Claude Next Leak: Anthropic's Unreleased Model Reportedly Running in Claude Code

On October 11, 2026, an account posting as @chetaslua on X claimed that a model called Claude Next is being tested inside Claude Code, and that it makes Claude Opus 5.5 "look primitive." The same post said it would take "a month to make image and video gen model obsolete," claimed the model is "SOTA," and promised more tests later. That is the entire public record as of this writing. There is no model ID, no rate card, no documentation row, no model card, no arena listing, and no statement from Anthropic. Anthropic's shipped lineup on the day this was posted was still exactly four models — Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 — and Claude Next is not one of them.

So this is not a launch story. It is a what-we-know-so-far piece about a claim that is unusually easy to separate into parts, because one part of it is a thing Anthropic demonstrably does, another part collides with what Anthropic's own documentation says its models can do, and the name itself does not fit the vendor's naming grammar. All three observations are checkable today, on your own machine, without waiting for a launch.

What the post actually claims — and what it doesn't

Stripping the enthusiasm, the post asserts four distinct things, and they have very different evidentiary weight:

• A model named Claude Next exists. Unverified. Nothing in Anthropic's public surface carries that string.

• It is being tested inside Claude Code. Unverified, but structurally plausible — see the next section.

• It beats Claude Opus 5.5 badly enough to make it look primitive. Unverified and currently unfalsifiable, because no number, task, harness or output is attached.

• It threatens image and video generation models. Unverified, and in tension with every Claude model Anthropic has shipped to date.

Note what is not claimed. The post does not say the model is released, does not give a date, does not name a version number, and does not offer a model identifier. The author explicitly says more tests are coming. Read literally, it is a report of early access, not a report of a launch — which is exactly the shape of claim that a leak article should handle and a launch article should not.

The source is a single X account with no corroborating post from a second tester, no screenshot of a model picker, and no artifact. Anthropic has not commented. Treat every performance word above as reported, never as measured.

The part of the claim that isn't absurd

Anthropic does run unreleased frontier models with selected users. That is documented, not rumoured, and it makes "a new model is being tested somewhere" a much weaker claim than it sounds.

The clearest precedent is Claude Mythos Preview, described on Anthropic's own Project Glasswing page as "a general-purpose, unreleased frontier model," tested with a limited set of partners. The same page states plainly: "We do not plan to make Claude Mythos Preview generally available." A model can be real, capable, actively used by a handful of people, and never ship under that name — all at once.

There is a second, more recent precedent. Anthropic's expansion of its Cyber Verification Program, announced October 6, 2026, describes access tiers that cover "Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1, and new models moving forward." That last clause is a vendor commitment that future models — including ones that do not exist publicly yet — will be folded into a limited-access programme rather than launched to everyone at once.

And the specific venue in the claim, Claude Code, is the most likely place for an internal alias to leak. Claude Code's model selection is not a fixed list baked into the binary. It reads aliases, environment variables, and a gateway's own model list — which is a lot of surface for a name to appear on before any announcement.

Why the name itself is the weakest part

Anthropic's model names come from a small, closed vocabulary of evocative words: Mythos, Fable, Opus, Sonnet, Haiku. Version numbers follow a documented scheme of claude-{name}-{major}[-{minor}] — claude-opus-5-5, claude-sonnet-5-5, claude-fable-5-1. "Next" fits neither half of that pattern. It is not a word from the naming set, and it carries no version segment, so there is nothing for it to increment from.

There is also a mundane explanation worth keeping in mind before anyone gets excited. Claude Code ships on npm under two dist-tags, stable and next; next is the prerelease channel. Anyone looking at Claude Code's own tooling metadata will encounter the word "next" attached to something Anthropic is testing before it is promoted to stable — and the object being tested there is the client, not the model. That does not make the claim false. It does mean "next" is a word already sitting in this ecosystem, which is exactly the kind of word that gets picked up and repeated as a product name.

An internal development codename is a different matter — Claude Marshmallow and Claude Melon surfaced as EAP identifiers in August 2026 and were plainly internal names, not shipping names. If "Claude Next" is anything, it is more likely that kind of placeholder than a product. But the honest position is that nobody outside Anthropic can currently tell a placeholder from a misreading from a fabrication, because the only evidence is one enthusiastic sentence.

Where "makes image and video gen obsolete" runs into a wall

This is the part of the post most likely to mislead a reader, and it deserves a direct answer.

Every Claude model Anthropic currently ships — Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 — takes text and images as input and produces text as output. There is no image generation in the lineup. There is no video generation in the lineup. Claude Opus 5.5 supports vision, tool use, JSON output, reasoning and a 1M-token context window with up to 128K tokens of output; it renders a chart as pixels only in the sense that it can write the code that draws one.

That makes the claim ambiguous in a way worth naming:

• If the post means a Claude model now generates images and video natively, that would be a new capability class for Anthropic and a much larger claim than "a better coding model," and it would almost certainly arrive as its own announcement with its own pricing, not as a whisper in a Claude Code thread.

• If it means the model writes code that produces images and video well enough to reduce the need for dedicated generators, that is a claim about orchestration quality — a real and measurable thing, and one that several shipped models already compete on. It is not the same claim, and it is not what "obsolete" suggests.

The post does not distinguish the two. Until it does, the honest reading is that a reported qualitative impression of a model nobody has benchmarked is being used to make a category-level prediction.

How to check it yourself this afternoon

Here is the useful part. Everything the claim depends on is testable on your own machine, and none of it requires believing anyone.

