Anthropic's Intent-Reading Model: What The Information Reported, and What Claude Fable 5.5 Has to Do With It

On 19 October 2026, The Information's Stephanie Palazzolo published a source line that describes an unreleased A​nthropic model entirely by temperament: A​nthropic, she reported, is preparing a model with stronger intuition and reasoning, one that “can discern underlying intentions” and is “better at thinking outside of the box.” No name. No identifier. No price, no context window, no benchmark. The same day, The Information's co-founder Amir Efrati put the obvious question on X — “Is A​nthropic nearing the release of Fable 5.5? Seems like it...” — with a question mark attached. That question mark is the most important character in this story. Claude Fable 5.5 is the leading candidate for the model being described; Claude Opus 5.5 is the model currently sitting at the top of the lineup and of Artificial Analysis's Intelligence Index. Neither fact makes the described model Fable 5.5, and nothing below should be read as A​nthropic having announced anything at all. The documented AI failures and model-launch incidents behind stories like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

This is a what-we-know-so-far piece, written one day into a leak. There is no launch to report, so treat every capability claim here as a reporter's characterisation of an internal model, relayed secondhand, and every product detail as something that does not exist yet on any public surface.

One sentence, four claims, and what each one is worth

The report is a single summary line, and it packs four separable assertions into it. Unpacking them is the only way to price the leak, because three of the four are ordinary and one is genuinely unusual.

• Better intuition — a qualitative claim with no unit attached. Every generation gets described this way; it tells you nothing you can grade.

• Better reasoning — likewise generic, and already the axis every lab competes on. Claude Opus 5.5 and Claude Fable 5.1 are both billed as reasoning models with adaptive thinking, so a successor scoring higher on a reasoning benchmark would be expected rather than newsworthy.

• Discerns underlying intentions — this is the outlier. It is a claim about inferring what a user meant rather than what a user typed, which is a different capability from instruction-following and is not something any current vendor measures on a public leaderboard.

• Thinks outside the box — the vaguest of the four, and the one most likely to have been a source's own phrasing rather than a lab's internal terminology.

Worth noting how unstable even that one sentence is in transit. Two captures of the aggregator line carrying the report, taken hours apart on 19 October, already differ by two words: the earlier snapshot reads “superior intuition and reasoning” and “discern unstated intentions,” the later one “better intuition and reasoning” and “underlying intentions.” That is a small thing, but it is the normal condition of leak text — a summary of a paywalled article, re-summarised by an aggregator, then quoted by accounts that never saw the original. Quote-level precision is not available here, and anyone presenting an exact quote as the report's wording is overstating what they hold.

Why “discern underlying intentions” is a strange thing to claim

Model launches are usually argued on numbers: an index score, a price per million tokens, a context window, a terminal-bench percentage. Intent inference does not have any of that. It is the kind of capability you notice in failure modes — the model that asks the question behind the question, or that stops and says the request as written would produce something the user plainly does not want.

It is also the kind of capability that cuts both ways, which is why the reaction to the line was more interesting than the line itself. Drew Breunig, replying on X, framed the tension cleanly: reading unstated intentions “will be good for the small details,” but if it leads the model to make significant, out-of-scope assumptions, the trade is unclear. That is the right question to ask of any model sold on intuition. A model that infers intent well is more useful on underspecified work and less predictable on work where the literal instruction was the point — a migration script, a billing calculation, a red-team prompt. There is no published evaluation that separates those two behaviours, and A​nthropic has published nothing at all about this model.

So the honest reading of the leak is narrower than the headline version: A​nthropic is reportedly optimising a frontier model along an axis that the industry currently has no scoreboard for. That is a real signal about direction. It is not a specification, and it cannot be converted into one until the vendor publishes something.

The name problem: three readings, and the test that separates them

The report does not name the model, so the name has to be inferred, and there are three plausible readings. They matter because they imply different prices and different upgrade paths.

Reading one: it is Claude Fable 5.5. This is the reading most of the leak-tracking ecosystem has settled on, and there is a body of prior, separately unverified reporting behind it: the outlet TestingCatalog reported partner access on 17 October, and early-access codenames — orzata, boba, vanilla, odango, verrine and onigiri — turning up in developer tooling through late September and early October, with onigiri the popular guess for the Fable line. If this reading holds, the new model slots into the $10 / $50 per million tier that Fable 5.1 occupies.

Reading two: it is an Opus-tier checkpoint. On 12 October, users reported Claude Opus 5.5 requests landing on a stronger, unannounced model — an observation about behaviour, not about naming. If the intent-reading model is that checkpoint, it would arrive in the $4 / $20 tier and replace the model most developers are currently defaulting to.

Reading three: it is something else entirely — a Mythos-class step, or a specialised line. The report gives no tier, and A​nthropic's documentation carries a Mythos family in its navigation that the current four-model comparison does not include.

