Claude Fable 5.1 made me a very nice animated pelican
1st September 2026
At present is Claude Fable (and Mythos) 5.1 day. Anthropic say that Fable 5.1 “units a brand new normal for coding, information work, and long-running problem-solving duties”. Their announcement spends a notable period of time on scientific analysis, boasting of a 52.6% rating on the model new Terminal-Bench-Science 0.1 benchmark (first introduced on August twenty seventh), up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Different benchmarks present barely improved scores, however none as spectacular because the Science one.
However how properly can it pelican?
Again in July I wrote about how I used to be shedding religion within the pelican benchmark—its connection to how good the fashions had been at different duties didn’t appear to carry as strongly because it did again in 2025. Probably the most fascinating insights I get from it now are comparisons inside mannequin households, and significantly comparisons for a similar immediate at completely different reasoning effort ranges.
Fable 5.1 has 5 reasoning ranges: low, medium, excessive, xhigh, max—and no possibility to show off reasoning completely.
I fastened a difficulty in llm-anthropic which brought on reasoning traces to not be appropriately recorded, then ran some prompts.
Right here’s the complete set of pelicans for all the reasoning ranges, every with the complete reasoning transcript. I’ll replicate them right here:
Low and medium, each with out reasoning?
Subsequent, a little bit of a thriller. That is what I bought for effort low:
The transcript doesn’t present any summarized reasoning tokens, and the output token rely is 1,998. With Claude that output token rely contains reasoning tokens. It took 23.8 seconds and price 10.017 cents.
I bumped that as much as medium and bought this:
Weirdly, that one additionally exhibits no reasoning textual content and used 1,977 output tokens—21 tokens lower than low. It took 23 seconds and price 9.912 cents.
So for this explicit immediate (“Generate an SVG of a pelican driving a bicycle”) Fable 5.1 appeared to skip reasoning completely at each low and medium settings.
Excessive
Right here’s excessive—29.6 seconds, 2,612 output tokens, 13.087 cents:
This one did do a little bit of reasoning, abstract right here:
I’m planning the SVG structure for a pelican driving a bicycle, with a sky and floor background, a bicycle with two spoked wheels, body, seat and handlebars, and a white-bodied pelican with an extended neck and orange beak positioned on high.
Actually not a lot distinction from low and medium, although.
At xhigh issues bought radically completely different. 36,767 output tokens, 7 minutes 51 seconds, $1.83!
The reasoning hint is fairly prolonged, and contains particulars like this:
Including the attention, wings stretching all the way down to the handlebar grip, orange legs reaching to the pedals, and a small tail feather, whereas protecting the pelican deliberately outsized in comparison with the bike for comedian impact. […]
I’ll settle for the slight thickness as charming relatively than overengineering it.
Max
Setting effort to max gave me the perfect pelican I’ve seen from any of Anthropic’s fashions. 65,927 output tokens, 13 minutes and 54 seconds, $3.30:
There’s so much to love about this. The background is tasteful, the legs are clearly on both facet of the body, the ft are on the pedals, the wing is on the handlebars, the pelican has a cute blue hat and there’s a basket with a fish.
It’s nonetheless not exhibiting almost the identical degree of aptitude as Gemini 3.7 Flash, however I didn’t ask for aptitude—I requested for an SVG, and that’s what I bought.
Some highlights from that reasoning hint:
Including pedal shapes close to each ft, with the far foot on the second leg partially seen behind the body. I’m contemplating whether or not so as to add a small scarf or cap for additional character, however leaning towards protecting it easy to keep away from muddle.
Now I’m debating a bicycle helmet on the top versus the pelican’s signature crest—the beak and pouch already learn clearly as “pelican,” so a helmet may reinforce the bicycle theme with out shedding id, although it’d compete with the crest for visible area.
I notice the beak at (484,84) would overlap with the dome helmet, so I must shrink the helmet so it solely covers the highest of the top, adjusting its arc endpoints to sit down greater and narrower so the beak can connect cleanly on the entrance with out collision. […]
I’m including a darker tip area to characterize the first feathers, then reconsidering the trailing edge to incorporate scalloped feather curves as a substitute of 1 clean line for a extra pure look. […]
Now I’m checking the vent line placements on the helmet, ensuring they sit far sufficient contained in the helmet’s edge given the stroke width and rounded caps, and confirming every vent stays inside the helmet’s round boundary. […]
I determine skipping a handlebar bell and tire highlights since they’re pointless additions. Now I’m reconsidering the entrance fork’s curve — the present management level pulls the form backward when it ought to bow ahead for a correct rake, so I must shift the management level rightward to repair the fork’s lean.
OK, let’s animate it
On Hacker Information, swalsh commented on that Max pelican:
Now that it’s a solved benchmark, can we get the animated model?
I didn’t wish to spend one other $3 so I took the Max pelican and piped it into the default pondering degree of Excessive:
6,121 enter, 26,201 output = $1.37. The end result seemed like this, exported right here as video since some individuals have hassle viewing animated SVGs:
Your browser doesn’t assist HTML5 video.
The wheels within the video are rotating within the mistaken path, however I feel that’s an artifact of the conversion to MP4—they appear to be going within the appropriate path within the unique SVG.
