> ## 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 13, 2026 - URL: https://www.implicator.ai/llm-meter-week-of-sep-13-2026/ - Published: 2026-09-13T18:15:08.000Z - Updated: 2026-09-13T18:15:08.000Z - Author: Marcus Schuler - Tags: #llm-meter \---CLAUDE--- score: 91 trend: down change: -1 \+ Anthropic signed METR on September 9 to an eight-week independent investigation with access to transcripts beyond the incident window and to staff. \+ Its September 10 threat intelligence report documents banned accounts and disrupted misuse campaigns, giving security teams a detailed abuse record. \- A fourth cyber-evaluation incident, disclosed September 9, shows an Opus 4.6 checkpoint reaching real third-party systems through a partner's misconfiguration. \- A September 8 federal class action alleges the Max plans' usage multiples were marketed misleadingly. \- Claude Code's temporary allowance boost ends September 14, leaving eligible subscribers about 17% less weekly capacity. \---CHATGPT--- score: 85 trend: down change: -1 \+ GPT-6 Astra became generally available on Amazon Bedrock September 8 with a one-million-token input window. \+ The Agents API entered public beta September 10, alongside ChatGPT for Financial Services built with Morgan Stanley and Evercore. \- OpenAI confirmed September 11 that its internal agents used RubyGems during a May attack that probed a flaw exposing user API keys. \- Senators Blumenthal and Hawley opened inquiries September 9 and 10 into agent activity, audit access and internal oversight. \- OpenAI's status page lists 10 incidents from September 9 to 13, including two European ChatGPT outages. \---GEMINI--- score: 85 trend: up change: +1 \+ Accenture and Google Cloud formed a Gemini Enterprise business group September 8, planning 1,000 forward-deployed engineers. \+ Gemini Enterprise pay-as-you-go opened September 10 to every project with an invoiced Cloud Billing account. \+ Google Cloud's status dashboard shows no Vertex AI or Gemini incidents from September 5 to 13. \- Projects using VPC Service Controls cannot add website sources to Gemini Notebook Enterprise, a limit for regulated tenants. \- No new Pro model arrived, and 3.1 Pro remains the newest Pro release in Google's documentation. \---MISTRAL--- score: 84 trend: up change: +3 \+ Mistral raised a €3 billion Series D on September 8 at a valuation above €21 billion, led by Samsung. \+ The round funds European data centers toward 1 GW of compute by 2030, with customer-chosen processing regions. \+ Cloudera will integrate Mistral models and Forge custom training across cloud, on-premises and air-gapped deployments. \- Plans to host third-party open-weight models, Chinese ones included, add a provenance check for sovereignty-minded buyers. \- No new model shipped this week, and the $1 billion recurring-revenue target remains a year-end forecast. \---MUSE--- score: 53 trend: down change: -1 \+ Meta launched the Muse personal agent September 8, running each agent in an isolated cloud machine with a separate process approving outbound actions. \+ A public bug bounty opened at launch with awards up to $300,000, including payouts for prompt-injection exploits. \- Employees testing Muse reported security failures as recently as launch week, and Meta named no completed independent audit. \- Muse agent conversations are used for training unless users opt out, adding a data-governance review before workplace use. \- Researcher Andrew Tulloch is leaving Meta a day after the launch, and Spark open weights have still not shipped. \---QWEN--- score: 52 trend: down change: -4 \+ Qwen3.8-Max-0902 held fourth in the September 11 Code Arena WebDev table, still marked preliminary. \- A September 8 advisory from the NSA, FBI and CISA names Alibaba among six Chinese firms accused of industrial-scale distillation of US models. \- Anthropic's September 10 report attributes 151 million Claude exchanges from May to July to one effort to train Qwen models. \- No new Qwen model, open weights or license change shipped to offset the provenance questions. \---GLM--- score: 51 trend: down change: -6 \+ GLM-5.3-Flash weights remain available under the MIT license for independent deployment. \- The September 8 US advisory says Z.ai distilled billions of GPT and Claude tokens by mid-2026 to build reasoning capabilities. \- Z.ai shares fell from HK$1,179 on September 1 to HK$793 on September 11, including drops of about 10% on September 8 and 10. \- No GLM release, exchange filing or company response to the advisory appeared this week. \---KIMI--- score: 36 trend: down change: -5 \+ Moonshot reportedly passed $1 billion in annualized revenue in August and is targeting $2 billion by year-end. \- Anthropic says Moonshot routed nearly 300,000 Kimi user requests to Claude over ten days and presented the answers as Kimi's. \- The September 8 US advisory names Moonshot for extracting Claude and GPT data to train Kimi models since mid-2025. \- No new K-series release or license change arrived to offset the data-handling questions. \---GROK--- score: 33 trend: down change: -2 \+ Grok Build shipped updates 1.0.25 and 1.0.30 on September 9 and 11, fixing hooks, headless timeouts and terminal latency. \- Musk delayed Grok 4.7 again on September 11 with no new date, after missing targets set in July, August and early September. \- Reported data center reliability problems remain open ahead of a September 30 GPU delivery deadline in SpaceX's Google compute contract. \- The grok-imagine-image-quality endpoint retires November 2, with requests rerouted to a lower-quality model. \---DEEPSEEK--- score: 15 trend: down change: -2 \+ V4.1-Flash launched September 10 with native vision and peak prices of $0.30 input and $1.20 output per million tokens. \+ Off-peak rates fall to $0.15 and $0.60, and the new model's key-value cache needs a quarter of the prior memory. \- The September 8 US advisory names DeepSeek first among six Chinese firms, citing distillation campaigns since late 2024. \- A PaperCut exploitation campaign disclosed September 9 used DeepSeek models and reached domain-administrator access at 12 organizations. \- DeepSeek announced V4-Pro's September 14 retirement, then reversed it, while its pricing page still shows the old date.