Qwen3.8-Max, premier modèle de classe Max à poids ouverts
Les poids ouverts ont inondé la frontière cette semaine (Qwen3.8-Max est devenu le premier modèle de classe Max jamais publié à poids ouverts et Qwen3.8-27B est arrivé deux jours plus tard), tandis que la guerre des prix est devenue structurelle dans les deux sens : Anthropic a verrouillé Sonnet 5 à $2/$10, et DeepSeek est passé en GA en augmentant ses prix jusqu'à 1 100 %. Le postulat du modèle bon marché sur lequel nous avons fait tourner ce répertoire est mort ; Grok 4.6 à $2/$6 est son remplaçant.
438 announcements scanned · 11 mattered · 3 verdicts changed
Verdicts modifiés
Conditional
Grok 4.6
Launched Aug 12 at $2/$6 with a 61 AA Intelligence Index (ties GPT-5.6 Sol, trails only Fable 5 Max) and leading GDPVal-AA. A day-one launch with a single independent evaluator, so conditional rather than recommended.
Conditional
Qwen3.8-27B
Open weights shipped Aug 14 under Apache 2.0 with 262K native context and single-GPU deployment; vendor-claimed coding/agent gains await independent reproduction.
Acquired
Cursor
SpaceX formally closed its $60B Cursor acquisition on Aug 15; the conditional verdict holds.
À traiter avant le prochain numéro
Qwen3.8-Max Open Weights Countdown: Announced, Not Shipped
Alibaba confirms the 'next week' delivery window for 2.4T-MoE + 27B companion open weights; repo and license still pending as of Aug 10.
Détails et étapesGPT-5.6-Cyber: Fewer Refusals, Days After Astra Pause
Purpose-trained cybersecurity model answers 95% of advanced exploit requests; Daybreak Blue/Red program with hardware security keys Sep 1.
Détails et étapesSpaceXAI Launches Grok Bot: Persistent Agents That Run 24/7
First persistent multi-agent platform from the SpaceX-Cursor merger: agents with their own cloud VMs, from $120/mo.
Détails et étapesAnthropic Locks Sonnet 5 at $2/$10 as Price War Intensifies
Scraps the $3/$15 September hike, confirms the frontier price war.
Détails et étapesGemini 3.7 Flash: Google's Coding Workhorse at Half Price
$0.75/$3.75 intro pricing; 43.6% FrontierCode and 65.3% DeepSWE.
Détails et étapesGLM-5.3: Z.ai's Model Tops CyberGym, Weights in Two Weeks
CyberGym 84.5%, ahead of Mythos 5; open weights staged ~2 weeks behind a safety gate.
Détails et étapesMAI-Code-1.1-Flash Ships, MAI-Code-1-Flash Retires Sep 10
Hard deprecation date for MAI-Code-1-Flash across Copilot: a concrete migration step.
Détails et étapesDeepSeek V4 Pro GA: API Prices Reported Up to 1,100%
Peak/off-peak billing breaks the 'cheapest frontier-adjacent' premise; both DeepSeek verdicts under re-review.
Détails et étapesSpaceX Closes Cursor Acquisition: the Deal Is Final
The $60B deal is closed; cursor directory status moves to acquired.
Détails et étapesSur notre radar
Meta Ships Muse Glimmer: 30B Open-Weight Agent Model
Apache 2.0 30B dense model for always-on local agents on a single consumer GPU.
Anthropic Watermarks All New Claude Output: No Opt-Out
Machine-readable invisible watermarks + C2PA metadata, EU AI Act §50, applied globally.
OpenAI Adds Premium Seats to ChatGPT Business at $125/Month
5x usage capacity, no 5-hour cap: enterprise cost-control tier.
Qwen3.8-Max API Pricing Ships at $2/$6 per 1M Tokens
The flagship's pricing is live; its 2.4T-A95B open weights shipped Aug 12 and the 27B companion Aug 14.
Anthropic August 2026 Risk Report
Misalignment upgraded very-low→low; internal 'Model 2' disclosed (62.8% CoBench); CoBench saturated; 11-month bioweapons-classifier gap.
Stealing Reasoning Traces: Cross-Lab API Flaw
MATS/Max Planck/ELLIS/Snyk paper decodes encrypted reasoning blocks from OpenAI/Anthropic/Google APIs: 182 credentials + 367 PII recovered from public trajectories.
