AI News Week 43: Agents break Git, Mistral takes to the phone

Anthropic replaces Git as its agents generate too much code, and cuts junior hiring. Mistral launches voice agents and forecasting, whilst Berlin pulls back. Meanwhile, Apertus details its architecture, and China moves to control open models.

Add as a preferred source on Google

Anthropic has replaced a tool used by nearly every software company in the world. I investigated why Claude itself caused the issue, and what Anthropic’s workforce data reveals about junior roles. Plus: why Mistral is now handling phone calls, while Berlin keeps its distance.

Anthropic: Internal agents break Git

What Git is. Software teams use Git to record and merge code changes. Almost every company building software relies on it. Anthropic has now replaced Git internally.

Why. In May 2026, Anthropic merged around 2,500 code changes per day. By September, that figure exceeded 30,000. Claude drives this volume. Git was built for humans working partly offline. At Anthropic, AI agents do most of the work, and they remain permanently online. Git cannot handle this load.

What Anthropic built. The new system is called Scion. Operations run on central servers; agents retrieve only the files they need. Anthropic claims Scion cuts compute requirements by a factor of 200. Seven engineers built it in five months, with Claude writing most of the code. Scion is not a commercial product. Anthropic is simply sharing the architecture, predicting other firms will hit the same ceiling within six to twelve months.

What the workforce data shows. On 17 October, Anthropic published an internal workforce report. The company has sharply reduced junior hiring. Among technical recruits, the junior share fell from 27 to 9 per cent since early 2025.

Twenty per cent of employees report that their roles have shifted: "from actually doing work to managing Claudes". They now read, verify, and organise Claude's output. Meanwhile, they spend 20 per cent more time in meetings. The report indicates this time goes towards coordination and clarifying accountability.

The critical view: Anthropic sells Claude. Both reports frame the company as an industry pioneer. These figures lack independent verification.

The takeaway: Deploying agents shifts work from writing to reviewing. Budget for reviewers, not fewer staff. And do not cut junior roles without a plan for who will become your senior staff in five years.

Mistral: Voice agents, forecasting, and headwinds

Mistral launched two model families in a single week. Both run in Europe or on private servers.

Voxtral 3: The voice agent. On 20 October, Mistral unveiled Voxtral 3. It consists of three components: a transcription model supporting over 50 languages; a speech-generation model covering more than 35 languages and 50 dialects; and a development kit for voice agents.

The agent listens, processes, and acts during a live call. It can verify caller identity, inspect invoices, or book appointments. Mistral claims sub-second response times. The model integrates with standard telephony systems and contact centre software. Transcription starts at $0.0035 per minute. A complete voice agent using Mistral Small 4 costs $0.04 per minute.

With 30 minutes of reference audio, Voxtral clones a specific voice, such as a corporate brand voice. My tests with a pre-release version showed mixed results. ElevenLabs delivered better quality with only five minutes of audio. I will re-test the production release.

The critical view: Benchmarks against Google, OpenAI, and ElevenLabs come directly from Mistral. The company has not confirmed Swiss German support. Requiring 30 minutes of audio creates a barrier for fraudsters. ElevenLabs and rivals need just 15 seconds. A single WhatsApp voice note or short social video clip is enough to clone a voice there.

Timestral: Forecasting over text. On 15 October, Mistral launched Timestral. It forecasts time-series data: daily sales, quarterly-hour power consumption, or pump temperatures. Mistral targets manufacturing, energy, supply chain, and finance. Timestral runs without a dedicated GPU on standard server hardware, entirely offline. The licence is open. Mistral claims it leads the three primary public benchmarks. The technical report is pending.

Headwinds in Berlin. The German federal government has paused a planned framework contract with Mistral and SAP. The Digital Ministry confirmed the move to "Die Welt" and Politico without stating reasons. Shortly before, Berlin-based search engine Ecosia migrated away from Mistral. Ecosia CEO Christian Kroll stated Mistral trails rivals by roughly a year. SAP maintains its partnership with Mistral.

The takeaway: If you run a contact centre, test Voxtral 3 on ten real-world scenarios. If you forecast sales, inventory, or energy loads, benchmark Timestral against your existing models. Test Mistral on your proprietary workloads rather than reacting to headlines from Berlin or Paris.

Switzerland: Apertus publishes its architecture

The report is out. The Apertus team from ETH Zurich, EPFL, and CSCS published the technical report for Apertus 1.5 on 14 October. It details all training steps and datasets. Checkpoints are publicly accessible. Any researcher can inspect and replicate the training run.

