OpenAI gives away the engine behind Codex
OpenAI has opened up the engine behind its Codex coding agent, publishing the app-server and the official Codex SDK as Apache-2.0 open-source components on 19 August. Together with the already-open codex exec command-line tool, they make up the “harness” — the runtime that drives the agent loop, reading a task, keeping context across a long session, streaming events, calling tools, and pausing for human approval — and the permissive licence lets developers modify, embed and commercialise it rather than being tied to a chat interface.
The interesting number is what the wrapper is worth. OpenAI says that two harness settings alone lifted GPT-5.6 Sol’s score on the ARC-AGI-3 benchmark from 13.3% to 38.3% while cutting token use roughly sixfold — evidence that how you wrap a model can matter as much as which model you pick. For a regulated buyer, an Apache-2.0 runtime that speaks to any OpenAI-compatible endpoint is one you can point at a sovereign or self-hosted model without rebuilding your tooling. The agent scaffolding keeps commoditising; the choice of where the tokens actually run stays yours.
Nvidia takes a stake in the company that lines up power for data centres
Nvidia is moving further up the supply chain. On 21 August it took a minority stake in Cloverleaf Infrastructure, a developer that assembles electricity and land for AI data centres and acts as a middleman between utilities and operators. Terms were not disclosed, though the Wall Street Journal reported the investment could run to several hundred million dollars. The projects will follow DSX, Nvidia’s reference design for racks, networking and storage, and Cloverleaf has already sold more than 7GW of capacity to developers, including sites in Wisconsin for Oracle and OpenAI.
The pattern is worth watching: the same company that sells the chips now also helps finance the power and land that create demand for them. For anyone weighing where regulated European workloads should run, it is another reminder of how concentrated the AI supply chain has become — chips, reference design and, increasingly, the capital behind the buildout all trace back to one vendor — and that the map of available sites is being drawn around megawatts, not cities.
The frontier models are bunched — so the race is now price and speed
August’s wave of releases has left the top of Artificial Analysis’s Intelligence Index tightly packed: Claude Opus 5 leads, with Claude Fable 5 and Grok 4.6 just behind and roughly half a dozen models — including Moonshot’s Kimi K3 and Z.ai’s GLM-5.3 — within a few points. When the best models are that close on raw capability, the thing that actually separates them is what you pay and how fast the tokens arrive. Google’s Gemini 3.7 Flash, launched 13 August, is the clearest example — Artificial Analysis clocks it among the fastest models it tracks, at about 340 tokens per second (373 on Google’s own API), and at half its predecessor’s price, though that is introductory pricing set to double at the start of 2027.
That is the same story last week’s hardware news told from the other end: Cerebras claiming 30x faster inference, and OpenAI’s own harness work squeezing more out of a fixed model. For a European buyer the practical takeaway is to stop shopping on leaderboard rank. When the top models are effectively interchangeable on quality, the decision that matters is price, latency and where the request is served — which is a routing question, not a model-picking one.
Quick Hits
- Anthropic posts the first profitable quarter in frontier AI. Preliminary Q2 revenue topped $11.5 billion — up more than 14-fold year over year and ahead of Q1’s $4.73 billion — with the company’s first positive adjusted operating income, per CNBC; the figures are preliminary and under review.
- Anthropic’s Amodei calls the AI backlash “a crisis of trust.” In comments reported 16 August, the CEO said public suspicion of AI is really distrust of companies and governments, and conceded the sharpest fair criticism is that the labs “haven’t yet delivered on our big promises to benefit the world.”
- Salesforce says enterprise agent counts are climbing steadily. Its Agentic Enterprise Index finds the average business on its platform went from five activated agents in early 2025 to 13 by April 2026, and its AI agents now resolve roughly seven in ten of the customer-service sessions they handle without a human.