Claude Sonnet 5.5 on Extra Effort: Why the "Next Claude Model" Claim Fails Both Its Own Tests

At 12:44 UTC on 5 November, an X account called @chetaslua posted a sentence that has been circulating in various forms since October: "Sonnet 5.5 Extra looks like routing to next claude model." By the following morning the post had 20,900 views, 384 likes and 38 replies. The model it names is one you can call today — Claude Sonnet 5.5, shipped on 28 September at $2 and $10 per million tokens. The word "Extra" is not a second model. It is the second-highest rung on Claude's effort menu, which Anthropic labels Extra high and the API calls xhigh.

Two tests are offered as evidence: a question about where Andrej Karpathy went in 2026, and a pelican on a bicycle drawn entirely in SVG. Both tests pass on the model Anthropic shipped in September. That is the whole difficulty with the claim. One test checks a fact that falls a month inside Claude Sonnet 5.5's own training window. The other measures a quantity the Extra setting exists to increase. Neither can separate "a better model answered" from "the setting I chose did what its documentation says it does."

That does not make the post worthless. It makes it a clean specimen of a test that feels decisive and is not — and the two moves it uses, a knowledge probe and an output-size comparison, are the same two moves behind most routing claims this autumn. Walking through them is more useful than another round of guessing at what Anthropic has not announced.

What was actually posted

The post describes a single chat session and two prompts. The first asked where Karpathy went in 2026 and instructed the model not to use the web; the reply said he joined Anthropic in May 2026 on the pretraining team. The second asked for "pelican riding bike as realistic as you can only with codes," and the result is described as 26,599 SVG elements, roughly 6,500 feather strokes, no embedded images, data URIs or scripts, and 1,860 lines of Python. From those two exchanges the poster concludes that the label is serving something newer than the model it names.

Nothing accompanies it. No model identifier, no rate card, no system card, no Anthropic statement. The successor it gestures at has been circulating under several names since October, none of them Anthropic's and none of them callable. Anthropic's own model list, read today, carries four current models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. There is no fifth row, and no identifier anywhere in the post that a reader could test.

What "Extra" is, before it becomes a mystery

The effort menu in Claude's apps lists five rungs — Low, Medium, High, Extra high (xhigh) and Max — with the model's recommended rung marked Default. "Extra" is how people shorten the fourth. Anthropic's own reference describes xhigh as extended capability for long-horizon work, aimed at "long-running agentic and coding tasks (over 30 minutes) with token budgets in the millions." That is the vendor telling you, in advance, that this rung is meant to consume a great deal more than the one below it.

The rung is also calibrated per model, not globally. Anthropic's migration note for this generation says plainly that Claude Sonnet 5.5's effort levels are recalibrated — a level "doesn't produce the same amount of thinking as it did on Claude Sonnet 5" — and tells developers to re-run their effort sweep rather than carry a setting over. Two consequences follow, both boring and both fatal to the inference. A request at Extra on Sonnet 5.5 is not the same request as Extra on its predecessor, so "it feels different from what I remember" is expected. And turning a dial from High to Extra is an instruction to spend more tokens, so a longer, denser, more elaborate drawing is the requested outcome.

There is one more surface detail worth pinning down, because it is where the labels stop matching. The Claude API's default effort for this model is high; Claude Code's default for the same model is medium. A response from the app at Extra, a response from the API at its default, and a response from Claude Code at its default are three different configurations of one model. Comparing any two of them and concluding that a new model has appeared is comparing defaults, not checkpoints.

Test one: a May fact checked against a June cutoff

The Karpathy question is the stronger of the two probes, because unlike most of these tests its answer can be checked against the public record. It checks out. Karpathy announced on X on 19 May 2026 that he had joined Anthropic; he started that week on the pre-training team, reporting to Nick Joseph, and said he would build a team focused on using Claude to accelerate pre-training research. The reply the post describes — joined Anthropic, May 2026, pretraining team — is accurate.

