Leaks

What might be coming next.

Unreleased models and features people have found in code, docs, configs and A/B tests. Every leak has a status, the dates and the source, so you can judge it yourself.

Updated 14 Oct 2026

Rumor

Claimed by someone, nothing to back it up yet.

Spotted

Real evidence: code strings, docs, configs or A/B tests. Name, date or details can still change.

Confirmed

The company announced it or shipped it.

Debunked

Shown to be false, or the company said no.

SpottedAnthropicFirst seen 11 Oct 2026 · updated 12 Oct 2026

A stronger model behind Claude Opus 5.5 requests

Several X users reported that Opus 5.5 requests were being answered by a stronger, unannounced model, with standout code and creative outputs using MCP tools. People are guessing it's Opus 5.6 or Fable 5.5, but nobody knows the name.

Evidence so far

  • Reports from multiple accounts (@marmaduke091, @kimmonismus, @chetaslua)
  • An Anthropic employee posted "The best is yet to come"

Reasons for doubt

Only the outputs have been seen. The model's name and release date are guesses.

TestingCatalog
SpottedGoogleFirst seen 12 Oct 2026 · updated 12 Oct 2026

Nano Banana 3 Pro (image model)

"GemPix 3 Pro", Google's internal name pattern for Nano Banana image models, appeared in Google's image citation feature. It's expected to be built on Gemini 4 Argon, possibly in November or December.

Evidence so far

  • "GemPix 3 Pro" string found in Google's image citation functionality

Reasons for doubt

The timing is a guess, and Google has announced nothing.

TestingCatalog
SpottedGoogleFirst seen 9 Oct 2026 · updated 11 Oct 2026

Gemini 4 "Carbon" checkpoint

Alongside Gemini 4 Argon, Google employees are reportedly testing a separate "Carbon" checkpoint. Argon references in Antigravity also showed 256K, 512K and 900K context options.

Evidence so far

  • Context-window options found in Antigravity
  • Reports of a separate Carbon checkpoint in testing

Reasons for doubt

What Carbon is (a bigger model, or just a checkpoint) is unknown.

TestingCatalog
SpottedAnthropicFirst seen 10 Oct 2026 · updated 10 Oct 2026

Anthropic's own voice models

A hidden "Voice" item marked Preview showed up in Claude's web navigation, and Claude recently asked users to share voice data to improve its models.

Evidence so far

  • Hidden Voice (Preview) entry in the Claude web app
  • Opt-in prompt to share voice data

Reasons for doubt

There is no announcement, and a voice UI doesn't prove Anthropic trained its own speech model.

TestingCatalog
SpottedAnthropicFirst seen 10 Oct 2026 · updated 10 Oct 2026

Claude Health for web and a mystery "iggy"

Unreleased entries for Claude Health on the web and something code-named "iggy" were found in Claude's navigation.

Evidence so far

  • Unreleased navigation entries in the Claude web app

Reasons for doubt

What "iggy" is remains unknown.

TestingCatalog
SpottedOpenAIFirst seen 10 Oct 2026 · updated 10 Oct 2026

gpt-rosalind-discovery (life sciences)

OpenAI's API pricing docs list a gpt-rosalind-discovery model that hasn't been announced. It appears to be aimed at life-sciences research.

Evidence so far

  • Listed in OpenAI's API documentation

Reasons for doubt

There is no launch post and no details about the model.

TestingCatalog
SpottedGoogleFirst seen 9 Oct 2026 · updated 9 Oct 2026

"Ultra" mode for AI Studio Build

An unreleased premium Ultra mode was found in Google AI Studio's Build feature.

Evidence so far

  • Hidden Ultra option in AI Studio Build

Reasons for doubt

There is no pricing or launch date yet.

TestingCatalog
RumorAnthropicFirst seen 13 Aug 2026 · updated 13 Oct 2026

Claude Fable 5.5

Posts on X since mid-August claim a new Fable model is coming. In early October, screenshots claimed some users were quietly routed to it while the app still said Fable 5.1. A reference was also reported in Azure configs.

Evidence so far

  • Aug 13: a social post claims a new pretraining run
  • Sep 5-8: posts call it Fable 5.2, then Fable 5.5
  • Oct 1: screenshots claim silent routing to a newer model
  • Oct 13: a reported 'Claude Fable 5.5' string in Azure configs (unverified)

Reasons for doubt

Anthropic's models page still lists Fable 5.1 as the latest Fable. There is no model ID, price or system card for a Fable 5.5.

AIToolsReview: what is verified so far Crypto Briefing: Azure config sighting Original X post (@Mr_Salio)
RumorAnthropicFirst seen 11 Oct 2026 · updated 11 Oct 2026

"Claude Next" inside Claude Code

One X post claims a model called Claude Next is being tested in Claude Code and makes Opus 5.5 "look primitive", including claims about image and video generation.

Evidence so far

  • A single post from @chetaslua, with no numbers, tasks or outputs

Reasons for doubt

The name doesn't follow Anthropic's naming (Fable, Opus, Sonnet, Haiku). There is no model ID, doc entry or statement.

OrcaRouter: the Claude Next claim
ConfirmedGoogleFirst seen 30 Sep 2026 · updated 13 Oct 2026

Gemini 4 Argon

Announced Sep 30 for a limited cyber-defense rollout, spotted in A/B tests in Antigravity, then launched broadly for developers on AI Studio and Vertex AI with a 1M-token context window.

Evidence so far

  • Limited announcement (Sep 30)
  • A/B tests in Antigravity (Oct 11)
  • Broad launch on AI Studio and Vertex AI (Oct 13)
TestingCatalog: A/B tests Google blog

Everything on this page is anticipated, not announced. These are early signals and community reports. Names, dates and features can change, and some leaks never ship at all. Until a company confirms it, treat it as a maybe.