Latest briefing
July 10, 2026 · 5 actus (site) · 5 actus (base)
Meta (the parent of Facebook, Instagram and WhatsApp) has just launched its first paid AI model dedicated to automated programming. Imagine an intern who not only writes code for you but can also take over your computer, open several windows in parallel, and run up to five of its...
π₯ Top story
01
Meta launches Muse Spark 1.1 and breaks its golden rule: a paid AI model to program on your behalf
Meta (the parent of Facebook, Instagram and WhatsApp) has just launched its first paid AI model dedicated to automated programming. Imagine an intern who not only writes code for you but can also take over your computer, open several windows in parallel, and run up to five of its virtual colleagues on the same problem - that is Muse Spark 1.1. The price: 1.25 dollars per million words read, 4.25 per million words written, with 20 dollars of free credits to test. It is the first time Meta, which has been living off the 'all free' philosophy with its Llama family, takes out its credit card to go after Anthropic and OpenAI clients. Mark Zuckerberg himself posted on X for the first time in 3 years to announce it - a signal of how strategic this is. For a developer or a small tech team, this is a new 'ready to use' assistant that arrives at a competitive price, especially compared to Claude Sonnet 5 which costs almost twice as much.
02
Microsoft confirms: OpenAI's GPT-5.6 will be the default engine for Word, Excel and PowerPoint
When you open Word tomorrow to draft a report, or Excel to analyse a spreadsheet, the help suggesting rewordings or formulas will come from GPT-5.6, the very latest OpenAI model. Microsoft confirmed it on July 9: GPT-5.6 becomes the 'preferred model' for the whole Microsoft 365 Copilot suite (Word, Excel, PowerPoint, Cowork). It also signals, in passing, that the 'breakup' rumours between Microsoft and OpenAI, repeated by Bloomberg in recent days, were overstated. The two houses remain tied for good. And good news on prices: the 'Luna' variant of GPT-5.6 is listed at 1 dollar per million words read - cheaper than the previous version. For a business, this is the signal that the editor of your favourite software is paying OpenAI to provide AI for you; for the curious, it is confirmation that the price war keeps pushing prices down.
03
OpenAI kills its Atlas browser: even with 400 million users, the AI browser did not catch on
If you have never heard of ChatGPT Atlas, you are not alone: it was OpenAI's web browser, launched in October 2025, which integrates ChatGPT directly into the sidebar and could theoretically browse and click for you. On July 9, 2026, OpenAI announced it will be definitively shut down on August 9 - less than a year after launch. The official reason: not enough adoption, despite 400 million weekly ChatGPT users who could have switched. The real lesson: OpenAI has decided to tighten the screw, to kill the 'extra products' (Sora 2 consumer already dead in April, adult mode on pause, and now Atlas) to focus on the enterprise business. For the general public, it is an admission that 'AI browsers' are not ready yet. For those hoping to delegate their Monday morning clicks to an AI, you will have to wait a bit more.
04
A 100% open source voice pipeline lets you build your own phone assistant 35 times faster
Today, if you want a voice assistant that takes your commands by phone, you pay OpenAI or ElevenLabs a few cents per minute. On July 1, 2026, Hugging Face and Cerebras posted a complete, free and open source blueprint to assemble the same kind of service yourself. The principle: you chain four free building blocks (voice activity detection, transcription, reasoning, speech response). The technical surprise: by running the AI brain on Cerebras specialised chips instead of a standard GPU, you get a response 35 times faster - therefore more natural to the ear, without the awkward silences. Even better: 9,000 Reachy Mini robots already use it in production, so it is proven. For a startup, a customer service line, an appointment booking chatbot, this is a credible alternative to paid vendors, without the per-minute cents cost.