The accuracy is the problem. Claude Sonnet 5.5's reliable knowledge cutoff is June 2026, and its training-data cutoff is the same. The hire happened roughly a month before that line. A shipped Claude Sonnet 5.5 asked this question with the web switched off is supposed to answer it, and answering it is what its documentation predicts. The probe produced the null result and was read as the finding.

It is worth being precise about what a knowledge probe can and cannot do, because this is where most of these claims go wrong:

• A correct answer about a pre-cutoff event is uninformative. Every model in the line has been trained on it.

• An incorrect answer would have been informative — in the opposite direction. If a model that should know a May 2026 hire denies it, that is evidence of a different, weaker checkpoint, not a stronger unannounced one.

• Only a post-cutoff fact discriminates. A fact published after June 2026, that no model in the current line could have been trained on, is the one shape of question that separates a new checkpoint from an old one. Karpathy published a workflow write-up on his own blog on 2 November. Asking for its contents with the web off would have been a real test. Asking where he went in May was not.

Test two: a pelican count that measures verbosity

The pelican prompt is Simon Willison's long-running ad-hoc eval, and he ran it against Claude Sonnet 5.5 on 28 September, the day the model shipped. His published results are the useful baseline here, because they are dated, attributed and measured rather than counted by eye.

• At max effort — one rung above the one the post used — the model spent 128,000 output tokens, hit the cap, cost $1.28, and never produced an SVG at all.

• At xhigh — the rung the post calls Extra — the same model produced a pelican in 41 seconds for 5.74 cents.

Read those two lines together and the post's second test collapses. The shipped model at Extra already draws the bird — that is the 41-second, 5.74-cent run — so a bird at Extra is not anomalous. The shipped model one rung above Extra draws nothing at all, so on this task the effort dial is the dominant variable by a wide margin, and the effort dial is the one thing the post changed. A result that moves with the setting you turned is evidence about the setting.

The element count deserves its own arithmetic, because it is the number most likely to be repeated. Anthropic states that 1M tokens is roughly 2.5M characters on the current tokenizer. If 26,599 elements averaged 40 characters each — a conservative figure once you include path data and attributes — the drawing would be about 1.06M characters, or roughly 424,000 tokens. That is more than three times the 128,000-token synchronous output ceiling, and above the 300,000-token batch ceiling as well. If the elements averaged seven characters, the same count fits comfortably in about 74,000 tokens.

So the count is bounded by markup density, not by model strength, and a terser model posts a bigger number. The post's own description points at the likelier reading: 1,860 lines of Python appear in the same chat, and a script that loops over feathers can emit 26,599 elements without the model writing 26,599 elements. The post does not say whether that code drew the pelican or was a separate artefact, which is exactly the detail a reader would need. What can be said without the file is that element count is a measure of verbosity on a rung whose documented job is to produce more.

The one genuinely interesting line in the description is the negative one: no embedded images, no data URIs, no scripts. A pelican at that level of detail built only from vector primitives is a real result. It is a result about output quality, though, and output quality at a higher effort setting is the least surprising thing a model can deliver.

The one documented way a different Claude answers

None of this proves that Anthropic would never serve a request under one label with different weights. It does mean the claim has to explain the mechanisms Anthropic does document, and both of them point the other way.

The first is refusal fallback. A declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area. Server-side fallback, set with fallbacks: "default" and in beta on the Claude API, retries cyber and frontier_llm declines on Claude Sonnet 5 — the older, cheaper model. That is a documented case of a Sonnet 5.5 request being answered by something else, and it is a downgrade that announces itself in the response.

The second is thinking-block binding. Thinking blocks record which model produced them, and when a conversation moves to a model that cannot read them the API drops them before the model sees the request, reporting the drop in input_transformations when the beta header is set. That is the API telling you the serving model changed. Nothing in either mechanism upgrades a claude-sonnet-5-5 request to a stronger unannounced model under the same string.

