Breaking

Claude Sonnet 5 Trails Opus 4.8 in Animation Generation Test

July 1, 2026 at 15:45 EDT

Foundation Models
Software Dev & Coding
AI Agents

A comparison of Anthropic's new Claude Sonnet 5, released on June 30, 2026, against the higher-tier Opus 4.8 across four canvas-style animation tasks found that Opus 4.8 produced working animations in all four cases, while Sonnet 5 returned two of them as static images and, in one instance, cost more.

June 30, 2026 · Anthropic

Claude Sonnet 5: when the "cheaper" model finishes more expensive

Sonnet 5 promised "Opus-level performance at a lower price." But on real animation tasks it delivered fewer finished results — and in one case cost twice as much as the pricier Opus 4.8 while producing only a static image.

2 / 4

tasks animated by Sonnet 5 — the other two returned as static images

4 / 4

tasks fully animated by Opus 4.8

2×

cost of Sonnet 5's failed shredder task vs Opus 4.8's success

The cost reversal — paper-shredder task

Cost per attempt (US$). Taller column = more expensive.

$0.36

Sonnet 5

static image only

$0.18

Opus 4.8

working animation

Is "cheaper" still cheaper if the job isn't finished?

List pricing · per 1M tokens

Sonnet 5Opus 4.8

Input$2$5

Output$10$25

Sonnet 5 intro pricing until Aug 31 — then $3 input / $15 output.

Agentic coding benchmark

Opus 4.869.2%

Sonnet 563.2%

Gap narrows further at medium–high effort on OSWorld-Verified and BrowseComp.

Positive — early-access partners

Now completes multi-step agentic tasks
Cost-efficient at the right effort setting
Opus-tier features: ultrathink, self-checking

Critical — developers

Static-image output on canvas animations
Data-pipeline test: more hallucinations, 2×+ usage
New tokenizer may raise token counts 1.0–1.35×
Cybersecurity tasks heavily restricted by default

Continue reading

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