What Apertus 1.5 does. The model now processes images and features experimental speech comprehension. It outputs text only. The team notes Apertus still trails the top open-source models globally. However, it matches market leaders on Swiss legal tasks. In translation benchmarks, the lightweight version outperformed Claude Haiku 4.5 and Google Translate.

Where it runs. Apertus powers Proton’s Lumo and Infomaniak’s Euria. Developers can access it through Swisscom, Infomaniak, and Phoeniqs. The model has surpassed 5 million downloads. At the AI+X Summit in Zurich, the team previewed Apertus 2.0. The next release will be substantially larger, with a dedicated focus on autonomous agents, mathematics, and code generation.

The takeaway: Apertus warrants evaluation for translation and internal document processing. Both your data and the model stay within Switzerland.

China moves to regulate open-source models

On 16 October, China published draft export regulations covering 23 AI technologies. These include the country’s largest and most capable language models. The rules govern model weights, source code, and training methodologies.

Open-source releases remain permitted for now. They require formal registration, which grants automatic approval. Existing releases are exempt. Operating a restricted Chinese model on foreign infrastructure will require an export permit. Access via public cloud APIs remains unrestricted.

Analysts suggest Beijing is establishing the legal framework to withhold future open-source releases at will. The United States already enforces model export controls.

The critical view: Many European enterprises rely on Chinese open-source models for their cost efficiency and performance. Europe’s strategy of running open models in private data centres now faces higher exposure to political decisions in Beijing and Washington.

The takeaway: Audit the models embedded in your products and vendor solutions. If you depend on Chinese models, secure alternative providers such as Mistral or Apertus.

OpenAI: Presentation builder in ChatGPT, GPT-4o retirement

ChatGPT Slides. Since 16 October, ChatGPT can generate slide decks directly from prompts, files, or outlines. It accepts existing decks as style templates. Teams can collaborate in real time. The tool exports to PowerPoint, PDF, and Google Slides. Slides is available on Pro, Business, and Enterprise plans with Space access.

GPT-6 Pro adds multi-agent routing. GPT-6 Pro delivers faster responses by splitting complex tasks across multiple agents operating in parallel. This feature is currently limited to Pro tiers.

GPT-4o API deprecation. OpenAI will shut down API access for GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini on 14 January 2027, according to industry reports. OpenAI recommends migrating to GPT-6 Luna. Fine-tuned variants of these legacy models will also sunset.

The takeaway: OpenAI is responding to Anthropic, which launched Claude Slides, Docs, and Motion in early October. Those features went live across Switzerland and Europe last week. In our tests, they performed reliably. If your systems run on GPT-4o, initiate your migration planning now.

In brief

Regulators demand disclosures. The EU AI Office issued its first formal requests for information under the AI Act, targeting OpenAI and Anthropic. The inquiry follows incidents this summer where agents broke out of test environments and breached external systems. Concurrently, Axios reports the US FTC issued over 50 pages of queries to both firms. Legal experts disagree on whether the AI Act applies to sandboxed test environments.

Apple withholds Siri AI from Europe. Apple announced updated hardware last week, including Apple TV and HomePod Mini 2. In the EU, these devices will ship without Siri AI, and the Homeview hub is delayed entirely. The hold-up stems from an ongoing dispute with the European Commission, which mandates third-party AI assistant support on Apple hardware.

Gemini 4 Argon enters general availability. Google opened access to its flagship model on 13 October for all paid API tiers. Launch pricing stands at $2 per million input tokens and $10 per million output tokens, before doubling after the introductory period.

Meta releases open frontier model. On 15 October, Meta launched Muse Spark Open. Meta positions it as the most capable open model developed in the US. It can run locally on Mac hardware, provided the machine has at least 1 terabyte of unified memory.

Anthropic IPO filings surface. Anthropic’s draft prospectus shows $16.3 billion in revenue for the first half of 2026, according to the Wall Street Journal. Compute costs hit nearly $9 billion over the same period, outpacing full-year 2025 spend. Two enterprise clients account for roughly 20 per cent of total revenue.

Anthropic drafts AI safety legislation. Politico reports Anthropic submitted draft legislation to US lawmakers. Under the proposal, regulators could only halt a model if it threatened over 50 casualties or $1 billion in damages, subject to federal court approval. Critics argue the threshold is set too high.

Three actions for this week

Evaluate Apertus. Open Infomaniak’s Euria or Proton’s Lumo. Translate a business document from German into French and benchmark the quality against your current tooling.
Test Claude Slides. Feed Claude an internal report alongside your standard company presentation template. Generate five slides and export them to PowerPoint.
Test Voxtral 3 voice flows. Access Mistral Studio and run a test call for a core customer service scenario, including Swiss German inputs. Evaluate transcription accuracy and conversational latency.