U.S. 'Pick Sides' AI Letter to Allies
State Dept. draft warns 35 signatories they can't join both the US-led and China-led AI coalitions.
SSI Sutskever Model
Still no release; Nvidia $5B stake is context only.
Gemini 4 / Brin-DeepMind Timeline
3.7 Flash is a workhorse, not the flagship; Gemini 4 has no date.
Cognition $40B Valuation Talks
Devin maker reportedly raising at $40B (unconfirmed).
Moonshot AI $50B / HK IPO
Contradictory valuations ($34.9B–$50B) across sources.
Microsoft Maia 300 Chip
Second-gen custom silicon; unveil expected as early as September.
NVIDIA Nemotron 3.5 Lightning + NeMo Switchyard
Smaller specialized model + multi-model router; secondary-only evidence.
Anthropic 'Model 2' (Internal)
Internal, no release plans: procedural, not safety.
Anthropic Q2 Revenue >$11.5B, Profitable
First profit; IPO-track context only.
LiteLLM Malicious PyPI (TeamPCP)
Supply-chain attack; ~2,500 orgs and 434K CI/CD pipelines exposed.
Écarté
- Opinion/commentary: Guardian 'nationalize OpenAI/Anthropic', Hinton/Fei-Fei/Ng debate, 'AI getting cheaper' pieces.
- Anthropic 'Model 2 surpasses Mythos 5' prediction-market rumor, speculative, superseded by the official risk report.
- M&A not AI: SpaceX/Decart $6B talks, VideoVerse $250M fraud.
- Consumer policy/tooling: Twitch training opt-out, MiniMax Music 3, Mistral hosting GLM-5.2, Claude OpenClaw gym anecdote.
- OpenAI exec churn (Brad Lightcap departs, Dali Rajic hired as CRO): personnel moves, no directory entity impact.
- Consumer hardware wave: Google Pixel 11, ChatGPT Linux app, Apple China AI model, Suno Studio 2.0, Mesh Android, Sandbar ring (consumer products, not tool/model directory).
- Funding rounds: River AI $1.1B, Lovable $13.3B, Blacksmith $550M, Thrive $2B, Accel India $550M, Applied Compute $3B (below-threshold or no shipping product).
- Claude watermarks backlash: re-coverage of the Aug 11 published item, no new increment.
- Zuckerberg 6,500-word AI manifesto: op-ed/vision statement, not a product event.
- Nvidia $500B Wall Street financing (Blackstone/BlackRock): infrastructure financing, no model/tool change.
- OpenAI $7B employee tender offer at $852B: internal financial event.
- Simon Willison personal projects + 'GitHub Models retired' + 'Quoting Dario Amodei': personal tooling/opinion, no new facts.
- Google consumer/marketing content: Gemini watermark toggle, Sheets canvas, AMIE medical, Ads AI (non-directory).
- Kiana mystery model on Code Arena: speculative, no confirmed identity.
- Apple Qwen China integration + Siri publisher payments: consumer/market-specific.
- China $28T capital-markets AI push: macro commentary.
- Nvidia $5B SSI investment: funding a lab with no shipping product.
- Grok CSAM / explicit-imagery lawsuit: legal/consumer.
- Microsoft Copilot app consolidation + Mico retirement: consumer Copilot, not the GitHub Copilot entity.
- IBM-OpenAI enterprise partnership: go-to-market, no directory entity.
Notre avis : Deux ans de "l'IA devient moins chère chaque mois" se sont inversés en une semaine. Anthropic a verrouillé Sonnet 5 à $2/$10 pour défendre ses parts, et OpenAI avait déjà réduit Luna de 80 %, mais DeepSeek, le labo qui a inventé le postulat de la frontière bon marché, a augmenté ses prix API jusqu'à 1 100 % au passage en GA avec une facturation heures pleines/heures creuses. Pendant ce temps, les poids ouverts sont devenus réellement compétitifs face aux modèles fermés : Qwen3.8-Max a livré la première sortie ouverte de classe Max, et GLM-5.3 a placé ses poids derrière une barrière de sécurité d'environ 2 semaines. Nous n'avons testé aucun des nouveaux modèles de cette semaine : Grok 4.6 et Qwen3.8-27B arrivent en conditionnel parce que leurs benchmarks sont du premier jour et minces. Lisez une grille tarifaire, pas un titre.
Recevez le numéro 22
Les changements de verdict et les échéances de la semaine prochaine, dans votre boîte mail.