Does superintelligence already exist? Evidence in 2026

Does superintelligence already exist? In the public sources examined at 10 October 2026 There is not enough evidence to attribute a product to a superhuman generality according to the framework adopted. This is not to say that every possible system has been examined: the conclusion is limited to the accessible evidence consulted.

There are important results in computer work, scientific research, reasoning and the use of tools. The point is to figure out what statement they support. A documented progress in a task and a test of general superintelligence have a different scope.

For ASI we use here the superhuman general reference of the paper Levels of AGI. The paper proposes a criterion; its product classifications are historical and are not used to describe 2026. The full definition is in the guide to what is artificial superintelligence. Morris and others, Levels of AGI.

What evidence have we examined and what conclusion do they allow?

Reconnaissance includes official OpenAI, Google DeepMind, Anthropic and Meta sources: product pages, technical announcements, architectural descriptions and a paper on the AlphaEvolve method. References describe different aspects and do not form a race with identical conditions.

The essential distinction is between documented fact, interpretation and limit of the test. The fact may be that a laboratory publishes a result on certain activities. The interpretation may be that the system has an interesting ability for another process. The limit lies in what that experiment does not prove.

We have not directly reproduced lab results, reviewed private evaluations or verified all existing models and releases. The pages cited allow you to check public announcements; they do not replace a complete independent assessment of superhuman generality.

The editorial conclusion is therefore precise: the evidence consulted is not enough to support the ASI attribution. We do not derive from this either an arrival date or a proof of impossibility. Even the absence, in the corpus, of a certification with a universal threshold must be described as the limit of reconnaissance.

This boundary makes the answer more useful than an absolute yes or no. It preserves the relevance of the observed results and prevents them from turning into a broader statement than the available evidence.

What do the 2026 ads document?

The cases in the table are selected examples of public evidence. Dates identify references, without presenting them as the complete inventory of the most recent releases.

Lines do not assign a winner. One source speaks mainly of model performance, another of infrastructure, another of a research method. Adding up these descriptions does not produce a single ASI test.

To understand how components can enter a system, the relevant insight is the technologies of OpenAI, Anthropic, Google and Meta. Here instead concerns the transition from single evidence to the thesis on the existence of superintelligence.

Are the results of GPT-6 Astra sufficient to demonstrate ASI?

OpenAI describes on the Astra page capabilities of computer use and professional work. The same page contains notes on the conditions of the September evaluations and updates: on September 29 it refers to GPT-6.1 Sol. Astra is therefore a documented reference from 2026, not a label to be used automatically for “the most recent model.” OpenAI, GPT-6 Astra.

A statement about a set of professional activities can be significant for research and for specific applications. The ASI classification requires, however, that the evidence supports the general superiority provided for in the criterion. The jump between the two claims must be argued, not presumed by the name of a benchmark.

Our reading does not reproduce a ranking of scores. For the detail of the published results, the blog already contains GPT-6 Astra official benchmarks. To interpret conditions and measures, we refer to how to read AI benchmarks.

A good result remains a good result even when it does not solve the ASI issue. Caution is about extending the conclusion, not rejecting the evidence.

What do Deep Think's scientific contributions prove?

Google DeepMind describes in February 2026 the use of Gemini Deep Think in mathematical and scientific research, including Aletheia and a cycle of generation, verification and review. The report involves experts and collaborative contexts: it should not be transformed into an indistinct description of entirely autonomous research. Google DeepMind, Deep Think in Search.

The distinction is substantial. A contribution to a discovery can be original and important. To describe it, one must identify the work of the system, the verification and the part done by humans. Removing these elements from the story changes what the case proves.

The Gemini 3.1 Deep Think page also features a table referring to February 2026 and distinguishes ratings with and without tools. The date and conditions belong to the test: they are not accessory notes to be deleted in a current comparison. Google DeepMind, Gemini 3.1 Deep Think.

From these cases we can observe evidence of documented scientific ability. We cannot infer, from them alone, that the same system exceeds persons in any other family of activity relevant to the ASI criterion. For the differences between amplitude and performance, the reference is AI, AGI and ASI: how to compare them.

Is Meta's "personal superintelligence" a certification?

In the April 2026 announcement, Meta presents Muse Spark as a step on the path to personal superintelligence. The source describes multimodal reasoning, instrument use, and orchestration of multiple agents. The correct attribution is "Meta presents" or "Meta describes", without transforming the positioning into an independent classification. Meta, Introducing Muse Spark.

The name of the laboratory and the stated objective indicate a direction of development. On the other hand, the ASI question requires an explicit relationship between criterion and evidence. Using the word in the communiqué does not resolve this relationship.

The functions of a product can still deserve a business evaluation. The question of what is usable and the question of the level of intelligence are separate. For ads, meaning attributed by the provider and availability, the dedicated piece is Meta and Muse Spark personal superintelligence.

Do AlphaEvolves and Persistent Agents Demonstrate Unlimited Capability?

