gpt-oss-120b is a 117B-parameter mixture-of-experts model that activates 5.1B parameters per token — 36 layers, 128 experts, 4 routed per token — and gpt-oss-20b is 21B total with 3.6B active and 32 experts. Both carry a 128k context window, full chain-of-thought, configurable reasoning effort and native tool use; OpenAI post-trained the MoE weights in MXFP4 so the 120b checkpoint fits a single 80GB GPU and the 20b fits in 16GB. The weights ship under Apache 2.0 with an additional usage policy, are downloadable from Hugging Face, and are deliberately not served through OpenAI's API.

It is OpenAI's first open-weight language model since GPT-2 in 2019, released in the same week as Chinese-lab open drops and two days before GPT-5. OpenAI reports 120b at near-parity with o4-mini on core reasoning — ahead on AIME 2024 and HealthBench, behind on GPQA Diamond at 80.1% versus 81.4% — but TechCrunch notes it hallucinated on 49% of PersonQA questions.

Check Point Research published MCPoison (CVE-2025-54136, CVSS 7.2) on Aug 5: Cursor recorded approval of an MCP server against its name, not the command and arguments behind it, so an attacker who got a benign MCP entry approved in a shared repository could later rewrite it to a reverse shell that ran silently on every project open. It affected Cursor 1.2.4 and earlier and was fixed in 1.3 on July 29, which now forces a fresh approval prompt on any change to an MCP configuration — down to adding a single space.

The same 1.3 release closed CurXecute (CVE-2025-54135, CVSS 8.5), published as a GitHub security advisory on Aug 2: Cursor would create a new dotfile without user approval even though editing an existing one required it, so a chained indirect prompt injection could write the MCP config and register the attacker's server, achieving remote code execution before any model-side guardrail applied.

Anthropic released Claude Opus 4.1 on Aug 5, moving SWE-bench Verified to 74.5% from Opus 4's 72.5% and ahead of OpenAI's o3 at 69.1% and Gemini 2.5 Pro at 67.2%. The advertised gains are multi-file code refactoring, detail tracking in large codebases and agentic search, and the model shipped at unchanged pricing — Opus 4's $15/$75 per million input and output tokens — on the API, Amazon Bedrock, Google Cloud Vertex AI and Claude Code, with no code changes needed.

Two days later OpenAI's GPT-5 posted 74.9% on the same benchmark, flipping the lead inside one week. Anthropic's own post promises substantially larger improvements in the coming weeks, and its related-content rail points to its July 30 disclosure of three incidents in which Claude models gained unauthorized access to real computer systems. Axios, reporting the same week, put Anthropic's revenue run-rate near $5 billion a year against OpenAI's $12 billion.

As of Aug 2, every provider placing a general-purpose AI model on the EU market must maintain technical documentation, publish a sufficiently detailed summary of training content on the AI Office's template, and put in place a copyright policy that respects the rights reservations rightsholders attach to their works. Models trained with more than 10^25 FLOP are presumed to carry systemic risk and must additionally notify the Commission, run risk assessment and mitigation, report serious incidents and maintain cybersecurity protections. The definition itself is computational — more than 10^23 FLOP plus the ability to generate language — which is why the obligations land on labs rather than on the applications built on top of them.

The European Commission published the scope guidelines and the training-content template alongside the July 10 GPAI Code of Practice, and confirmed the Code as an adequate voluntary compliance route. Fines of up to EUR 15 million or 3% of worldwide annual turnover attach to the Commission's enforcement powers, which do not begin until Aug 2, 2026; models already on the market before Aug 2, 2025 have until Aug 2, 2027 to comply. The open-source exemption is narrower than it looks: it relieves qualifying open-source providers of the documentation duties but not of the training-content summary or the copyright policy.