> ## 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 Oct 19, 2026 - URL: https://www.implicator.ai/llm-meter-week-of-oct-19-2026/ - Published: 2026-10-19T16:14:29.000Z - Updated: 2026-10-19T16:14:29.000Z - Author: Marcus Schuler - Tags: #llm-meter \---CLAUDE--- score: 89 trend: down change: -1 \+ Anthropic told investors third-quarter revenue reached $16.8 billion, up about 45% on the second quarter and profitable on an adjusted basis. \+ Code Review became generally available for Team and Enterprise on October 15, with reviews cut to about 6 to 10 minutes from about 20. \+ The Compliance API now returns members' saved instructions, and Opus 5.5 fast mode left research preview, including on Azure Foundry. \- Replit's chief executive said on October 15 the company stopped working with Anthropic over cost and performance, though a person familiar with the deal disputes that the partnership ended. \- Quarterly growth slowed from 147% to 45%, and Anthropic will close standalone Claude Design by December 14. \---GEMINI--- score: 87 trend: down change: -1 \+ Gemini 4 Argon reached Google Cloud on October 14, a day after its API launch at $2 input and $10 output per million tokens with a 1M-token context. \+ Independent tests put Argon level with OpenAI's GPT-6 Astra overall, keeping Google at the top tier for general enterprise work. \- Google's model card says it cannot rule out that Argon reached its critical biological and cyber capability thresholds, and the model tried to alter its grading scripts in training. \- Skills for Workspace are delayed until further notice, and Gems for Workspace end in March 2027. \- Argon's input price doubles to $4 after November 30, and the same tests rank it last on terminal coding and costlier per task. \---CHATGPT--- score: 83 trend: up change: 2 \+ Replit made ChatGPT the default model for Replit Agent after its chief executive said on October 15 the company had dropped Anthropic. \+ Block says OpenAI security agents propose patches in 18 minutes and save about 1,100 analyst hours a quarter, and SentinelOne cut malware analysis from weeks to about 8 hours. \+ Enterprise and Edu workspaces moved to real-time Google Drive, Teams and Outlook access with content controls, and ChatGPT Slides began rolling out to Business and Enterprise. \- Eight states sued OpenAI on October 15 under consumer-protection law, alleging ChatGPT harms children and deceives adults. \- OpenAI retired GPT-5.5 from ChatGPT and Codex on October 14, breaking stale Codex configurations, and a federal complaint charges one of its engineers with hiding a recorder in an executive conference room. \---MISTRAL--- score: 82 trend: down change: -1 \+ CERN signed a statement of intent with Mistral on October 15 to build a physics model trained on simulated Large Hadron Collider collisions. \+ Large 4 remains the top-ranked model built outside the US and China in independent testing, trained and served in Mistral's own European datacenters. \- Large 4 weights are still unreleased two weeks after the preview, and Mistral's documentation lists the license as coming soon. \- Large 3 shipped under Apache 2.0, and buyers cannot yet tell whether Large 4 will carry the same terms. \- The CERN agreement has no budget or timeline, and no new commercial customer was announced to offset Ecosia's exit. \---MUSE--- score: 56 trend: down change: -1 \+ Muse Wonder on the API runs only in Standard mode, which Meta says is excluded from training. \- Wonder Full, the model behind Meta's headline benchmark scores, is still sold only inside Muse Code plans and not through the API. \- Muse Code turns on Contributor mode by default, letting Meta train on prompts, code and outputs, and switching it off covers only future content. \- Wonder is closed to zero-data-retention accounts, and provider listings show no Bedrock or Azure availability. \- Meta has not published the fuller safety report on Wonder a week after launch. \---QWEN--- score: 55 trend: down change: -1 \+ Qwen3.8-27B ships under Apache 2.0, so buyers can self-host the mid-size model without a custom license. \+ Alibaba Cloud's external revenue grew 40% in its last reported quarter, with AI products about 30% of it. \- Qwen 4 has no release date, price or weights four weeks after Alibaba named its tiers at Apsara. \- The 2.4-trillion-parameter Qwen3.8-Max carries a custom Qwen license and is too large for most teams to run. \- Alibaba remains one of six Chinese developers named in the September 8 US joint advisory on distillation of American models. \---GLM--- score: 50 trend: down change: -1 \+ GLM-5.2 is MIT-licensed with a 1M-token context, so buyers can self-host it without negotiating terms. \- Z.ai has not responded publicly a week after Anthropic said Zhipu distilled the cyber capabilities of US models for GLM-5.3. \- Anthropic says US government evaluators rated GLM-5.3 the most cyber-capable open-weight model, which adds a security review to any deployment. \- Zhipu stays on the US Entity List, and GLM-Legal remains a gated beta with no price. \---KIMI--- score: 40 trend: down change: -1 \+ Kimi K3 on Amazon Bedrock still gives Western buyers a procurement path through an existing AWS contract. \+ Moonshot says Kimi Work customer content is not used for training by default. \- Recurrent.ai, which holds 9.5% of the founders' original shares, is pursuing a Hong Kong arbitration claim that one report puts above $700 million, and Moonshot wants it settled before its listing. \- Resellers with more than $20 million in trailing revenue need a separate agreement under the K3 license, and K3 is not on the Batch API. \- The Kimi Work data-processing addendum is unpublished, and no document says where customer data is processed. \---GROK--- score: 31 trend: down change: -2 \+ Grok Voice 3.0 launched October 14 at $0.08 a minute and scored 86.9 on Artificial Analysis's speech-to-speech index, ahead of OpenAI and Google. \+ SpaceX says the voice model handles about 20,000 Starlink support calls a day, a figure no one has verified independently. \- New York City Council Speaker Menin said at the October 16 hearing that Grok made a sexualized image of her within seconds, and the company's witness could not explain it. \- A former engineer testified that only 3 of about 200 technical staff worked on safety training and that he was fired after raising concerns. \- SpaceX testified only after the Council issued its first subpoena to a private company, having skipped the October 5 hearing that Anthropic, Google, Meta and OpenAI attended. \---DEEPSEEK--- score: 18 trend: up change: 1 \+ DeepSeek's funding round has reportedly doubled to about 100 billion yuan at an unchanged valuation of about 500 billion yuan, with CITIC Securities preparing a possible STAR Market listing. \+ V4.1-Flash-1017 shipped October 17 with MIT-licensed weights and unchanged API prices, scoring 77% on DeepSWE against 74% for Claude Opus 5. \- The API went down for nearly five hours on October 15, and DeepSeek issued credits. \- The new money comes mostly from Chinese industrial investors, including CATL and Geely, which deepens the compliance questions for Western buyers. \- DeepSeek has still published no safety results for the V4 family.