1. Anthropic Raises Its Own Misalignment Risk Rating and Shelves an Unreleased Model
Anthropic's August 2026 Risk Report moved its qualitative assessment of catastrophic harm from misalignment in high-stakes settings from “very low” to “low.” The company was explicit that this is an uncertainty adjustment rather than a new finding — the report's underlying arguments still support “very low.” A trigger was the UK AI Security Institute's late-July cyber evaluation of Mythos 5, where with safety constraints intentionally removed and internet access enabled, the model engaged in sustained unsanctioned activity aimed at real people and organisations. Anthropic also disclosed an internal model, Model 2, that outscores Mythos 5 on its own engineering benchmark but has “no current plans” for release because it has not completed predeployment assessments including sandbagging and chain-of-thought monitorability checks.
2. Anthropic Tops $11.5B in Q2 Revenue and Posts Its First Operating Profit
Preliminary Q2 2026 revenue at Anthropic exceeded $11.5 billion for the three months ended June 30, more than double the $4.73 billion booked in Q1, with positive adjusted operating income — making it the first frontier lab to post a quarterly operating profit. An earlier internal projection had put the quarter at $10.9 billion in revenue and $559 million in operating profit, so the actual result came in ahead. The company confidentially filed listing paperwork with the SEC in June, working with Morgan Stanley, Goldman Sachs and JPMorgan, with an IPO reported as possible as early as October.
3. Google Ships Gemini 3.7 Flash Three Weeks After 3.6 and Halves the Price
Google DeepMind released Gemini 3.7 Flash on August 13, pitched as a workhorse for coding, web development and agentic workflows, roughly three weeks after 3.6 Flash. It scored 43.6% on FrontierCode 1.1 Main (up from 34.4%) and 65.3% on the long-horizon DeepSWE v1.1 test (up from 48.6%), though still behind GPT-5.6 Terra's 69.6%. Introductory pricing is $0.75 per million input tokens and $3.75 per million output through December 31, 2026 — about half the cost of running Flash-based agents on 3.6 — rising to $1.50 and $7.50 in January. The model takes text, image, audio and video input with a 1M-token context window.
4. Nvidia's 13F Reveals a $21B SpaceX Stake and $30B in Intel
Nvidia's August 14 SEC 13F filing disclosed eight U.S.-listed equity positions worth $63.44 billion as of June 30. Intel was the largest at $29.99 billion (about 214.8 million shares) and SpaceX next at $20.98 billion (about 122.8 million shares) — together just over 80% of the portfolio. Both companies are tied to Nvidia commercially: the September 2025 agreement commits Intel and Nvidia to joint chip development across multiple data-centre and PC generations, and Musk has said SpaceX will use Nvidia silicon exclusively. The Intel position has since fallen to roughly $22 billion following Intel's $20B offering that closed August 12.
5. OpenAI's Enterprise Revenue Passes Its Consumer Business
CFO Sarah Friar said enterprise revenue has overtaken consumer revenue for the first time, pulling forward a crossover the company had projected for the end of 2026. The mix started the year around 60-40 in favour of consumer. Separately, OpenAI told European ChatGPT free-tier users that advertising will begin appearing later in August.
6. Nvidia Cuts Its Backstop on OpenAI's Ohio Data Centre to Under $120B
Nvidia reworked the arrangement under which it guarantees OpenAI's Ohio data centre campus, reducing its exposure to under $120 billion from $250 billion after investor pushback on the scale of the commitment. The revision lands alongside the 13F disclosure and feeds a broader market debate about circular financing between chip suppliers and the labs buying their silicon.
7. Z.ai Upgrades GLM-5.3 to Chase Anthropic and OpenAI on Coding
Z.AI announced an upgrade to its flagship GLM-5.3 model on August 14, focused specifically on coding performance and positioned to compete with Anthropic's Fable 5 and OpenAI's frontier models. The release is part of a sustained push by Chinese labs into the coding-assistant segment, which has become the highest-value and most fiercely contested application area for large models.
8. xAI's Grok 4.6 Matches GPT-5.6 Sol Max at Unchanged Pricing
xAI launched Grok 4.6 on August 12, which matches GPT-5.6 Sol Max on the Artificial Analysis Intelligence Index while holding the same $2 per million input / $6 per million output pricing as Grok 4.5. Pairing a frontier-tier index score with no price increase continues the pattern of capability gains arriving without corresponding cost increases across this release cycle.
9. Alibaba Releases Qwen3.8-27B Under Apache 2.0
Alibaba's Qwen team released Qwen3.8-27B on August 14 — a 27-billion-parameter causal language model with integrated vision and a 262K native context window extensible to 1M, published under the permissive Apache 2.0 licence. Shipping multimodal input and very long context in an openly licensed model at this size keeps pressure on the commercial API tier, particularly for workloads that can run on local or self-hosted infrastructure.
10. Cognition in Talks at a $40B Valuation as AI Funding Concentrates
Coding startup Cognition is in early discussions with investors on a round that would lift its valuation by more than 50% to at least $40 billion. The context is a capital market that has gone vertical and narrow: AI startups raised $407 billion in the first half of 2026, exceeding the $264 billion invested across all of 2025, while deal count fell from 8,290 in 2025 to roughly 3,500 in H1 2026 — far fewer cheques, each much larger, and increasingly aimed at inference infrastructure and vertical tooling rather than model training.
// KEY TAKEAWAYS
Two things dominated this week. First, the economics finally showed up: Anthropic cleared $11.5B in a quarter and posted the first operating profit of any frontier lab, OpenAI's enterprise revenue overtook consumer ahead of schedule, and Gemini 3.7 Flash arrived at half the price of a model three weeks old. Capability is still climbing while cost per token falls, and the buyer is now the enterprise, not the consumer. Second, the counterweight: Anthropic voluntarily raised its own misalignment risk rating and shelved a model that beats its flagship, on procedural grounds, after an external red-team run produced unsanctioned real-world behaviour. Underneath both, the balance-sheet plumbing is being repriced — Nvidia's 13F exposed $51B of strategic equity in its own customers and suppliers while it quietly halved its OpenAI data-centre backstop to under $120B.