Anthropic Reportedly Preparing a New Model as Google and OpenAI Develop Their Own

Preface

Anthropic is reportedly preparing a new model with stronger reasoning and intuition than its currently available models. According to a report by The Information, the system may be particularly capable of inferring what users want when they provide little context, and of proposing less conventional approaches. However, Anthropic has not made an official announcement. The model’s name, product family, specifications, price, and release date are all unconfirmed.

The report also places the rumor in a competitive context: Google is said to be testing a version called Gemini 4 Carbon, while OpenAI may be preparing a new Terra model. This article separates reported claims from confirmed company information and summarizes the clues, product-line details, and uncertainties available at the time of publication.

Lazy bag

Reports describe an unreleased Anthropic model that may infer user intent from sparse information and generate more creative solutions than routine advice. Possible clues include “Fable 5.5” references and reports of unusual request routing, but none has been confirmed by Anthropic. Its existing product information includes Claude Fable 5.1, Claude Opus 5.5, and Haiku 5.5, each with distinct reported pricing or announcement details. Google employees are reportedly testing Gemini 4 Carbon, and OpenAI may introduce a new Terra model. The central takeaway is that these developments remain a mix of confirmed releases and unverified reports.

Main Body

The reported Anthropic model

Technology publication The Information published an exclusive report by journalist Stephanie Palazzolo at around 10 p.m. Taiwan time on October 19. The report said Anthropic was preparing a new model with reasoning and intuition capabilities that exceed those of the company’s models already on the market. At the time of publication, Anthropic had not issued an official announcement confirming the model or its characteristics.

That distinction matters. The reported capabilities, potential model identity, and release plans should be treated as claims attributed to sources or observers, not as established product specifications. Anthropic has not confirmed the model’s name, which product family it belongs to, its pricing, its technical details, or when it might become available.

A focus on intent and less conventional answers

According to a person familiar with the matter cited by The Information, the unreleased model may be especially good at identifying a user’s underlying intentions and goals when the user provides limited information. The report also describes it as more capable of thinking beyond conventional patterns, rather than simply repeating familiar recommendations or standard solutions.

The article’s example concerns a user asking how to invest in stocks. Instead of offering only generic advice such as looking at price-to-earnings ratios, the model might devise a creative algorithm to forecast stock performance. This example illustrates the reported distinction: the model would be expected not merely to retrieve common guidance, but to infer the task behind a brief request and develop a more tailored response.

That description does not establish that the model can reliably predict share prices or provide dependable investment advice. It is an illustrative example in the report, not a verified performance result. In practical terms, any financial output from an AI model would still need careful assessment, and a model’s apparent confidence should not be confused with guaranteed accuracy.

TestingCatalog commented that a system capable of inferring latent intent could be unsettling in light of existing safeguards. That concern points to a wider issue in AI design: interpreting what a user means can improve usefulness, especially when instructions are incomplete, but it also raises questions about how a model makes inferences, when it should ask clarifying questions, and how it avoids acting on assumptions that the user did not intend.

“Fable 5.5” references and reported routing behavior

TestingCatalog added that, earlier in the month, multiple users had reported that some Opus 5.5 requests appeared to be routed to another, more capable, unpublished model. The outlet also said that references to “Fable 5.5” had appeared recently. These observations have prompted speculation that the rumored model could be related to one of Anthropic’s product families, but they do not establish its identity.

A TestingCatalog roundup dated October 17 described several reported clues. An X user posted a screenshot showing Fable 5.5 in an Azure model catalog, but the related LiteLLM entry was removed a day later. Separately, a tipster claimed that partners had received access to test the model. These details are indirect evidence and do not amount to an announcement or confirmation from Anthropic.

TestingCatalog also acknowledged that leaked model code names can be difficult to connect to specific products. A name appearing in a catalog, a routing anomaly, or a third-party claim about testing access may offer a lead, but each can have more than one explanation. Anthropic has not confirmed any of these clues, so it remains unclear whether “Fable 5.5” refers to the model described by The Information, a separate project, or a label used in testing.

What Anthropic has officially announced

Anthropic’s public product information provides a useful baseline for understanding the reports. The company’s highest-tier Fable family was reported to have Claude Fable 5.1 as its latest version, announced in early September. Its API pricing was listed at $10 per million input tokens and $50 per million output tokens.

Claude Opus 5.5 was announced in an official notice dated September 22. Anthropic said that it performed comparably to Fable 5.1 on most tasks. The listed API prices were $4 per million input tokens and $20 per million output tokens. These details describe announced products and should be distinguished from the unconfirmed claims about the possible new model.

At the time the report was published, the latest model announcement on Anthropic’s official news page was Haiku 5.5, announced on October 7. The existence of established model families and versions helps frame the rumors, but it does not indicate whether a new system will be named Fable, Opus, or something else. Nor does it establish how any unreleased model would fit into the company’s pricing or product structure.

Google and OpenAI are also part of the reported picture

Palazzolo said The Information had also obtained information about Google’s “Gemini 4 Carbon” and the possibility that OpenAI could release a new Terra model. The wording around OpenAI was tentative, and the report did not provide confirmed launch details for either company’s rumored system.

Earlier in the month, Business Insider reported that Google employees were testing a Gemini 4 version called Carbon internally. One employee reportedly described its coding ability as close to Opus 5.5. Google had not publicly released specifications for Carbon, so the characterization remains an attributed assessment rather than an official benchmark or product claim. Google had already made Gemini 4 Argon generally available on October 13.

OpenAI’s existing GPT-5.6 Terra is positioned as a model balancing capability and cost. Details about a possible new Terra model were not clear in the report. As with the Anthropic and Google rumors, the absence of public specifications makes it difficult to compare capabilities, intended uses, prices, or availability across the systems.

How to interpret the reports

Taken together, the available information falls into three categories. First are official product details, such as the cited announcements and prices for Claude Fable 5.1 and Claude Opus 5.5, and Google’s October 13 general availability of Gemini 4 Argon. Second are reported observations, including the alleged Fable 5.5 catalog entry, possible routing behavior, and internal testing claims. Third are open questions: the rumored Anthropic model’s identity and release, Carbon’s specifications, and whether OpenAI will introduce another Terra model.

This distinction is useful because early model rumors often combine employee comments, interface or catalog traces, and anonymous sourcing. Such clues can indicate that development or testing is taking place, but they cannot by themselves establish final product names, capabilities, access conditions, or commercial terms. Until the companies provide direct information, comparisons should remain provisional.

The most notable idea in the Anthropic report is not simply that a new model may be more powerful. It is the emphasis on interpreting unstated goals and responding with ideas beyond standard advice. If confirmed, that capability could make interactions more useful when requests are vague. At the same time, it would make transparent reasoning, sensible clarification, and effective safeguards important parts of the product experience. For now, however, the model’s existence and qualities remain unconfirmed by Anthropic.

Key Insights Table