Fugu Ultra v2 vs GPT-5.5.
model vs model · prices as of
GPT-5.5 takes 2 of 5 rounds against Fugu Ultra v2. It is also the cheaper of the two to run, at $40.91 per million tokens against $45.
Sakana AIFugu Ultra v2
$45/M tokens
Sakana AI list price, nothing below it is accounted for · $4.80 per 1,000 pages
1M context · tool calling · reads images
List price in (lower wins)
$5/M
$5/M
List price out (lower wins)
$30/M
$30/M
Cheapest we can explain (lower wins)
none below list
$40.91/M at Poe
Standard price (lower wins)
$45/M
$45/M
Cost per 1,000 pages (lower wins)
$4.80
$4.36
Context
1M tokens
1.1M tokens
Open weights
not stated
no
Tool calling
yes
yes
Structured output
yes
yes
Reads images
yes
yes
Reasoning
yes
yes
Released
11 Sept 2026
23 Apr 2026
Providers selling it
5
49
Price we measure against
the maker's own
the maker's own
Trains on your prompts (no is better)
no provider to check
not stated
Providers selling both · 4
Kilo Gateway$45 / $45
Nous Portal$45 / $45
OpenRouter$45 / $45
Vercel AI Gateway$45 / $45
Fugu Ultra v2 then GPT-5.5, per million tokens
Go further
Add a third or fourth model, or switch the weighting, in the tool.
What moved in the last seven days
- Fugu Ultra v2Nous Portal$4 / $24 $5 / $3025% dearer ·
- GPT-5.5Nous Portal$4 / $24 $5 / $3025% dearer ·
Frequently asked
- Which is cheaper, Fugu Ultra v2 or GPT-5.5?
- GPT-5.5, at $40.91 per million tokens against $45 for Fugu Ultra v2, three parts input to one part output, read on 16 Sept 2026.
- Where is each one cheapest?
- Fugu Ultra v2: nobody goes below the maker's own price with a reason we can point at, so $45 is the number to beat. GPT-5.5: $40.91 at Poe.
- Can I buy both from one provider?
- Yes. 4 providers sell both: Kilo Gateway, Nous Portal, OpenRouter, Vercel AI Gateway. The cheapest of them for the two together is Kilo Gateway.
- Which holds more context?
- GPT-5.5, at 1.1M tokens against 1M.
Prices exclude tax and are per million tokens, read 16 Sept 2026 from each provider's own catalog. The standard price blends three parts input to one part output; a page is 600 tokens in and 60 out. Bars are scaled to the larger value in each row.