Latest briefing
July 9, 2026 · 5 actus (site) · 6 actus (base)
When you talk to Siri or a hotline bot today, you have learned to stay quiet while it answers - otherwise it does not hear you. It is like a walkie-talkie: you press, you speak, you release, the other speaks. On July 8, 2026, OpenAI launched 'GPT-Live', a new generation of voice ...
π₯ Top story
01
OpenAI launches GPT-Live: no more walkie-talkie, your voice assistants will now speak AND listen at the same time
When you talk to Siri or a hotline bot today, you have learned to stay quiet while it answers - otherwise it does not hear you. It is like a walkie-talkie: you press, you speak, you release, the other speaks. On July 8, 2026, OpenAI launched 'GPT-Live', a new generation of voice agents that break this rule: they listen AND reply at the same time, like in a real conversation between humans. You can even throw in a small 'mhmm' to nudge them without cutting them off. The technique: two layers of AI in parallel - a fast one that handles the voice, a deeper one (GPT-5.5) that prepares the answer in the background. The result: 150 million people already use ChatGPT by voice every week, and this is the new model that every customer service line, hotel booking desk, or bank call center will need to copy to avoid looking like a relic. For a company deploying phone agents, this is the new technical standard to bake in.
02
xAI and Cursor release Grok 4.5: an 'AI' model built on real interactions of millions of developers (and not just for code)
Imagine a car maker wanting to build the ideal car - but instead of testing it in the lab, it hands the prototype to taxi drivers for a year, and watches what they do, what jams, what they would fix. That is exactly what xAI (Elon Musk's company behind Grok) and Cursor (the coding tool used by millions of developers worldwide) did. On July 8, 2026, they launched Grok 4.5, an AI model trained on trillions of real interactions from coders. The surprise: it is not ONLY for coding. It also targets finance, law, data analysis, multi-day long-running tasks. It is available today in Cursor at 2 dollars per million input words and 6 per million output words. Early feedback shows Cursor subscribers doubled their usage in a single week. For a tech team, this is a new 'off the shelf' model that can replace several specialised tools.
03
SambaNova raises 1 billion dollars and signs JPMorgan: banks want their own AI machines, not OpenAI's
When you deposit money at the bank, you would rather the IT system managing your accounts run INSIDE the bank, behind real walls, than on a faraway server nobody knows where it is. On July 8, 2026, SambaNova, a California company that builds specialised chips to run AI, announced a major funding round - 1 billion dollars, doubling its value in 5 months - and crucially a deal with JPMorgan Chase, the largest American bank. The deal: JPMorgan will install SambaNova machines on its own premises to run its AI agents without sending sensitive data outside. For a bank, an insurer, a hospital, a government ministry, this is the start of a heavy trend: 'sovereign AI infrastructures' (Norwegian, Quebec, French, you name it) running locally, without depending on an American cloud. For citizens, it is often good news: your banking data can now be processed by AI without ever leaving the country.
04
A small US community bank puts 40% of its loan processing in the 'hands' of AI agents - and keeps humans in the driver's seat
When you apply for a bank loan - for a car, a house - there is today one employee who checks your file, another who assesses the risk, another who drafts the contract. In the United States, over 2,400 small banks use the software of the company Abrigo (based in North Carolina). On July 8, 2026, Abrigo launched APX, a new platform where 'fleets of AI agents' cooperate to automate this whole journey - pulling the file, cross-checking controls, drafting the offer, tracking repayment. The key point: every step is logged in a journal readable by banking regulators, and a human can take back control at any time. Announced result: -40% manual work on files. For a broker, a small loan officer, a bank executive, this is a signal that 'vertical' platforms specialised in one single profession have reached a maturity that all-purpose demos never did.
