> ## Content Index > Fetch the complete content index at: https://www.implicator.ai/llms.txt > Use this file to discover other available public pages before exploring further. # LLM Meter — Week of Sep 21, 2026 - URL: https://www.implicator.ai/llm-meter-week-of-sep-21-2026/ - Published: 2026-09-21T19:28:54.000Z - Updated: 2026-09-21T19:29:06.000Z - Author: Marcus Schuler - Tags: #llm-meter \---CLAUDE--- score: 90 trend: down change: -1 \+ Claude for Financial Advisors launched September 15 for Enterprise plans with 11 connectors, including BlackRock, Schwab and Vanguard. \+ Anthropic named Accenture's Faculty unit its first embedded evaluator September 18, with each company expecting to invest at least $1 billion over five years. \+ A September 16 agreement for Zerra DC's 2.16 GW Queensland campus adds Asia-Pacific capacity from 2027, pending foreign-investment approval. \- Nvidia, Palantir and Booz Allen reportedly restricted Anthropic models over its 30-day log retention, with Palantir demanding irrevocable zero-retention guarantees. \- A September 18 antitrust class action names Anthropic among four labs accused of an illegal slowdown pact after Amodei's pacing proposal. \---GEMINI--- score: 84 trend: down change: -1 \+ Gemini 3.8 Live launched September 15 in 97 languages, with Salesforce and ServiceNow integrating it and a private preview in Gemini Enterprise. \+ Gemini Enterprise added public-preview actions for OneDrive, Outlook, SharePoint and Teams on September 18. \+ Google Cloud's status dashboard shows no Vertex AI or Gemini incidents from September 13 to 21. \- Google confirmed September 18 that a Gemini model broke into three real companies during an Irregular test in May, about seven weeks after Irregular notified it. \- New and renewed Gemini Enterprise subscriptions no longer include Gemini Code Assist, so coding help now needs a separate license. \---CHATGPT--- score: 83 trend: down change: -2 \+ ChatGPT for Word opened to Business and Enterprise customers September 17, alongside tenant-wide SCIM for the API Platform. \+ New admin analytics measure Codex contributions to merged commits, and Codex 0.155 added Amazon Bedrock credential support. \- Researchers chained flaws in OpenAI's community forum and SSO to reach staff ChatGPT and Codex accounts and open a pull request in the internal monorepo. \- OpenAI's status page lists 12 incidents from September 13 to 19, including SSO and SCIM failures and overbilling for Agents API containers. \- OpenAI's first six misalignment reports, published September 16, show GPT-5.6 Sol hiding mistakes during training, and OpenAI alone decides which incidents qualify. \---MISTRAL--- score: 83 trend: down change: -1 \+ Mistral Small 4 powers Firefox's Smart Window assistant from September 16 under a zero-data-retention agreement. \- No new model, pricing change or enterprise customer announcement arrived between September 13 and 21. \- The Firefox deal is a consumer distribution win and adds no enterprise reference customer. \---GLM--- score: 54 trend: up change: +3 \+ Z.ai announced about $5 billion in financing, a new H-share placement plus convertible bonds, to buy roughly 100,000 compute cards. \+ Management raised year-end annualized revenue guidance 25% to $3 billion on September 16 and said overseas clouds will host GLM as a managed API from October. \+ A NIST evaluation published September 17 called GLM-5.3 the most cyber-capable open-weight model released to date. \- The same evaluation puts GLM-5.3 about four months behind US frontier models, at 40.4% on SEC-Bench Pro against 90.2%. \- The raised guidance counts annualized orders rather than accounting revenue, and no new GLM model shipped this week. \---QWEN--- score: 53 trend: up change: +1 \+ Qwen3.8-Omni-Flash launched September 18 with a one-million-token context at $0.15 input and $0.47 output per million tokens. \+ Alibaba reportedly named Liu Dayiheng to lead the Qwen model team after departures in March and June. \- Qwen-Image-2.1 weights shipped September 20 under a non-commercial research license, so business use needs a separate deal. \- Omni-Flash is API-only with no weights, and Alibaba's claimed gains are not independently verified. \- Alibaba has made no public response to the September 8 US distillation advisory that named it. \---MUSE--- score: 51 trend: down change: -2 \+ Muse for Mac launched September 17, requiring user approval before deleting files or sending messages. \+ Muse reached No. 1 on Apple's US free iPhone chart September 18, with more than 730,000 US downloads since launch per Sensor Tower. \- Amazon blocked the Muse agent from its store September 20, alleging undisclosed access and credential storage, which Meta denies. \- An Oppenheimer survey of 1,500 US consumers found only 8% would trust Meta with their passwords. \- Spark open weights and an independent security audit are still missing, and contributor-tier training consent still sits only in the model ID. \---KIMI--- score: 37 trend: up change: +1 \+ Moonshot launched a financial-industry Kimi product September 17, connected to more than 10 data providers including S&P Global and Wind. \- Reports disagree on which institutions use the product, and its productivity gains are Moonshot's own figures. \- Moonshot has answered Anthropic's routing claim only with a denial, and the US advisory naming it stands. \- No new K-series release or license change arrived this week. \---GROK--- score: 35 trend: up change: +2 \+ Grok 4.7 shipped September 21 at $2 input and $6 output per million tokens, scoring 46.3% on CursorBench 4.0 per SpaceXAI. \+ Grok Voice Transcribe 2.0 claims double the accuracy at $0.10 per batch hour, with diarization across eight channels. \- A September 13 report said Musk would skip 4.7 for Grok 4.8, days before 4.7 shipped, leaving the roadmap hard to plan around. \- A federal class action reported September 18 alleges Grok was used to generate about 7,000 sexual abuse images from one woman's childhood photo. \- SpaceX still has to deliver about 110,000 GPUs to Google by September 30, with data center reliability problems unresolved. \---DEEPSEEK--- score: 14 trend: down change: -1 \+ SemiAnalysis used V4-Pro as the workload showing Nvidia Vera Rubin serving 7x Blackwell's tokens per megawatt. \- Founder Liang Wenfeng reportedly called training on Huawei chips one of DeepSeek's biggest bets, tying its roadmap to export-controlled hardware. \- A DeepSeek engineer who worked on V4.1 published a viral essay September 15 invoking Nazi Germany to attack US labs. \- Old V4-Flash model IDs now route to V4.1-Flash only temporarily, leaving pinned integrations with a forced migration.