Matt Wolfe compares two same-day model launches: Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Sol and Luna. He weighs their capabilities against the price of completing useful work.

Cheaper tokens can still produce a large bill (00:21–05:01)

  • Wolfe reports Opus 5.5 API prices of $4 per million input tokens and $20 per million output tokens. These are recording-time prices.
  • In the Artificial Analysis comparison he shows, Opus 5.5 uses about 119,000 output tokens per task, versus 78,000 for Fable 5.1.
  • The displayed weighted cost per Artificial Analysis Intelligence Index task is $5.98 versus $7.63. More generated text absorbs part of the lower token price.
  • He also shows different reasoning settings with different costs and scores. The most expensive setting does not automatically win every test.

The animation demos show possibility, not repeatability (05:01–09:05)

  • Wolfe collects JavaScript animations, a playable handheld-console interface, games and a Blender animation with a claymation look.
  • Two demos come from people he identifies as Anthropic employees. He does not know how the Blender example connected the model to the application.
  • These are selected examples from other creators. Their polish does not reveal the number of attempts, edits or failed runs behind them.

OpenAI’s launch targets a cheaper tier (09:05–14:22)

  • Wolfe reports $2/$10 per million input/output tokens for GPT-6 Sol and $0.10/$0.50 for Luna.
  • He presents Sol as a lower-cost option below Astra’s capability tier. The displayed comparison puts Sol at $1.06 per Intelligence Index task, with roughly 31,000 output tokens.
  • He finds fewer public demos for OpenAI’s launch. He suspects a smaller early-access rollout, but offers no evidence that this explains the difference.

A model judge can disagree with the viewer (14:22–15:17)

Wolfe’s own image-generation exercise puts Sol first and Opus 5.5 eighth when another model grades the results. His visual preference differs from that ranking. A narrow creative prompt and an automated judge can produce a different answer from a broad benchmark suite.

Caveat: this is a launch-day comparison with selected demos and reported scores, not a controlled test of production software.