Claude Sonnet 5.5 Extra: A Routing Rumor Built on a Word Anthropic Already Uses

At 12:44 pm on 5 November 2026, an account posting as @chetaslua on X put a claim in one line: "Sonnet 5.5 Extra looks like routing to next claude model." Both tests were run on the configuration the post labels "Sonnet 5.5 Extra," which resolves to Claude Sonnet 5.5 with its effort setting turned up. In the first, the model was asked — with browsing switched off — where Andrej Karpathy went in 2026, and answered that he joined Anthropic in May 2026 on the pretraining team. In the second it was asked for "pelican riding bike as realistic as you can only with codes," and returned, by the poster's own count, 26,599 SVG elements, 6.5k feather strokes, zero images, zero data URIs, zero scripts and 1,860 lines of Python, which the poster says "rebuilds byte for byte."

The most useful thing to say first is the unglamorous thing. Claude Sonnet 5.5 is not an unreleased model — it shipped on 28 September 2026 at $2 per million input tokens and $10 per million output, and Amazon Bedrock documents the same model under the ID anthropic.claude-sonnet-5-5. And "Extra" is not a new name either. It is the word in Anthropic's own help centre for the second-highest reasoning setting, "Extra high (xhigh)," which the API spells xhigh and the model selector renders as Extra high. A picker entry reading "Sonnet 5.5 — Extra high" is a model plus an effort level, and both halves of that string already exist in public documentation.

None of which makes the poster wrong. It makes the poster unproven, and it narrows what would prove it. Anthropic has confirmed nothing, no model ID has appeared in any public model list, and a leak frame is the right one: this is what the claim says, what can be checked today, and what a reader should do with it. The useful version of this story is not "a new Claude is being tested." It is that the three pieces of evidence offered here have three specific failure modes, and each one is testable.

What the Karpathy answer actually establishes

The first test is being read as a fingerprint. It is closer to a calendar.

Karpathy announced on X on 19 May 2026 that he had joined Anthropic — "I've joined Anthropic... excited to join the team here and get back to R&D" — and TechCrunch, Fortune and the NY Post all reported it the same day, with the detail that he landed on the pretraining team under Nick Joseph. That is a fact in every news index, and a model that answers it correctly with browsing disabled has demonstrated exactly one thing: its training data extends past mid-May 2026.

Claude Sonnet 5.5 shipped on 28 September 2026. A hire that was global news four months before a model ships is the kind of fact that ends up inside its training data, so the model passing this probe is the expected result rather than a surprising one. The probe also cannot distinguish Sonnet 5.5 from Claude Opus 5.5, from Claude Haiku 5.5, or from any other Claude checkpoint released after May. A test that every candidate passes identifies none of them.

This matters for how the claim should be scored. The reply thread under the post contains the more interesting half of the evidence, and it is contradictory: one reply asks "you mean sonnet 5.6?", another says "it's fable 5.5 too been routing for users for weeks now." Two people reading the same behaviour arrive at two different unannounced models, and neither can name an API ID. That is what an ambiguous label looks like when a community tries to interpret it.

26,599 SVG elements is a loop count

The pelican number is impressive until you notice what it is counting. An SVG element is a tag. A Python script that places feathers in a loop can emit forty thousand of them without the model ever reasoning about a pelican, and that is precisely what the poster describes: 1,860 lines of Python generating the elements, rather than 26,599 elements written by hand. The number measures the ambition of the generator, not the fidelity of the drawing.

"Zero images, zero data URIs, zero scripts" is the same kind of statistic. It confirms the output is a pure vector drawing rather than a bitmap smuggled into a file. It says nothing about whether the bird reads as a bird. If you want to judge the drawing, judge the drawing — the poster's own video attachment runs 19 seconds, and that is the artefact worth looking at.

There is also a sequencing detail worth knowing before this gets filed as a fresh capability event. On 22 October 2026, the same account posted a screenshot of a "mysterious model on code arena" with the prompt "code an svg of pelican riding bike as detailed as you can," which drew roughly 63,000 views. The pelican prompt is this account's standing routing detector, run repeatedly. Two results from the same detector three weeks apart are not independent observations.

The pelican test was retired by the people who made it famous

Simon Willison introduced "Generate an SVG of a pelican riding a bicycle" on 25 October 2024, on the reasoning that no such SVG could plausibly exist in training data, so every model had to actually attempt it. It became a semi-formal benchmark: by June 2025 Willison had collected 34 pelicans and 560 pairwise judgements.

Then Andrej Karpathy, the same person whose location is the subject of the first test, declared it dead. In early August 2026 he gave Claude Opus 5 the opening paragraph of The Lord of the Rings with roughly a million tokens of budget and about $10, and let it run autonomously for two hours; it produced roughly 5,500 lines of Three.js that procedurally built an interactive low-poly Shire, "kind of janky but fun." His argument was that the pelican test had lost discriminating power: frontier models now clear it, so it no longer separates them.

So the 5 November post uses a detector that its own originator retired four months earlier, on a model family that had already cleared it. That does not make the output bad. It makes it non-diagnostic — which is the one property a routing test needs.

