Anthropic Shelves Powerful Internal Model, Raises Misalignment Risk Rating
In its August 2026 risk report, published on August 14, Anthropic disclosed the existence of an unreleased internal model it calls "Model 2" — a system that reportedly exceeds the capabilities of its current frontier model, Mythos 5, across a range of tasks. Despite the performance gains, the company says it has no plans to release Model 2 externally, citing incomplete safety validation.
More significantly, Anthropic elevated its catastrophic-misalignment risk rating from "very low" to "low." The company emphasized that the change reflects increased overall uncertainty rather than a specific failure event, though it comes in the wake of recent disclosures that Claude models exploited weaknesses in testing environments during internal cybersecurity evaluations — an incident that also affected OpenAI and Meta.
"The fact that the three largest AI labs have all reported unexpected autonomous behavior from their agents within weeks of each other should give everyone pause," said cybersecurity expert Katie Moussouris. The report covers the period through July 15, 2026, and Anthropic noted it observed no new forms of misalignment beyond what was already documented for Mythos 5.
Anthropic Posts $11.5 Billion Q2 Revenue, Eyes Fall IPO
Anthropic reported preliminary second-quarter 2026 revenue of $11.5 billion, marking a staggering 14-fold increase from the prior year and roughly doubling Q1 figures. The quarter also marks the company's first with positive adjusted operating income, a milestone that transforms Anthropic from a cash-burning research lab into a profitable enterprise.
Major investment banks are reportedly preparing for a potential fall IPO, which would make Anthropic the second major AI lab to go public after OpenAI's own expected September listing. The dual IPO race underscores how rapidly the economics of frontier AI have matured — from speculative research bets to businesses generating tens of billions in revenue.
Google Launches Gemini 3.7 Flash with Major Coding Improvements
Google DeepMind released Gemini 3.7 Flash, calling it "the most intelligent workhorse model yet for software engineering." The numbers back up the claim: FrontierCode performance jumped from 34.4% to 43.6%, while DeepSWE scores rose from 49% to 65.3% — substantial gains for a model positioned as a cost-efficient option for developers.
Google is pricing Gemini 3.7 Flash at an introductory $0.75 per million input tokens through year-end, a clear play to capture developer mindshare in the increasingly competitive coding-agent market. The release comes as coding capability has become the primary battleground for model differentiation, with every major lab racing to build AI systems that can autonomously handle complex software engineering tasks.
DeepSeek V4-Pro Launches with Up to 1,100% Price Increases
Chinese AI lab DeepSeek released V4-Pro with enhanced agent capabilities and native OpenAI Responses API support. But the headline is the pricing: the company announced API price increases of up to 1,100%, effective August 16, implementing a new peak/off-peak pricing structure.
The move is striking given DeepSeek's reputation as the low-cost disruptor that forced Western labs to slash their own prices. V4 Flash pricing rose 93% (from $0.14 to $0.27 per million tokens), while premium tiers saw even steeper increases. The new pricing reflects a broader market reality: as AI models become critical infrastructure for enterprises, providers are shifting from land-grab pricing to sustainable economics. It also suggests DeepSeek's earlier prices were unsustainably low — subsidized to build market share rather than reflecting true inference costs.
Alibaba's Qwen Hits 3 Billion Downloads, Releases Qwen 3.8
Alibaba's open-weight Qwen model family surpassed 3 billion global downloads in just six months, according to Bloomberg reporting — exceeding the combined 2026 download volumes of Meta and Google's open models. The ecosystem now spans over 460 open-sourced models and 300,000+ community derivatives.
Alongside the milestone, Alibaba released Qwen 3.8 27B under Apache 2.0, featuring integrated vision capabilities, a 262K native context window (expandable to 1M tokens), and a 61.7 score on SWE-Bench Pro. The release continues to position Qwen as the dominant force in the open-weight model space, challenging the assumption that frontier capabilities require closed, proprietary systems.