The tests that would separate these are concrete, and none of them has fired. A new API identifier appearing in A​nthropic's models overview would name the tier immediately, because the identifiers carry it: claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5. A published rate card would separate a $10/$50 launch from a $4/$20 one. A row on a cloud marketplace with the pricing inherited from the predecessor would suggest a point release, not a new line. And an entry in Artificial Analysis with effort levels attached would tell you which family's reasoning profile it follows, since Fable and Opus are configured differently there.

The two datapoints everyone quotes have already gone stale

Leak coverage compounds, and by this week the Fable 5.5 case had accumulated two pieces of “evidence” that both evaporated within a day of appearing — which is worth saying plainly, because each is still being cited as if it were standing proof.

The first was an Azure East US 2 catalogue entry for Fable 5.5, surfaced on 17 October and reported by TestingCatalog, reportedly seen through a LiteLLM listing at $10 per million input and $50 per million output with a 1M-token context window and 128K output. Every one of those values matches Claude Fable 5.1 exactly, which is what you would expect from a row that copied its predecessor's metadata rather than from a finished product listing. The entry was removed a day later when queries against Azure came back empty.

The second was a codename, claude-onigiri-eap, spotted in a leaderboard frontend with four effort levels and a fallback to Claude Opus 4.8 — a configuration pattern consistent with the Fable and Opus tiers. It was pulled within hours.

Neither is evidence that a model does not exist. Both are evidence that a model does not yet exist in a form anyone can call, which is the only form that matters to a reader deciding what to build on. The pattern — metadata that mirrors the predecessor, artefacts that vanish on contact — is what staging looks like, and staging is not a release.

What it would land into, if it lands

A​nthropic's current lineup, as documented on 20 October, is four models that all share a 1M-token context window, a 128K maximum output and a June 2026 knowledge cutoff. The differences are price, speed and score:

• Claude Fable 5.1 — $10 per million input, $50 per million output; the demanding-reasoning tier; retirement floor not sooner than 1 September 2027.

• Claude Opus 5.5 — $4 / $20; the default recommendation for most workloads, and the model A​nthropic points developers to first.

• Claude Sonnet 5.5 — $2 / $10; the speed-and-intelligence balance, with five effort levels whose calibration changed from Sonnet 5.

• Claude Haiku 5.5 — from $0.10 / $0.50; high-volume classification, extraction and routing, shipped 7 October.

The scoreboard has an oddity in it that the leak makes more interesting. On Artificial Analysis's Intelligence Index, Claude Opus 5.5 sits at the top at 58, with Claude Fable 5.1 next at 53 — so the $4/$20 model currently outranks the $10/$50 one, which is exactly why A​nthropic's own docs tell people to start with Opus 5.5 and only move to Fable when their evals on Opus at higher effort still fall short. A new model whose selling point is intuition rather than index position would slot into that gap awkwardly: it would be hard to justify at Fable pricing unless the intent behaviour shows up in work people actually run, and easy to justify at Opus pricing if it does.

What you can call today, and the honest answer about this one

Nothing new. There is no model identifier for the model in the report, and if you ask our catalogue for anthropic/claude-fable-5.5 it returns model not found — the string is not in the 209-model catalogue at all, while anthropic/claude-fable-5.1 resolves with a 1M context window, a 128K output cap and $10 / $50 per million pricing. Any product offering access to “A​nthropic's new intent model” right now is selling a label.

What is live is the four-model lineup above, at A​nthropic's list price, through one API. If the practical question is how to be ready for a model that reads intent, the answer is not to speculate on the name — it is to make the swap cheap. A​nthropic price changes and new releases land on OrcaRouter the same day at 0% markup, so a rate card that moves from $10/$50 to something else does not require a new contract; and where a new model is unproven on your traffic, an automatic failover chain lets you send it a share of requests with a proven model behind it rather than betting a production path on a leak. You can see exactly what the current top of the lineup costs on the Claude Opus 5.5 model page, and compare it against Fable 5.1 before deciding which one your workload should default to.

What would change this piece

Three things, in ascending order of how much they would tell you.

A first-party mention. A​nthropic's newsroom has not published a model announcement since Claude Haiku 5.5 on 7 October; between 8 and 19 October the posts are evaluator policy, Singapore, the midterms and a public-benefit essay. The day a model post appears, the name, the price and the positioning all arrive together and the leak becomes history.

A rate card or an identifier. Either one resolves the tier question in a single line, and the identifier resolves the family question too, because the naming convention has been stable across Fable 5, Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.

An independent evaluation of intent inference — which does not exist yet for any model, and is the real reason this leak is hard to grade. Until someone publishes a task set that separates “inferred what I meant” from “invented what I did not say,” a claim about discerning underlying intentions is a description of a design goal, not a capability you can compare. When that evaluation appears, the model it is run on will be worth taking seriously. Until then, the interesting fact is not that A​nthropic is building toward intent. It is that a frontier lab is now marketing along an axis nobody has learned to measure.