1. Look at the model picker. In Claude Code, run /model with no argument. That opens the picker, and it reflects what your account can actually reach. A model that is genuinely being offered to testers on a channel you have access to would show up here. Run it twice on different days — pickers update out of band.

2. Try to address it by name. Claude Code accepts a model string from four places, in this order of precedence: /model during a session, the --model flag at startup, the ANTHROPIC_MODEL environment variable, and the model field in your settings file. Setting ANTHROPIC_MODEL to a string your provider does not serve produces an error on the first request — and the important detail is that Claude Code does not validate the string up front when it comes from the flag, the environment variable or settings. It sends it and lets the request fail. That is precisely how a leaked name would behave if the name were real but the access were not yours: it would look like a typo rather than a permission denial.

3. Add a custom entry to the picker. ANTHROPIC_CUSTOM_MODEL_OPTION inserts one extra model into the picker without disturbing the built-in aliases, with optional ANTHROPIC_CUSTOM_MODEL_OPTION_NAME and ANTHROPIC_CUSTOM_MODEL_OPTION_DESCRIPTION for the label. Anthropic's documentation is explicit that Claude Code skips validation for that ID. If you are on an account that has been given a name you were not given a list for, this is how you would point the client at it.

4. Ask an endpoint what exists. This is the decisive check, because model lists are server-side facts. Anthropic's Models API returns every model available to your key along with max_input_tokens, max_tokens, a capabilities object, and a line field naming the family the model belongs to. If a model is callable by you, it is in that response. If it is not in that response and not in the documentation's compare table, then no application code can reach it — whatever a post says about its quality.

5. Watch the two artifacts that always arrive with a real Anthropic model. Every shipped Claude model has a row in the models overview table and an entry in the Claude Code changelog on the release that adds it — Claude Opus 5.5 was added in version 2.1.280 and Claude Haiku 5.5 in 2.1.293, both with the model ID and the price stated in the changelog line itself. Those are the two places to watch. Neither currently contains anything named "Claude Next."

What would turn this from a rumour into a fact

Five artifacts would settle it, and they arrive in a fairly predictable order:

• A model ID. Something in the claude-{name}-{major} shape, or a variant that resolves on the Anthropic API. Without one there is nothing to call.

• A row in the models overview table. Anthropic's own comparison table currently has four columns. A fifth is the vendor confirming existence.

• A rate card. Price per million tokens, cache-read rate, any long-context tiering. Anthropic prices are published alongside the model, not after it.

• A model card or system card. This is where capability claims get bounded — including whether image or video output exists at all.

• A changelog line in Claude Code. The client states the ID and the price in the entry that adds a model, which makes it the single most reliable early confirmation surface for anything reachable from a coding session.

An LMArena or Artificial Analysis entry would be a useful independent signal on top of those — and, importantly, it would be a comparison against a fixed harness rather than a tester's impression. Note that at the time of writing, neither arena carries an Anthropic listing under this name — the 23 Anthropic entries on the text leaderboard top out at claude-opus-5.5-high at rank 2, and none of them is flagged preliminary.

What to do about it in the meantime

Nothing, if the goal is to ship something this week. The right posture toward an unbenchmarked model is the posture you would take toward any other unproven dependency: do not put it in the critical path, and do not build a roadmap on a sentence.

If and when a model like this does appear, the practical questions change shape fast, and they are the same three questions every new Claude release raises: what does it cost per million tokens, what does it do worse than the model it replaces, and what happens to your request when the provider is degraded. Claude Opus 5.5 is the current baseline at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million; Claude Fable 5.1 sits above it at $10/$50 for demanding reasoning and long-horizon agentic work, with cache reads at 2.5% of the input price. A successor to either would be evaluated against those numbers, not against a vibe.

The other thing that matters early, before any benchmark exists, is not betting a production path on a model nobody has measured. Running a new model behind a router lets you send it a share of real traffic and let automatic failover drop back to Claude Opus 5.5 or Claude Fable 5.1 when it errors, stalls or gets rate-limited — so the experiment costs you a fraction of your requests instead of all of them. On OrcaRouter that is one OpenAI-compatible endpoint in front of 200+ models with no second contract and no code change on the failover path, and provider list prices pass through unchanged, so a vendor price cut is live on our side the same day it lands. We do not serve Claude Next — no one does, because as far as any public surface shows, it is not a shipped model. Claude Opus 5.5 and the rest of the current Claude line are callable today and priced at list.

img src="4.png" alt="Screenshot of the OrcaRouter model page for anthropic/claude-opus-5.5, headed 'Claude Opus 5.5' under a Home > Models > Anthropic breadcrumb, with the identifier anthropic/claude-opus-5.5, 'by Anthropic — 2026-09-22', text, image and file input with vision, tools, JSON and reasoning, 128K max output, a 1M-token context window, $4.00 input and $20.00 output per million tokens, a p50 time to first token of 3.88 seconds, 126.3M tokens of traffic, and side-by-side code samples for the two SDKs against the same base URL." />

The open question

The most likely outcomes, in order: the post describes an internal or EAP-only build that never ships under that name, in which case no reader will ever be able to verify it and the whole thing evaporates; or an Anthropic model arrives under a proper name and ID with the usual artifacts attached, at which point this claim becomes a footnote about a leak that got the direction right and the name wrong.

Either way, the test is the same, and it takes about ten minutes on your own keys: open the picker, query the Models API, and look at whether a fifth column has appeared in the vendor's own table. Until one of those three changes, "Claude Next" is a name in a post — and the four models Anthropic actually ships are the ones worth your engineering time.