Brief IA

Anthropic Opus 4.7: Surge in Token Costs

🤖 Models & LLM·Tom Levy·

Anthropic Opus 4.7: Surge in Token Costs

Anthropic Opus 4.7: Surge in Token Costs
Key Takeaways
1Anthropic's Opus 4.7 consumes more tokens per request than 4.6, despite having the same listed price.
2Measurements from Abhishek Ray show a token consumption increase of up to 1.47x for certain content.
3A community assessment reveals a 37.4% rise in tokens and costs per request across 483 submissions.
💡Why it mattersUsers of Opus 4.7 may see their actual costs increase, impacting the budgets allocated for using this technology.
Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisis ton rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

📄
Full Analysis

Opus 4.7, the latest version of Anthropic's artificial intelligence model, has an intriguing feature: while its listed price remains unchanged from Opus 4.6, it consumes significantly more tokens per request. This increase was highlighted by Abhishek Ray during the Claude Code Camp, where he shared his observations on this new version.

The migration guide provided by Anthropic mentions an increase in token consumption ranging from 1.0 to 1.35x. Measurements taken by Ray confirm this trend, with figures falling within this range. However, certain types of content exceed these values: for instance, the actual content from Claude Code shows an average increase of 1.325x, while a CLAUDE.md file reaches 1.445x, and the technical documentation climbs up to 1.47x.

A community assessment, available on the site tokens.billchambers.me, has compiled data from 483 submissions, revealing a 37.4% increase in both tokens used and costs per request when transitioning from Opus 4.6 to version 4.7.

Abhishek Ray also noted that code experiences a more pronounced increase than prose, while texts in Chinese and Japanese appear to be less affected. For a sampling session of 80 turns, he estimates that costs could rise by 20 to 30%, raising the bill from $6.65 to a range between $7.86 and $8.76.

Despite this increase in consumption, Opus 4.7 offers notable improvements. A test using the IFEval benchmark on 20 prompts showed that Opus 4.7 adheres to strict instructions with a reliability that is five percentage points higher than its predecessor.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.