Breaking

Sonnet 5.5 Matches a GPT-6 Astra Chip Zoom at One-Fifth the Token Price

September 29, 2026 at 18:48 EDT

Foundation Models
Software Dev & Coding

Claude Sonnet 5.5, run at its xhigh effort setting, and OpenAI's GPT-6 Astra, run in Ultra mode, both produced an interactive zoom that follows a Cerebras chip's wiring down to the micron scale from the same prompt. The list-price gap is exact. Anthropic charges $2 per million input tokens and $10 per million output tokens for Sonnet 5.5. OpenAI charges $10 and $50 for GPT-6 Astra, five times as much on both lines.

September 29, 2026 · Anthropic and OpenAI

Same chip zoom, five times the token price

Claude Sonnet 5.5 at xhigh and GPT-6 Astra in Ultra both followed a Cerebras chip's wiring to the micron scale. On the list rate, Sonnet costs one-fifth as much per token.

Input price per million tokens

Each block is $2. Output is the same multiple: $10 versus $50.

Sonnet 5.5

$10

GPT-6 Astra

5× on the rate card. Cache reads follow the same ratio, $0.20 versus $1.

Output tokens per index task, at max

Column height matches the token count. About 27,000 versus about 193,000.

27k

Astra, max

193k

Sonnet, max

About 7× the output tokens. That is why the cheaper rate can cost more per finished task: $7.60 versus $3.26.

5×

List price, input and output

56–53

Intelligence Index at max, Sonnet then Astra

$7.60

Sonnet cost per index task at max, versus $3.26

What held up

Both models built the micron-scale wiring zoom from one prompt. Sonnet's $2 and $10 rates are a fifth of Astra's, and at max it leads the index 56 to 53.

What the clip does not prove

Ultra uses subagents; xhigh does not. Viewers split on which render looked better. At max, Sonnet's task cost is higher because it writes far more tokens.

Read the settings before the winner

Same prompt, different controls. xhigh is one step below Sonnet's max. Ultra splits the job across subagents. This run did not publish tokens or time, so the smaller rate is not yet the smaller bill.

Continue reading

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