05
OpenAI audits the reference test for 'coding agents': almost a third of the exercises are broken
When we say 'this AI model is as good as a senior developer', we usually rely on public test batteries. One of the most used is called SWE-Bench Pro. The problem: on July 8, 2026, OpenAI did the commendable but awkward thing of auditing this test. Result: about 30% of the 731 proposed exercises are flawed (bad subject, 'right answer' that is not one, badly specified tests). OpenAI is withdrawing its recommendation to use this benchmark and invites the community to build better ones, evaluated by experienced software engineers. For a company picking an AI model for its developer team, this is a brutal reminder: the performance figures published by vendors (Cursor, Anthropic, Google, OpenAI...) on this test are partly artefacts. The 'my AI is the best at coding' leaderboard has become 'my AI is the least bad at the exercises that remain'.
π‘ To watch
Will business voice agents adopt OpenAI's full-duplex architecture within 60 days?
Companies like Vapi, Retell or Bland build voice agents for businesses (hotel booking, customer support, appointment setting). If they integrate simultaneous listen-and-speak by end of August 2026, the market standard flips and GPT-Live becomes the reference. If not, OpenAI keeps a year of lead. To watch: the changelogs of these platforms over the next 30-60 days.
Two vertical 'agent fleet' platforms launched the same day - signal or coincidence?
On July 8, Abrigo launched its platform for banking loans, and Akeneo its agentic layer for product information management (PIM). Two established vertical vendors, same pattern, same day. If a third vertical platform (healthcare, legal, logistics) follows within 14 days, this is a real market move. If nothing emerges, it is a one-off convergence.
Will OpenAI's audit of SWE-Bench Pro become a new industry norm?
OpenAI published its audit methodology on July 8. The question is whether Anthropic, Google DeepMind and Cursor will in turn publish audits of their internal benchmarks. If yes, this is a new industry hygiene. If no, OpenAI risks paying for its transparency effort by looking like 'the only one self-criticising'.
Will SambaNova sign other banks or insurers after JPMorgan?
The JPMorgan-SambaNova announcement came alongside a Qatar Investment Authority backer in the funding round. This is a signal that international finance takes 'on-prem AI' seriously. If 2 to 5 other big names (bank, defense, healthcare) follow by end of 2026, this is a 30-50 billion dollar investment vertical materialising. If only JPMorgan signs, it stays an isolated - but very symbolic - case.
Akeneo Agentic Ziggy in commerce: who signs first?
Akeneo launched on July 8 Agentic Ziggy, agent orchestration for product information management, in a market where Google Shopping Optimization and ChatGPT Shopping are rewriting the rules of product visibility. The first retailer or brand signings in the next 30 days will validate (or not) that 'vertical agent fleets' have become the new commerce norm. To watch: Akeneo customer changelogs, responses from Pimcore, Informatica or Salsify.
π Trend
July 9, 2026 sees three major axes of agentic AI converging. First axis: voice becomes natural - OpenAI launches GPT-Live that speaks AND listens at the same time, removes the walkie-talkie barrier and ships the technical blueprint that every business voice agent will now copy. Second axis: the model is no longer a standalone product - xAI and Cursor co-launch Grok 4.5, training paid for by real-usage data and distributed exclusively through the leading agentic editor, the equivalent of the Microsoft-OpenAI partnership for the agent era. Third axis: the infrastructure is getting verticalised - SambaNova raises 1 billion dollars and signs with JPMorgan to put AI behind bank walls, while Akeneo and Abrigo launch on the same day their 'agent fleet' platforms for commerce and credit. The common point: everyone moves from 'one-off demo' to 'platform in production'. And behind it, OpenAI makes a rare act of transparency: auditing publicly a reference benchmark, showing that 30% of the exercises are broken, and inviting the industry to do better. The cross-cutting lesson: AI agents in 2026 are no longer a gadget to test - they are critical infrastructure that forces vendors to be at once more performant (models), more reliable (benchmarks), more audible (vertical platforms) and more respectful of local constraints (on-prem). Tomorrow's watchlist: B2B voice agent adoption, audit responses from Anthropic and Google, and post-JPMorgan bank signings.