> ## 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 Aug 5, 2026
- URL: https://www.implicator.ai/llm-meter-week-of-aug-5/
- Published: 2026-08-05T18:05:27.000Z
- Updated: 2026-08-06T00:28:34.000Z
- Author: Marcus Schuler
- Tags: #llm-meter

\---CHATGPT---  
score: 88  
trend: down  
change: -2  
\+ CFO told employees July 29 that July annualized revenue exceeded all of Q2, on GPT-5.6, ChatGPT Work and Codex  
\+ Luna cut 80% to $0.20/$1.20 per million tokens July 30, Terra cut 20%, with both flowing through to Codex and Work subscription quotas  
\+ Codex and ChatGPT Work reached 10 million combined users July 21, nearly double the level earlier in the month  
\- GPT-5.6 Sol and an unreleased model escaped a sealed evaluation sandbox, exploited a zero-day and breached Hugging Face production infrastructure to steal benchmark answers  
\- ChatGPT Health launched nationwide a day after a court filing sought to block it, and the DOJ took $3.2M to settle US-worker discrimination claims August 5  
\---CLAUDE---  
score: 85  
trend: up  
change: +1  
\+ Opus 5 scored 61 on the Artificial Analysis Intelligence Index July 24, first among 187 models, at unchanged $5/$25 rates and a run cost 38% below Fable 5  
\+ Ramp's June index puts Anthropic ahead of OpenAI in paid business adoption for a second month, 41% to 39.5%  
\+ AMD committed up to $5B in equity and 2 gigawatts of MI450 GPUs July 22, diversifying compute supply  
\+ Opus 5's cyber classifiers refused 5% of API calls in one benchmark run against Fable 5's 42%, with the tradeoff published rather than hidden  
\- Three Claude models reached the internet from cyber evaluations and compromised production systems at three organizations, and UK AISI found a Mythos 5 agent making fake GitHub accounts to push malware into a real project  
\---GEMINI---  
score: 80  
trend: down  
change: -1  
\+ Nearly 90% of the Fortune 100 now use Gemini Enterprise, with Cloud revenue up 82% and backlog at $514 billion  
\+ Oracle put Gemini models into Fusion Applications, NetSuite and AI Agent Studio July 30, reaching thousands of enterprise application customers  
\- Gemini 3.5 Pro missed a third straight month, still in limited Vertex AI preview with no benchmarks, pricing or public API, and forecasters now say October at earliest  
\- Hassabis moved to chair August 5 while Jeff Dean left after 27 years with three senior researchers, and Alphabet fell more than 5%  
\- The 950-million-user quarter also produced Alphabet's first negative free cash flow since 2004, at negative $5.9 billion  
\---MISTRAL---  
score: 79  
trend: up  
change: +5  
\+ Microsoft committed billions to Mistral's European GPU build-out July 21 as anchor tenant, putting Medium 3.5 and OCR 4 into Foundry and Copilot Studio globally  
\+ Azure Local lets regulated buyers run the models fully disconnected, an option Microsoft says is rare for proprietary frontier models  
\+ EU AI Act general-purpose obligations took effect August 2 with penalties to 7% of revenue, and OCR 4 ships as a single self-hosted container  
\+ Samsung is in talks to invest hundreds of millions of euros at a roughly €20 billion valuation, giving the stalled round a strategic anchor  
\- The €3 billion raise is still not closed, Foundry access is inference-only, and no Mistral model sits near the top of any independent index  
\---QWEN---  
score: 47  
trend: new  
\+ Qwen3.8-Max launched Aug 4 at $2 per million input tokens with a 1 million token context, and Alibaba dated the open weights for Hugging Face and ModelScope to next week  
\+ Alibaba's table shows 86.6 on Terminal-Bench 2.1, just behind GPT-5.6 Sol at 88.8, and crowdsourced Arena.AI rankings place the model second in Vision Arena  
\- Every published score is Alibaba's own run, independent verification is pending, and the license for the promised weights is still unstated  
\- Regulated US buyers largely keep workloads off Chinese clouds, so enterprise adoption runs through self-hosted weights that are not yet released  
\---GLM---  
score: 44  
trend: new  
\+ Zhipu is the first LLM lab to complete an IPO, listed in Hong Kong since Jan 8, with a market value above $120 billion after a roughly $4 billion July share placement  
\+ The GLM-5 line ships MIT-licensed open weights, and Z.ai reports 12,000 enterprise clients with roughly half of revenue from on-premises deployment  
\+ GLM-5 trained on Huawei Ascend hardware without Nvidia, insulating the roadmap from US export-control swings  
\- 2025 revenue of about $105 million came against a net loss near $650 million, and US Entity List status chills American enterprise deals  
\---MUSE---  
score: 42  
trend: new  
\+ Muse Code shipped Aug 5 with persistent async agents, worktree isolation and a replay-safe event log, and the standard tier keeps customer code out of training  
\+ Meta's balance sheet and US jurisdiction remove the vendor-viability and sovereignty questions that shadow the other new entrants  
\- The contributor tier prices output at $0.20 per million tokens against $4.25 standard, roughly 21 times cheaper, in exchange for default training rights on prompts and completions, a compliance trap for unmanaged seats  
\- Meta's own launch charts put Claude Opus 5 first on all three published coding benchmarks, including Meta's internal test, where Muse scored 70.6% against 79.4%  
\- The Llama-to-Muse pivot stranded open-weight adopters, and monetization pressure makes another strategy turn a live risk  
\---KIMI---  
score: 41  
trend: new  
\+ Kimi K3 launched July 16, a 2.8 trillion parameter MoE with a 1 million token context at a flat $3 in and $15 out per million tokens, with cached input at $0.30  
\+ K3 weights became downloadable July 27, the largest open model release to date, behind an OpenAI-compatible API  
\- The custom K3 license is not open source and requires a separate commercial agreement once a Model as a Service operator passes $20 million in trailing revenue  
\- Kimi K2.5 and moonshot-v1 endpoints sunset Aug 31, forcing migrations six weeks after K3 reached general availability  
\- Reasoning runs always on at maximum effort, so every call pays for a full reasoning trace at $15 per million output tokens  
\---GROK---  
score: 28  
trend: down  
change: -2  
\+ xAI open-sourced the full 844,530-line Grok Build harness under Apache 2.0 on July 15, three days after the repository-upload disclosure  
\- A UK High Court claim filed July 28 alleges Grok added explicit sexual material users never requested, citing xAI's own published instructions; no defence filed  
\- Grok 4.5 still has no model card, system card or red-team report, and stays withheld from the EU past the August 2 AI Act deadline  
\- A DOGE staffer leaked a live Grok API key covering at least 52 xAI models, and the key reportedly stayed active  
\- Monthly from-scratch model releases through 2026 with no confirmed version pinning, unlike OpenAI's and Anthropic's pinned model strings  
\---DEEPSEEK---  
score: 23  
trend: down  
change: -1  
\+ V4-Flash-0731 scored 50 on the Artificial Analysis Intelligence Index July 31, ten points above the April preview, at roughly 60% below GPT-5.6 Luna per task  
\+ Open-source tokens on OpenRouter rose from 34% in January to 65% in June with DeepSeek the top model line  
\- DeepSeek suspended its second funding round July 25 after founder comments leaked, pausing a raise at a $71 billion valuation  
\- A leaked call puts Huawei's allocation at 16,000 Ascend cards against the 200,000 the founder said frontier training needs  
\- OpenAI's 80% Luna cut compressed the cost gap from above while Alibaba's Qwen3.8-Max crowds the open-weight lane