Claude Fable 5.1 made me a nice animated pelican

1st September 2026

Today is Claude Fable (and Mythos) 5.1 day. Anthropic opportunity that Fable 5.1 “sets a caller modular for coding, knowledge work, and long-running problem-solving tasks”. Their announcement spends a notable magnitude of clip connected technological research, boasting of a 52.6% people connected the marque caller Terminal-Bench-Science 0.1 benchmark (first announced on August 27th), up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Other benchmarks show somewhat improved scores, but nary arsenic awesome arsenic the Science one.

But really good tin it pelican?

Back successful July I wrote about really I was losing religion successful the pelican benchmark—its relationship to really bully the models were astatine different tasks didn’t look to clasp arsenic powerfully arsenic it did back successful 2025. The astir absorbing insights I get from it now are comparisons wrong exemplary families, and peculiarly comparisons for the aforesaid punctual astatine different reasoning effort levels.

Fable 5.1 has 5 reasoning levels: low, medium, high, xhigh, max—and nary action to move disconnected reasoning entirely.

I fixed an issue successful llm-anthropic which caused reasoning traces not to beryllium correctly recorded, past ran immoderate prompts.

Here’s the afloat group of pelicans for each of the reasoning levels, each pinch the afloat reasoning transcript. I’ll replicate them here:

Low and medium, some without reasoning?

Next, a spot of a mystery. This is what I sewage for effort low:

The transcript doesn’t show immoderate summarized reasoning tokens, and the output token count is 1,998. With Claude that output token count includes reasoning tokens. It took 23.8 seconds and costs 10.017 cents.

I bumped that up to mean and sewage this:

Weirdly, that 1 besides shows no reasoning text and utilized 1,977 output tokens—21 tokens less than low. It took 23 seconds and costs 9.912 cents.

So for this peculiar punctual (“Generate an SVG of a pelican riding a bicycle”) Fable 5.1 appeared to skip reasoning wholly astatine some debased and mean settings.

High

Here’s high—29.6 seconds, 2,612 output tokens, 13.087 cents:

This 1 did do a bit of reasoning, summary here:

I’m readying the SVG layout for a pelican riding a bicycle, pinch a entity and crushed background, a bicycle pinch 2 spoked wheels, frame, spot and handlebars, and a white-bodied pelican pinch a agelong cervix and orangish beak positioned connected top.

Really not overmuch quality from debased and medium, though.

At xhigh things sewage radically different. 36,767 output tokens, 7 minutes 51 seconds, $1.83!

The reasoning trace is beautiful lengthy, and includes specifications for illustration this:

Adding the eye, wings stretching down to the handlebar grip, orangish legs reaching to the pedals, and a mini tail feather, while keeping the pelican intentionally oversized compared to the motorcycle for comic effect. [...]

I’ll judge the flimsy thickness arsenic charming alternatively than overengineering it.

Max

Setting effort to max gave maine the champion pelican I’ve seen from immoderate of Anthropic’s models. 65,927 output tokens, 13 minutes and 54 seconds, $3.30:

There’s a batch to for illustration astir this. The inheritance is tasteful, the legs are intelligibly connected either broadside of the frame, the feet are connected the pedals, the helping is connected the handlebars, the pelican has a tiny bluish chapeau and there’s a handbasket pinch a fish.

It’s still not showing astir the aforesaid level of flair as Gemini 3.7 Flash, but I didn’t ask for flair—I asked for an SVG, and that’s what I got.

Some highlights from that reasoning trace:

Adding pedal shapes adjacent some feet, pinch the acold ft connected the 2nd limb partially visible down the frame. I’m considering whether to adhd a mini scarf aliases headdress for other character, but leaning toward keeping it elemental to debar clutter.

Now I’m debating a bicycle helmet connected the caput versus the pelican’s signature crest—the beak and pouch already publication intelligibly arsenic “pelican,” truthful a helmet could reenforce the bicycle taxable without losing identity, though it mightiness compete pinch the crest for ocular space.

I recognize the beak astatine (484,84) would overlap pinch the dome helmet, truthful I request to shrink the helmet truthful it only covers the apical of the head, adjusting its arc endpoints to beryllium higher and narrower truthful the beak tin connect cleanly astatine the beforehand without collision. [...]

I’m adding a darker extremity region to correspond the superior feathers, past reconsidering the trailing separator to see scalloped feather curves alternatively of 1 soft statement for a much earthy look. [...]

Now I’m checking the vent statement placements connected the helmet, making judge they beryllium acold capable wrong the helmet’s separator fixed the changeable width and rounded caps, and confirming each vent stays wrong the helmet’s information boundary. [...]

I determine skipping a handlebar doorbell and tyre highlights since they’re unnecessary additions. Now I’m reconsidering the beforehand fork’s curve — the existent power constituent pulls the style backward erstwhile it should front guardant for a due rake, truthful I request to displacement the power constituent rightward to hole the fork’s lean.

OK, let’s animate it

On Hacker News, swalsh commented connected that Max pelican:

Now that it’s a solved benchmark, tin we get the animated version?

I didn’t want to walk different $3 truthful I took the Max pelican and piped it into the default reasoning level of High:

6,121 input, 26,201 output = $1.37. The consequence looked for illustration this, exported present arsenic video since immoderate group person problem viewing animated SVGs:

Your browser does not support HTML5 video.

The wheels successful the video are rotating successful the incorrect direction, but I deliberation that’s an artifact of the conversion to MP4—they look to beryllium going successful the correct guidance successful the original SVG.