AI Model Weekly Roundup — June 9–15, 2026

This was the week AI went from launch to lockdown in 72 hours. Anthropic shipped its most capable public model ever, only to see a US government export directive pull it offline before the weekend. Meanwhile, Mistral refreshed a workhorse MoE with a half-million-token context window, and the industry-wide price war showed no sign of cooling. Here is what moved the leaderboard conversation this week.

Fable 5's three-day reign

On June 9, Anthropic released Claude Fable 5, the first publicly available "Mythos-class" model positioned above the Opus line. Independent testers quickly reported roughly 80% on SWE-Bench Pro and strong showings on CursorBench and FrontierCode, with a 1 million-token context window and $10/$50 per-million-token pricing. For three days, Fable 5 looked like the new default for agentic coding and long-horizon knowledge work.

Then, at 5:21 p.m. ET on June 12, a US Department of Commerce export-control directive ordered Anthropic to suspend Fable 5 and its research-only sibling Mythos 5 for any foreign national, anywhere. Because Anthropic could not reliably distinguish foreign nationals in real time, it disabled both models globally. Claude Opus 4.8, Claude Sonnet 4.6, and Haiku remain available, but the episode marks the first time Washington has used export-control authority against a live, broadly distributed language model.

Mixtral 8x22B v2 goes long

On June 10, Mistral shipped Mixtral 8x22B v2, a 512K-context MoE refresh with a 40% price cut. The model is aimed squarely at long-document RAG, codebase-wide reasoning, and agentic workflows that need to keep hundreds of thousands of tokens in working memory. While it does not sit at the very top of the frontier tier, it undercuts most 1M-context competitors on price and gives Mistral a credible answer to DeepSeek V4 Flash and Grok 4.3 in the long-context value lane. It also reinforces Mistral's recent push into agentic coding alongside Mistral Medium 3.5 and Codestral 2508.

The price war accelerates

Pricing pressure kept building across the board. DeepSeek cut V4.1 API pricing by another 15%, keeping its reputation as the budget disruptor at the frontier. OpenAI rolled out 90% prompt-cache discounts for GPT-5.5 and GPT-5.5 Pro, making repeated long-context workloads dramatically cheaper for production users. Google, not to be left behind, pushed select Gemini tiers toward $0.10 per million input tokens in promotional tiers. The message from all three is the same: margin is being traded for volume, and the cost of intelligent inference keeps falling.

Score changes

No LMRank scores were updated this week. The leaderboard remains unchanged while the team waits for fresh independent benchmarks on Fable 5, Mixtral 8x22B v2, and the next wave of reasoning models.

\n

\n At lmrank.com, we track live scores, pricing, and capabilities for frontier AI models. Check the leaderboard for the latest rankings, or read last week's roundup for the run-up to Fable 5.