Absence of a mechanism is not proof that nothing is happening. It is, however, the standard the claim has to meet: a routing story that only ever describes an improvement has to explain why the one path Anthropic publishes runs downhill and leaves a receipt.

The closest thing to an outside check is the index that independent evaluators run against the model Anthropic ships. Artificial Analysis lists Claude Sonnet 5.5 at 55 on its Intelligence Index v5.0.1, ninth of 69 models against a median of 37. That is a reading of the string as served today, and it is the surface on which a materially stronger checkpoint hiding under the same name would be expected to show up eventually.

What a test that could settle it looks like

If you want to know whether the label in front of you is being served by something else, the experiment is not hard. It is just less fun than posting a pelican.

• Pin the identifier, not the label. Call claude-sonnet-5-5 with an explicit effort value, so the request is the same request every time. A label in a menu is a product surface; an identifier is a target.

• Sample more than once. One response at one rung is an anecdote. Ten at high and ten at xhigh, same system prompt, same max_tokens, gives you two distributions. Effort produces a smooth, monotone shift in output length. A different checkpoint tends to produce a discontinuity in content that length does not explain.

• Use a post-cutoff fact. A June 2026 cutoff means everything published after June is fair game and everything before it is worthless as a probe. This is the single change that would have made the 5 November test worth running.

• Read the response object, then discount it. The model field, usage, stop_reason and stop_details are worth logging, and a refusal fallback is visible there. But those fields echo what the server says it did. They are a statement, not an inspection — the same category of evidence as asking a model to describe itself.

• Price it before you run it. At $2 and $10 per million tokens, a 128,000-token response costs $1.28, which is Willison's figure and also just the rate card. Twenty responses allowed to run to the cap is about $25.

What Extra costs while you wait for an answer

There is a practical question underneath the rumour that does not depend on resolving it. Claude Sonnet 5.5 bills thinking as output, thinking counts toward max_tokens, and Extra is documented for token budgets in the millions. A single response that spends a million output tokens costs $10 on the output side alone, at a rate card that many teams picked precisely because it was the cheap tier.

• One rung up from Extra, the measured pelican run hit the 128,000-token cap and cost $1.28 without producing a drawing. The cost of a failed request on this model is not zero.

• One rung down, the same task finished in 41 seconds for 5.74 cents. The gap between the two is more than twentyfold, and it is set by a dropdown.

• A team running 1,000 Extra requests a day at 30,000 output tokens each is spending 30M output tokens a day, or $300, before input is counted at all.

That is the part a reader can act on today regardless of what Anthropic ships next. If Extra is a different model, the windfall is real but unplannable. If Extra is what its name says, then a workload that has not been swept against high and medium is paying for thoroughness nobody measured.

OrcaRouter routes Claude Sonnet 5.5 at the vendor's own $2 and $10 with no markup on top, and effort is a per-request parameter rather than a separate contract, so an effort sweep across rungs is a loop in a script. That is also the honest hedge against a claim like this one: pin the model identifier you meant, run the same prompt at each rung, and compare your own token counts instead of a screenshot. When a routing question cannot be settled from outside, the next best thing is a setup where the answer is measurable on your own traffic — and where failing over to high or to a different model is a config change rather than a migration.

What would change this page

A model identifier, a rate card, a system card, or a sentence from Anthropic naming the successor. Any one of those settles the question in an afternoon, and none of them exists today. Until one does, the honest description of the thing behind the label is the one Anthropic publishes: Claude Sonnet 5.5, a June 2026 cutoff, five effort rungs, and a documented reason to believe that a bigger pelican at a higher setting is a bigger pelican at a higher setting.

The post's own framing gives the game away. It compares Extra against an expectation, not against a measurement — the same rung on Sonnet 5, or the same rung at max, or a published baseline from the launch week. Those comparisons exist and are cheap. Until someone runs them, "looks like routing" is a description of a dropdown doing its job.