Anthropic released Claude Sonnet 5, which ranks fifth in independent benchmarks but consumes roughly 40 percent more tokens per task than its predecessor, nearly doubling the real cost to $2.29 per task despite keeping official token prices unchanged. This marks the second time Anthropic has raised effective prices without changing published rates; the earlier Opus 4.7 launch saw a tokenizer change inflate token counts by approximately 30%, and developers measured a 1.325x to 1.47x cost increase. For businesses choosing between AI models, the hidden cost creep makes transparent pricing harder to compare, particularly against cheaper competitors in the mid-range segment.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Anthropic released Claude Sonnet 5, which ranked fifth in Artificial Analysis's Intelligence Index with 53 points, but the model consumes roughly 40 percent more output tokens per task at maximum performance than its predecessor Sonnet 4.6, and runs about three times as many agent loops on knowledge work benchmarks. Despite keeping official token prices flat at $3 per million input tokens and $15 per million output tokens, the real cost per task nearly doubled from about $1.20 to $2.29—now more expensive than the pricier Opus 4.8 at $1.97 per task.
Why it matters
Anthropic has repeated this pattern before; when Opus 4.7 launched, a new tokenizer inflated token counts by approximately 30%, and developer measurements found a 1.325x to 1.47x cost increase. With Sonnet 5, the combination of higher token consumption and more agentic behavior compounds the hidden cost creep. For developers and businesses evaluating AI models, real per-task pricing is now substantially higher than the official rate card suggests, making model selection less transparent—especially compared to competitors like Deepseek V4 Pro and GLM-5.2 that offer competitive performance at a lower cost in the mid-range segment.
What to watch
Anthropic is running a promotional rate of $2 or $10 per million tokens through September 1, but analyses cited in the article are based on regular prices. Sonnet 5 still lags larger models on reasoning-heavy benchmarks like CritPt (17 percent), where it scored 14 points above Sonnet 4.6 but below GLM-5.2, Claude Opus, Fable, and GPT-5.5 in their higher configurations.
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