05
Bespoke Labs raises 40 million dollars to train AI agents in hyper-realistic virtual offices
When you entrust a complex task to an AI agent today - 'prepare this loan file for me', 'fix this bug in production' - it often gets stuck after a few steps. American startup Bespoke Labs has a thesis: this is not a model problem, it is a training problem. So it builds hyper-realistic virtual offices (millions of lines of code, microservices, logs, emails, Slack, Jira tickets) where agents train to chain the right actions on long tasks. On July 6, it raised 40 million dollars combining its seed and Series A, with heavyweights from Anthropic, OpenAI and Meta as backers. Why are the labs themselves betting on it? Because they need third-party 'sandboxes' to test and improve their own agents. For a team that wants to deploy agents in production, this is the signal that the ecosystem is maturing around a well-identified problem: reliability.
π‘ To watch
Hugging Face makes inference 5 to 20 times faster without changing a single line of code
On July 8, Hugging Face published two structuring pieces. First: 'Data for Agents', a reflection on the data needed to train a quality agent (trajectories, failures, tool calls, world state). Second: a 'vLLM backend for Transformers' that lets you keep your model loading code but benefit from vLLM speed - up to 20 times faster. For teams serving models in production without time to migrate everything, this is an immediate gain.
OpenAI's product chief steps down: signal of discipline or risk for the IPO?
Fidji Simo, CEO of AGI Deployment at OpenAI (formerly CEO of Instacart), moved to part-time advisor on July 9 for medical reasons (POTS, a neuro-immune condition). Greg Brockman, co-founder, takes back product leadership. The announcement lands on the same day as Atlas's death and GPT-5.6's launch - three aligned signals confirming that OpenAI is tightening the B2B focus. For enterprise customers, this is a discipline signal (end of products that don't work) but also a 3-6 month period of uncertainty before a successor is named.
Tencent integrates its open-weight Hy3 model into 50 internal products in 48 hours
The official @TencentHunyuan account confirmed on July 8 that 50 Tencent products (WorkBuddy, CodeBuddy, Yuanbao, QQ Browser, ima, Marvis) already integrate the Hy3 model - officially launched 48 hours earlier. 90% completion rate on internal apps, Apache 2.0 commercial licence, floor price 0.14 dollar per million words read. For teams looking for a cheap Chinese open-weight alternative, this is enterprise validation by the largest deployer of Chinese models in the world.
Will Anthropic respond on prices after August 31?
Anthropic keeps Claude Sonnet 5 at 2 dollars per million words read as a promotional rate until August 31, 2026, then switches to 3 dollars standard. With GPT-5.6 Luna at 1 dollar and Muse Spark 1.1 at 1.25 dollars, the pressure is huge for Anthropic to extend or renew the promo. To watch in the next 30 days: Anthropic pricing announcement, and possible similar drop from Google Gemini 2.5 Pro.
π Trend
July 9, 2026 will be remembered as a turning point for the AI agent industry, with three signs coming down on the same day. First sign: Meta finally enters the paid coding agent market with Muse Spark 1.1 - the group's first paid model, 1.25 dollar per million words read, distributed from the start in Instagram/WhatsApp/Facebook (3 billion users potentially reached). Second sign: OpenAI locks in its relationship with Microsoft by officialising GPT-5.6 as the preferred model of Microsoft 365 Copilot, and kills ChatGPT Atlas the same day - an admission that AI browsers are not ready, even with 400 million users available. Third sign: OpenAI's product chief, Fidji Simo, leaves her operational role the same day, confirming the 'kill the side quests, bet on B2B' discipline. In parallel, two structuring pieces of news for the months ahead: (a) Hugging Face and Cerebras show that you can now run a voice agent 35 times faster in open source, which democratises the creation of phone assistants; (b) Bespoke Labs raises 40 million to become the 'third-party laboratory' where labs (Anthropic, OpenAI, Meta) test and train their agents before production. The cross-cutting lesson: AI agents in July 2026 have moved out of the 'impressive demo' phase into the 'platform in production' phase - with, in the background, a price war that systematically pushes tariffs down.