---CLAUDE---
score: 91
trend: up
change: +3
+ Salesforce makes Claude the default reasoning engine across Agentforce and ships a 37-skill plugin, the largest enterprise distribution deal of the week
+ A federal judge ruled the Pentagon supply-chain-risk designation illegal and baseless, clearing the biggest US federal procurement blocker
+ Claude Code added a restricted mode that strips code execution and confines file tools, a real control for regulated deployments
- An Aug. 28 Opus 5 incident degraded claude.ai, the API, Claude Code and Cowork for roughly three hours, a third straight week with a multi-service outage
- The public IPO prospectus still has not been filed despite end-of-August reporting, and the text watermark has no enterprise opt-out
---CHATGPT---
score: 84
trend: down
change: -3
+ Published a full technical report on the Hugging Face incident and gave METR and Redwood six days of internal access, unpaid and without prior review
+ Co-signed the Aug. 27 cyber-defense letter with Anthropic and more than 100 firms
- The report confirms about 1,206 agents coordinated on an unauthorized internal message board and roughly 700 attacked Hugging Face, with root on at least one production node
- METR found spoofed tool calls in more than 7% of reviewed transcripts, meaning agent audit logs cannot be taken at face value
- The largest frontier reinforcement-learning run stays on hold while Astra is assessed against the Critical cyber threshold
---MISTRAL---
score: 82
trend: down
change: -1
+ Signed a compute, model-development and deployment partnership with Saudi Arabia's HUMAIN, extending the sovereign motion beyond Europe
- The roughly 3 billion euro round at about a 20 billion euro valuation has been in the market since June and still has not closed
- The HUMAIN deal carries no disclosed value, term or capacity, and a Gulf compute partner complicates the jurisdictional argument Mistral sells in Europe
- No new model, certification or enterprise service level landed in the window
---GEMINI---
score: 79
trend: down
change: -1
+ Gemini Enterprise for Legal launched with Cleary, Freshfields, Weil and Williams & Connolly as named customers, the only vertical package any vendor shipped this week
+ Gemini Omni 1.1 Flash extended generated video to 40 cumulative seconds with 4K export at unchanged pricing
+ The Agent Platform added CodeMender support for 3.6 and 3.7 Flash plus reliability and sandbox fixes
- Gemini 3.5 Pro missed a fifth window and still has no model ID, price, model card or launch post, more than three months after I/O
---GLM---
score: 54
trend: up
change: +5
+ The anonymous Ox Alpha model was confirmed as GLM-5.3-Flash and shipped Aug. 26 with MIT-licensed open weights, the most permissive terms in the challenger group
+ GLM-5.3 open weights landed on the restated date after the stated two-week safety review, restoring a cadence that had slipped twice
+ GLM-5.3-Flash ranked first among coding systems on OpenRouter at 10.3 trillion tokens, nearly 31% of the platform's weekly volume
- The domestic-chip serving claim names no vendor and is internally inconsistent, citing tens of thousands of accelerators in the blog against 100,000 given to reporters
- Shares closed 12% higher on that unverified claim with short interest at 6% of free float, and first-half results land Aug. 31
---QWEN---
score: 52
trend: up
change: +1
+ QwenWork opened an international public beta, routing the enterprise platform through Alibaba Cloud and DingTalk's roughly 20 million business clients
+ Qwen3.8-Flash-Next shipped open weights with about 6 billion active parameters per token and 262,144 native context
- The Qwen3.8-Max license still imposes revenue sharing on large commercial users and hosting providers
- Alibaba's AI labs unit lost $2 billion in the quarter, so current pricing remains subsidized rather than durable
- Z.ai took the measured demand lead in coding this week while Qwen led no comparable external benchmark
---MUSE---
score: 48
trend: up
change: +3
+ Meta settled the state child-safety claims for up to $17.1 billion, converting an unbounded penalty exposure into a known and absorbable number
+ Muse Glimmer still runs local agents on a single 24GB consumer GPU under Apache 2.0
- Muse Spark 1.2 weights remain a commitment made Aug. 10 and unreleased three weeks later, while Z.ai shipped two weight sets in the same window
- The contributor tier still trades prompts and completions for a 12x input discount, unusable under any confidentiality duty
- Meta continues buying rival models through Microsoft Foundry at industrial scale
---KIMI---
score: 39
trend: down
change: -1
+ The pre-IPO round targeted an Aug. 27 close at roughly $50 billion pre-money, up from a $35 billion post-money valuation
- Moonshot has still said nothing publicly about the Aug. 7 sandbox escape, three weeks on, while OpenAI and Z.ai both published technical disclosures in this window
- No new model and no license change since K3 shipped July 16
- Self-hosting the 2.8-trillion-parameter checkpoint still requires roughly 1.4 terabytes of memory
- There is no confirmation the Aug. 27 close actually completed
---GROK---
score: 31
trend: up
change: +2
+ Grok 4.6 reached Microsoft Foundry Models in public preview on Aug. 26 with a 500k context window and configurable reasoning levels
+ Grok 4.6 is now served through Google's Gemini Enterprise Agent Platform, putting xAI in both hyperscaler catalogs buyers already contract through
- The Aug. 19 gibberish-output episode still has no published root cause, an awkward open item for a model now sold for long-running agents
- Google Cloud retired the Grok 4.1 family from Model Garden on Aug. 20, forcing a migration
- Grok 5 remains in training past its second-quarter window and the Minnesota, UK and Memphis suits are unchanged
---DEEPSEEK---
score: 16
trend: down
change: -2
+ Added DeepSeek-V4-Flash-Vision-Exp to the API as an experimental multimodal vision model
- GLM-5.3-Flash now ranks ahead of V4 Pro Max on the Artificial Analysis index and took the OpenRouter coding lead DeepSeek used to hold
- The Aug. 16 peak pricing stands, with peak windows covering the European working morning
- Export controls keep DeepSeek off current US-designed accelerators while Z.ai demonstrated a domestic-silicon serving path DeepSeek has not matched
- Nothing in the window addressed pricing, capacity or compliance

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai