#41 What is actually coming in the next six months (and what nobody can promise)

Every AI talk ends with the same slide: a list of model names still to come. It is the least useful passage and the most watched. Here is the honest version, separating what is documented from what is announced, and replacing model names with the three movements that will actually reach your organisation.

On the models themselves, be wary of roadmaps

The autumn 2026 picture fits into ten days. On 1 September, Anthropic released Claude Fable 5.1 and its invitation-only twin Mythos 5.1, and cut cache-read pricing by 75 %. On the 2nd, Google shipped Gemini 3.8 Flash, with a restricted-access Cyber variant. On the 3rd, OpenAI released GPT-6 under the name Astra. Five frontier launches in ten days, in a month where every major lab shipped a model with cybersecurity capability.

Let us be precise about the nature of this acceleration, because it is often misread. What the labs ship are language models. Generic capability, not ready-made operational systems. Genuinely autonomous AI remains confined today to narrow, access-controlled domains, defence and research. Between raw capability and a system actually running inside your organisation sits all the integration, framing and verification work, and nobody is going to deliver that part to you.

One detail deserves your attention more than the names. GPT-6 Astra is the first model to cross the « critical cyber » safeguard threshold that its own maker had set. In other words, the labs are starting to run into the limits they drew themselves. This is not a reason to panic, it is an indicator of pace.

Two cautions. First, these calendars slip constantly and the names change every quarter: the paragraph you have just read will be out of date before the end of the year. Second, a large share of the available information on roadmaps comes from specialist blogs and trackers, not from official communications. Any conference that gives you a precise date for an unannounced model is selling you a prediction, not a fact.

The useful point lies elsewhere. Over the past eighteen months, the capability gap between the best model and the third best has narrowed while prices collapsed. For a small business or a local authority, this means the choice of supplier has become a reversible and secondary decision. That is not where your success will be decided.

Movement 1: agents are leaving the demo stage

The capability known as « computer use » lets a model drive a browser or a workstation the way a human would: look at the screen, click, fill in a form, read the result. It exists at Anthropic and at OpenAI, and Astra is explicitly presented as oriented towards this use.

In 2026, this capability has moved from demonstration to supervised use. The friction point is no longer model performance, it is operational: authentication, session stability, traceability of actions, and the question of who answers when the agent gets it wrong. Remember the phrase circulating in the sector, probably the most accurate observation of the year: models are now improving faster than organisations can adopt them.

Movement 2: voice becomes a commodity

On 7 May 2026, ElevenLabs cut its prices by 55 % on speech synthesis, 45 % on recognition and 20 % on its agents. This is not a promotion, it is a signal of commoditisation.

Look at the technical specifications rather than the prices. The Flash v2.5 model claims 75 milliseconds of latency, and Scribe v2 Realtime under 150 milliseconds across more than 90 languages. The threshold that matters sits around 200 to 300 milliseconds: below it, an exchange stops being perceived as an interaction with a machine and becomes a conversation. We have just crossed that threshold, in production, at a price a town of 8,000 people can afford.

And this is not only a developer API story. Live Translation on Apple's AirPods, delayed in the European Union for Digital Markets Act compliance and then opened with iOS 26.2 in December, runs in French with on-device processing. The threshold is being crossed in consumer hardware too, already in your users' pockets. It is the most important technical fact of the year for public-facing services, and it went almost unnoticed.

Movement 3: real adoption, far behind the talk

Adoption figures depend heavily on who publishes them and on what counts as « using AI », so treat them as orders of magnitude. The France Num barometer puts AI use at around 34 % of French small businesses, against roughly 13 % a year earlier. Bpifrance expects close to 58 % of companies to have a project under way in 2026. The progression is real and fast, but it mostly describes individual, untooled use, not transformed processes.

In other words: most organisations are still letting their staff use a consumer assistant in a personal browser, with no framework, no policy and no traceability. That is where the immediate risk sits, long before the science-fiction scenarios.

The 2 August 2026 deadline, and what it actually covers

It has already passed, but it does not say what people make it say. Since 2 August 2026, the Article 50 transparency obligations apply, along with the rules on general-purpose models, the governance and penalties regime, and the enforcement powers of national authorities. The AI literacy obligation, in force since February 2025, becomes fully enforceable: you must be able to demonstrate that your teams have been briefed on the possibilities, the limits and the risks.

The heavy part, however, has moved. The Digital Omnibus, in force since 27 July 2026, defers the obligations applying to high-risk systems to 2 December 2027 for stand-alone systems and 2 August 2028 for those embedded in products. The reason is prosaic: the harmonised standards meant to let providers demonstrate conformity were not ready, and neither was the assessment infrastructure. Deferred does not mean cancelled, but it does mean the heaviest part of the regulation is not yet in front of you.

As for actual enforcement, it is uneven. Twenty-four of the twenty-seven member states have designated their competent authority, the CNIL in France, but only a minority show advanced implementation. The penalty ceilings circulate widely, up to 15 million euros or 3 % of worldwide turnover on transparency and general-purpose models. For a local authority or a small business, though, the issue is not the fine: it is the burden of proof. Proof is built before the incident, never after. If you deploy a voice reception agent this autumn, you should already be able to answer three questions: is it disclosed, is it logged, who has been trained.

So what now?

Take three things from this article. The model name carries no strategic importance, and what arrives from the labs remains generic capability, not a turnkey system. Real-time voice has become affordable, on the API side as well as in consumer hardware, and that is the real change of the year. And your regulatory constraint is real but more spread out than people say: what concerns you today is transparency and training, the rest arrives in 2027 and 2028.

The next article looks at the ground where all of this is already visible, and where computing is starting to disappear for good: healthcare.