GPT-6.1 Sol vs Claude Sonnet 5.5: same price, 2.1x the output

GPT-6.1 Sol and Claude Sonnet 5.5 have identical prices on FastMetal: $2.11 per million input tokens and $10.55 per million output tokens. Price and context length won't pick one for you, so we ran both on the same 41 prompts and measured what came back.

The short version:

• Sonnet 5.5 produced about 2.1x the output tokens, and more on 40 of the 41 prompts. With the same per-token price, that is roughly 2.1x the output bill.
• Sonnet 5.5 was still faster. Median response time was 18.5 s for Sol and 15.2 s for Sonnet 5.5.
• The visible text only grew 1.12x. Most of the token gap is not in the answer you read.

How we measured

We used the 41 prompts from FastMetal's comparison set (20 in Japanese, 21 in English): debugging, code review, SQL optimization, business email, translation, Japanese proofreading, meeting-notes summaries and similar tasks. Both models got identical prompts on September 30, 2026, with default settings. Every figure below comes from the responses saved in that run.

This is one run of a small sample. Change the reasoning effort and the numbers move. Read it as a tendency, not a law.

Results

• Sonnet 5.5 used more output tokens on 40 of 41 prompts. The exception was the Japanese math word problem: 676 tokens for Sol, 666 for Sonnet 5.5.
• Sonnet 5.5 was slower on 16 prompts and faster on 25.
• Across all 41 answers, Sol wrote 136,292 characters and Sonnet 5.5 wrote 152,627. The token ratio is 2.14x; the character ratio is 1.12x. That works out to about 0.39 tokens per character for Sol and 0.75 for Sonnet 5.5.

At today's USD price, the output tokens alone for all 41 prompts come to about $0.57 for Sol and $1.21 for Sonnet 5.5 (input and cache not included), or roughly $0.014 and $0.030 per prompt. For a single call that's noise. For an agent that makes dozens of calls per task, it's a straight 2x.

The split held in both languages. On the 20 Japanese prompts, mean output was 1,590 tokens for Sol and 3,165 for Sonnet 5.5 (about 2.0x). On the 21 English prompts it was 1,038 against 2,454 (about 2.4x).

More tokens, yet faster

Twice the tokens should mean slower answers. It didn't, because Sol spends a long time on the prompts it thinks hard about:

• Pros and cons analysis (Japanese): Sol 55.7 s, Sonnet 5.5 30.3 s
• Pros and cons analysis (English): Sol 59.2 s, Sonnet 5.5 30.8 s
• SQL query optimization (Japanese): Sol 37.0 s, Sonnet 5.5 24.4 s

On short prompts Sol can be the quicker one. The nuance-preserving translation prompt took 3.1 s on Sol and 7.2 s on Sonnet 5.5. On average Sonnet 5.5 comes out slightly ahead, and that's all.

Where the extra output goes: annotations

Three of the Japanese prompts show the difference clearly.

Translating a Japanese product review into English. Sol returned the translation and nothing else: 118 tokens, 476 characters. Sonnet 5.5 added translator's notes: 797 tokens, 2,167 characters. One of the notes:

「贅沢な悩み」→ "luxury problem"

The source sentence says, roughly, "this may be a luxury problem." Sol rendered it as "but that's really just nitpicking." The meaning survives, but the "luxury" shade of the original doesn't. Sonnet 5.5 explained why it chose its wording. If you want only the translation, that's Sol; if you want the reasoning behind it, that's Sonnet 5.5.

Rewriting a long sentence about the CAP theorem for readability. Sol returned a three-sentence rewrite: 327 tokens. Sonnet 5.5 returned a rewrite, an explanation of the changes and a shorter alternative: 771 tokens. Its explanation included this:

元の「3つのうち最大2つしか満たせない」は俗説的な説明です。

That is, "the original's 'you can satisfy at most two of the three' is the folk version of the theorem." The prompt asked for a rewrite that kept the technical accuracy, and only Sonnet 5.5 went as far as correcting the accuracy of the original. It's also a note nobody asked for, so if you're pasting the result straight into a document, it's output you'll delete.

Proofreading a Japanese business notice. Here the two were close. Both fixed the same typo and the same overly casual verb form, and both took about 15 seconds (15.7 s and 15.2 s). The differences were the output size (924 against 2,110 tokens) and how each handled the ambiguous parts. Sol left one formal phrase as it was and said it was grammatically fine; Sonnet 5.5 rewrote it. Sonnet 5.5 also flagged that "2 a.m. on March 25" is ambiguous about which night it means, which Sol didn't mention.

What the vendors say

From each vendor's own pages, checked on September 30, 2026:

• Anthropic released Claude Sonnet 5.5 on September 28, 2026 and describes it as "more than 30% faster" than Sonnet 5. In our run the median went from 17.7 s on Sonnet 5 to 15.2 s on 5.5, about 14%. The conditions differ, so the numbers won't match. Anthropic: Introducing Claude Sonnet 5.5
• The same page says the default thinking effort is Medium in the Claude apps and High on the Claude Platform.
• OpenAI's API docs list GPT-6.1 Sol's default reasoning effort as Medium. OpenAI: GPT-6.1 Sol

This run can't tell you whether the output gap comes from the models themselves or from their default thinking settings.

Which one to use

• Default to GPT-6.1 Sol for high-volume work where you only want the answer: conversion, rewriting, summarization, extraction. Same price, about half the output. On these 41 prompts the output cost was $0.57 against $1.21.
• Choose Claude Sonnet 5.5 when you want the answer explained. Translator's notes, the reasons for each change, things you didn't ask about but should know. When a person reads the reasoning, as in review or proofreading, the extra tokens earn their keep. And in this run it was faster on both mean and median.
• We didn't grade quality. In the answers quoted here, both handled Japanese naturally, but we didn't score all 41. Try them on your own work.

Identical prices don't mean identical bills. The reliable test is to run a few dozen of your own prompts through both and look at the output token counts.

All 41 answers from both models are side by side on the comparison page. Details are on the GPT-6.1 Sol and Claude Sonnet 5.5 pages, and current prices on pricing. The API model ids are gpt-6.1-sol and anthropic-claude-sonnet-5-5.