rate — pushing Sonnet 5 above Opus 4.8 at high effort.

Where it shines

Reliably finishes multi-step tasks prior Sonnet failed
Self-checks its own outputs
Strong price-performance in coding, legal research & data analysis
Solid on complex PRs and bug fixes

The caveats

Opus still wins on quality in some cases
Can feel overpriced vs expectations
Cybersecurity capability lower than Opus
Token bloat can flip the cost advantage

Continue reading

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