The AlphaEvolve paper describes an algorithmic search method in which code proposals are evaluated according to prepared criteria. The evidence relates to measurable improvements within that problem; it does not alone demonstrate unlimited general self-improvement. Google DeepMind, paper AlphaEvolve.

The DeepMind update of May 2026 reports applications of the approach. These are cases communicated by the supplier: we do not automatically transfer them to the performance of an Italian SME. Analysis of the method is in AlphaEvolve: an AI that improves itself?.

Another advance concerns continuity of work. Anthropic describes Managed Agents through the separation of a persistent session, the agent loop and the execution environment. This is an engineering result for managing prolonged tasks and distinct components, not independent proof of superhuman generality. Anthropic, Scaling Managed Agents.

The two evidences show why the term “advanced” can relate to different aspects. Optimizing an algorithm and maintaining work after an interruption are real problems; solving them contributes to a useful system without exhausting the ASI classification.

How to distinguish a local test in a packaging manufacturer?

Consider one hypothetical scenario: a small packaging manufacturer receives an AI proposal to prepare production plans. The demo finds a job order that meets the constraints presented and looks better than the manual schedule used as a comparison.

The result may be useful, but you need to understand what it supports. If the test includes only homogeneous orders, it still does not show the behaviour with missing materials, an unavailable machine or incompatible requests. These situations belong to the promised ability in the process, even if they do not appear in the demo.

The firm may ask to keep the constraints visible and to explain which plan is discarded. A proposal that reduces a cost criterion but violates a material condition is not an acceptable improvement. The definition of "best" must include real needs, not just the value that the system optimizes.

If the supplier only proposes plans to be approved, the test must judge the proposals. If the system also needs to change the programming, the effect of the change should be observed. The new permit changes the verification required.

A success on these cases can justify a circumscribed function. It does not constitute a demonstration that the system has superhuman general capabilities. The company can accept the local result without formulating the ASI thesis.

What evidence does a travel agency need?

One second hypothetical scenario concerns a small agency that wants to prepare route drafts using traveller requests and supplier documents. The system produces an ordered text, brings the conditions together and proposes a plausible path.

The ability to observe is consistency with sources and demand. If a document contains a limitation, the itinerary must retain it. If the client's desire is incompatible with a condition, the system must signal the conflict instead of hiding it in a convincing proposal.

The agency could prove a complete request, one with contradictory preferences and one with missing information. Not to measure ASI, but to see if the draft supports the work of the operator. The verification considers what needs to be corrected before using it.

Preparation and booking must also be separated. A consistent draft does not authorize you to commit an availability or communicate a warranty to the customer. Actions involving the supplier require an explicit perimeter and corresponding evidence.

In this scenario, a less autonomous assistant can already have a useful role. The evidence to look for is that it improves preparation of the work without losing essential conditions. For the boundary between action and intelligence, it is useful AI Agents and Superintelligence in SMEs.

Three possible outcomes when examining a test

The test supports the proposed function: the result meets the criteria of the process and the necessary support is stated. You can evaluate that function in the observed perimeter.

The evidence is interesting but incomplete: Relevant cases or essential conditions are missing. Ask for the missing evidence before extending the promise.

The evidence relates to another statement: a specialist result is used to support general superiority. It preserves the result, but separates it from the wider conclusion.

What errors lead to declaring ASI too soon?

Using the test name as a certification. A signature that contains AGI does not automatically attribute that level to the system that gets a good result. You have to read what has been measured and what conclusion is argued.

Turning the result of the system into a quality of the model only. Tools, verifiers, and human work can contribute to the outcome. If they are eliminated from the story, the reader receives a different description from the documented experiment.

Reading a historical table as a complete photograph of the present. A February comparison keeps its date even if the page is accessible in October. The blog deepens because supplier benchmarks are not automatically comparable.

Counting only the successful attempt. Anthropic distinguishes measures of success achieved at least once and constant success. For a repeated function in the company, the case that requires correction remains part of the test. Anthropic, Demystifying evals for AI agents.

Exchanging a circumscribed conclusion for universal certainty. Failing to find sufficient ASI evidence in the corpus does not prove that any possible assessment fails. On the contrary, a successful case does not automatically close the generality problem. Both extensions exceed the evidence described.

How to use this evidence in the decision of an SME?

The question of the existence of ASI is broad; the acceptance of a business application can be limited. The useful choice is to link every promise to a test of the same perimeter.

If the proposal promises to read documents, it brings relevant documents and observes the result. If he promises to act, he also observes the final state of the process. If it promises to handle exceptions, ask for cases where data is missing or conditions come into conflict.

Human support should be described as part of the service. An operator who selects sources, corrects a draft, or decides on a suspension performs work necessary for success. Making that contribution visible helps you choose the project and compare it with alternatives.

For authorizations and verification of actions, the reference is superintelligence, alignment and control. To decide whether to try an available function or prepare the process first, continue with waiting for superintelligence or using AI in SMEs.

The public evidence of 2026 deserves attention for what they document. A business decision can recognize that progress without declaring ASI or predicting its arrival. If you want to verify a function on your work, a project of AI agents can start from real examples, expected results and explicit acceptance criteria.