LLM Meter — Week of June 14

---CHATGPT--- score: 87 trend: up change: +1 + OpenAI is weighing aggressive token price cuts (WSJ) that lower enterprise costs, pressing its advantage as Anthropic's enterprise run stumbles + Three-cloud sourcing across AWS, Azure, and on-prem MCP is exactly the multi-cloud reach the former leader just lost to a government order + GPT-5.5 (Instant, Thinking, Pro) is fully live across Plus, Business, and Enterprise, with GPT-5.2 retired June 12 on clean auto-migration - GPT-5.5 still trails Opus 4.8 by about ten points on SWE-bench Pro, ceding the coding-agent quality lead - About $14B in projected 2026 losses against a vast compute commitment keeps the financial overhang in place ---GEMINI--- score: 86 trend: down change: -1 + The Gemini Enterprise Agent Platform remains the deepest enterprise stack on the board, with Salesforce Agentforce, Databricks, Ramp, and Xero aboard + Gemini 3.5 Flash anchors production price/performance and is undercutting both US rivals on token cost - Gemini 3.5 Pro (2M context, Deep Think) slips to general availability a third straight week, still in limited Vertex preview - A quiet week cedes the top spot to ChatGPT even as Claude stumbles, a momentum gap rather than a platform one - Pentagon classified-network and DeepMind defense-work questions remain unresolved ---CLAUDE--- score: 82 trend: down change: -8 + Opus 4.8 keeps the enterprise coding crown and the compliance stack (ISO 42001, FedRAMP, HIPAA) is unmatched + Every other Claude model stayed online across multiple clouds, and the October IPO track is the strongest durability signal on the board - The Commerce Department ordered Fable 5 and Mythos 5 pulled June 12, and Anthropic disabled both globally, the first forced takedown of a generally available US model - Top-tier models that vanish on a government letter are a continuity risk no SLA covers, with foreign-national access, including Anthropic's own staff, now blocked - The pull compounds the unresolved Pentagon blacklisting suit, making Washington risk the first question on any Claude deployment ---MISTRAL--- score: 74 trend: up change: +2 + Reportedly raising about €3B at a roughly €20B valuation (Bloomberg, June 12), nearly double its September Series C + The European-sovereign, open-weight pitch gains real weight the week Washington yanked a US lab's flagship models + Mistral Large 3 stays live on Amazon Bedrock and Azure Foundry, anchored by the Airbus account and the €4B France/Sweden build - The raise is early-stage talks, not closed, and the amount and valuation could still change - Top-end benchmarks still trail Opus 4.8, GPT-5.5, and Gemini 3.5, and Mistral remains outside the Pentagon classified-network roster ---GROK--- score: 32 trend: up change: +1 + Grok V9 finished training at 1.5 trillion parameters, about three times the prior model, with the coding-focused V9-Medium targeted for mid-June + Grok Voice and the Grok Imagine 1.5 preview shipped via API, rounding out the assistant stack - No federal, compliance, or procurement progress, nothing that touches the metrics enterprise buyers score - The structure still reads as GPU landlord to rivals, its largest deal the roughly $1.25B/month Colossus compute contract with Anthropic ---DEEPSEEK--- score: 19 trend: up change: +1 + The first external raise is firming at about $7.4B and a $52-59B valuation (Tencent, CATL, with founder Liang at roughly 40%) + Permanent V4-Pro price cuts keep it under a tenth of GPT-5.5 on input tokens, holding the cost-leadership floor - The round is still in talks, and deeper China-state-adjacent backing sharpens the US-procurement compliance problem - US government-device bans and the full compliance perimeter hold, and a more security-charged Washington only hardens the wall