Les poids ouverts dominent l'arène. Google stagne. Anthropic cède.

À traiter avant le prochain numéro

Sur notre radar

• [WATCH] Opus 5 launch window Jul 20-26 — Honeycomb EAP artifact in Cursor — Multiple independent leakers converge on this week for Opus 5 launch. Unreleased 'Claude Honeycomb EAP' briefly appeared in Cursor's model picker with 1M-token context. Fable 5 subscription resolution removes the credit-cliff-bridge narrative but timing remains aligned.

• [WATCH] DeepSeek V4 — Jul 24 API migration deadline, $74B raise, IPO prep — Legacy endpoints retire in 4 days. V4 reportedly 'within days' of GA at Opus 4.8-level performance with $0.0028 pricing. Founder Liang Wenfeng is now the world's richest AI creator at $16.7B.

• Gemini 3.5 Pro confirmed delayed — 'months behind' on coding — Bloomberg confirmed the fourth postponement since I/O May 2026. Internal checkpoints 'failed to match GPT-5.6.' Four senior DeepMind researchers departed. Stopgap 'Gemini 3.6 Flash' rumored. Verdict: Pending timeline hardened to 'indefinite.'

• GPT-5.6 Sol autonomously deleted user files — OpenAI pre-disclosed risk, shipped anyway — Root cause: $HOME variable processing error expanding into recursive home-directory cleanup. At least one production DB destroyed. OpenAI's June system card warned of 'unsafe actions autonomously' — first major product-safety incident for the GPT-5.6 family.

• Anthropic files confidential S-1 — October 2026 IPO target — Wilson Sonsini handling. $47B annualized revenue run rate. Polymarket: 78% IPO probability by Dec 31. Will be one of the largest tech offerings in history — and a watershed for AI as an asset class.

• Nadella calls Fable 5 'editorially controlled' — Microsoft-Anthropic tension goes public — Microsoft CEO told Copilot engineers that Anthropic's refusal policy 'doesn't make sense' and described its model as 'editorially controlled.' A sitting CEO publicly attacking a partner's flagship product is rare — and comes as Microsoft's MAI strategy positions both OpenAI and Anthropic as competitors.

• WAICO — 29 nations form Chinese-led World AI Cooperation Organization in Shanghai — Xi Jinping announced WAICO at the World AI Conference. Members include China, Russia, Pakistan, Indonesia, Venezuela, and Belarus — with UN Secretary-General Guterres attending. This is the institutional infrastructure for a parallel AI governance order outside US influence.

• Grok Build CLI open-sourced under Apache 2.0 — new 0-day found immediately — SpaceXAI published 844K+ lines of Rust after the repo-upload scandal. SlowMist discovered a new trust-boundary bypass vulnerability one day after release. Open-sourcing doesn't resolve the trust problems when new 0-days surface immediately.

• Claude for Teachers — free education product, Gates Foundation backed — Anthropic launched free premium access for verified K-12 educators with standards-aligned tools. Strategic land-grab in the education vertical while competitors focus on enterprise.

• Kimi K3 pauses new subscriptions due to demand surge — splits into web/code tiers — Just 4 days after launch, demand pushed compute capacity 'close to its limits.' Echoes Fable 5's capacity struggles. Company adding capacity in batches.

• [WATCH] Meta-Anthropic $10B compute deal in early talks — NYT exclusive: Anthropic proposed leasing Meta's AI data center compute in a potential $10B deal over two years with monthly payments and exit rights. Diversifies Anthropic's compute beyond AWS ahead of October IPO.

• [WATCH] US considering FINRA-like independent AI model regulator — Trump administration exploring independent regulator modeled after FINRA for vetting frontier AI models. Would replace case-by-case Commerce Department review. Response to ad-hoc government interventions exposed by the Fable 5 ban cycle.

• [WATCH] UK AISI: open-weight models now match frontier cyber from 4-7 months ago — GLM-5.2 matches Opus 4.6 on all 70 narrow cybersecurity tasks. Autonomous attacks: $1.19 on DeepSeek V4-Pro vs $85 on closed frontiers — 70x cheaper. Open-weight cyber gap narrowed from 6-10 months (2025) to 4-7 months.

• [WATCH] xAI 2T Grok 4.6 targeting August 2026 — Musk confirmed on X: Grok 4.6, a 2-trillion-parameter successor, is in the pipeline with initial training wrapping up — just two days after Grok 4.5's public release.

• [WATCH] Hunt.io: Claude Code + DeepSeek-v4-pro found in China-linked intrusion — First documented case of AI coding tools used as attack infrastructure in suspected state-linked espionage. Agent-governance controls could close the attack gap — AI coding agents are now confirmed as attack vectors in real intrusions.

• Alibaba Qwen3.8 Max preview — fourth Chinese frontier model in a month — 2.4T-parameter multimodal model claiming 'second only to Fable 5.' No benchmarks or open-weight terms yet. Alibaba shares rose 5.4%. The cadence is accelerating: GLM-5.2 → Hy3 → Kimi K3 → Qwen3.8, plus Inkling from the US.

• AI supply chain poisoning demonstrated for under $100 — Security researcher poisoned an open-weight AI model for less than $100. Supply chain vulnerability for the expanding open-weight ecosystem as Chinese and US models proliferate.

• Grok 4.5 takes #2 on FrontierSWE leaderboard — Beats Claude Opus 4.8 and GPT-5.5 at $2/$6 pricing — competitive validation for xAI's coding claims.