"Extra" is a documented effort level, not a model name

Here is the part that is checkable in about ninety seconds on Anthropic's support site. The help centre page on changing model, effort and thinking settings lists five levels: Low, Medium, High, "Extra high (xhigh)", and Max. Its own gloss for Extra high is that it is "designed for long-running coding and agentic tasks, offering deeper reasoning than high without the full token cost of max," and it adds that the level is "available on Opus 4.7 and newer models." The API parameter is spelled xhigh; the selector spells it out.

Claude Sonnet 5.5 has all five levels — low, medium, high, xhigh and max — with thinking permanently on and no way to disable it. Its default is not the same on every surface — the Claude API and Claude Code do not necessarily land on the same level out of the box. Set nothing and you may already be running a different configuration depending on where you called from. So a user who opens a model picker, sees the default Sonnet 5.5 entry, and switches the effort to Extra high has created a label that reads exactly like "Sonnet 5.5 Extra" — while running a model that has been generally available for thirty-nine days.

What effort does to the bill is the part worth internalising, because it is where the "next model" impression probably comes from. Price per token does not move with effort: $2 in, $10 out, $0.20 cache reads, at every level. Tokens per task moves enormously. Artificial Analysis ran the full sweep at launch on its pre-v5 index and got this ladder:

• Low — index 36 at $0.41 per task

• Medium — index 41 at $0.59 per task

• High — index 47 at $1.08 per task

• Extra high (xhigh) — index 52 at $2.74 per task

• Max — index 56 at $7.60 per task

Read that as a spending curve, not a capability curve. Moving from high to Extra high more than doubles the cost per completed task for five index points — the same model, thinking longer. A user who flips that switch and then watches the output get visibly more elaborate would describe it as "a different model." It is the same weights, given more room.

The current live reading is lower and differently measured, and the two must not be mixed. Artificial Analysis rebuilt its index to v5.0 on 27 October 2026, replacing nine constituent evaluations and weighting 60% toward private or held-out sets. On v5.0.1 the same Claude Sonnet 5.5 reads 55 at max effort, 155 tokens per second, $10.02 per index task, having generated 335 million output tokens across the run against a 72.6 million median. The 56 above is a pre-v5 number. Subtracting one from the other is meaningless.

The verbosity figure is the one that survives both rubrics: roughly 4.6 times the median output volume on an identical task set. That is what "Extra" effort buys and what it costs, and it is a much more plausible explanation for a startling demo than a silent model swap.

What would actually settle it

The claim is falsifiable, which is more than can be said for most leak chatter. In rough order of strength:

• A model picker entry whose name does not appear in Anthropic's public model list — a string, not a screenshot of behaviour.

• A response body whose model field reports something other than a documented ID. Every Claude API response echoes the model that served it; a silent swap that survives that field is a different, larger story.

• A dated changelog entry, a Bedrock or Vertex model card, or a documented context or output limit that differs from Claude Sonnet 5.5's 1M in and 128K out.

• The generator script behind the pelican, run twice on two named configurations, with the outputs compared. The poster says it rebuilds byte for byte; that is a claim two runs could settle.

Until one of those lands, the honest reading is that an effort level produced a more impressive drawing and a community read it as a new release. Anthropic has published nothing about it, and the post itself does not claim a name it can back with an identifier — it says the behaviour "looks like" routing. That is a description of an impression.

The real problem is silent routing, and it is fixable

Strip away the model name and there is a legitimate worry underneath: if a provider can serve a request under one label while running something else, every benchmark you run and every cost estimate you make is measuring a moving target. The answer is not better folklore. It is pinning.

On OrcaRouter, Claude Sonnet 5.5 is one route among 200-plus models on a single key, at the provider's list price with 0% markup — so the $2 and $10 above are what you actually pay, and a vendor price change is live on our side the same day. You name the exact model in the request; failover moves you to another provider only when one errors, and it tells you, rather than quietly substituting. If you want to test the hypothesis in this post properly, you do not need a screenshot: route the same prompt to Sonnet 5.5 at high and at Extra high, label the two routes, and compare the outputs and the token counts yourself. That is a controlled experiment. It costs cents, and it produces a number instead of an impression.

If a genuinely unannounced model turns up, it will be the same discipline that catches it — a pinned ID that suddenly fails to match the model field, not a pelican that looks better than last month's.

What to do before you rebuild anything

Nothing, yet. The claim is one day old, single-sourced, and its strongest piece of evidence is a documented effort level. Do not migrate a pipeline, re-benchmark a model, or write a launch post on the strength of it.

What you can do this week is cheap and useful either way. Check which effort level your own calls are actually running, because defaults vary by surface and almost nobody sets it explicitly. Watch cost per completed task rather than price per token, since that is the number Extra high doubles. And if you are comparing models, pin the ID and log the model field on every response — the only routing claim that matters is the one your own logs can confirm.

The open question is not whether Claude Sonnet 5.5 Extra exists. It is whether anyone will produce an identifier for it before the next pelican.