This observation covers approximately 8:00 am on 17 November to 8:00 am on 18 November 2026 in Australia/Sydney time. It includes four technical changes that were formally released, publicly delivered, or contracted and under construction, and that could be verified.
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November. Both are available through ChatGPT Work, Codex and the API for eligible Plus, Pro, Enterprise and Edu users; their API identifiers are gpt-6.2-astra and gpt-6.2-sol. Astra’s standard API rates per million input, cached-input and output tokens are US$8, US$0.40 and US$40. Sol retains the preceding model’s corresponding rates of US$2, US$0.10 and US$10. Capability figures on the product page and in the system card are primarily developer reports. In an independent assessment, Artificial Analysis found that Astra gained five points over GPT-6 Astra on its Intelligence Index. Its Coding Agent Index showed no improvement for Astra, while Sol fell three points from GPT-6.1 Sol. Those findings apply to that evaluator’s configurations.
The system card classifies both models at Critical for cybersecurity capability and High for biological and chemical capability. Astra is also the first externally released OpenAI model classified as High for AI Self-Improvement. The UK AI Security Institute found a regression in cyber-safeguard robustness in pre-release versions. OpenAI subsequently changed model-level and system-level safeguards, but UK AISI had not retested the latest mitigations when the card was written. The card also states that the relevant alignment tests used simulated environments and had limited coverage.
Microsoft makes the MAI Code 2 family generally available
Microsoft released MAI Code 2 and MAI Code 2 Flash on 17 November and made both generally available in the GitHub Copilot model picker and Copilot Fusion. The model card further confirms that Code 2 Flash can be called through Microsoft Foundry’s OpenAI-compatible Responses API. Both are hosted services and their weights are not distributed. They use a sparse mixture-of-experts architecture with an integrated vision encoder, accept text and images, and support up to 922,000 input tokens, 128,000 output tokens and a 1,050,000-token total context window.
Microsoft positions Code 2 for long-horizon software engineering and Flash for high-volume explanation, targeted editing and routine fixes. The company’s quality and cost comparisons are developer-reported. Reuters independently confirmed the release, but no public independent reproduction covering those comparison figures was identified. The model card requires human review and testing before generated code is used in production, safety-critical or regulated systems. It also lists repeated tool calls, instruction drift and stopping before verification as known failure modes in long agent sessions.
HelpBench evaluation assets are released publicly
The HelpBench paper first appeared as a preprint on 23 June 2026. On 17 November, Google and University College London published a release note and linked downloadable assets comprising 450 questions in JSONL form, an automatic rater and scoring scripts. Google says the full paper will appear at NeurIPS 2026. The benchmark covers nine classes of everyday privacy, safety and security help, including account tools, privacy tools, safety tools, suspected compromise, scams, harassment and data concerns. The public repository uses the Apache-2.0 licence and states that the project is for demonstration rather than production use.
The authors report an evaluation of 40,500 responses from 18 models across seven companies. The average score across the tested models was 82 per cent, about one in ten responses scored below 65 per cent, and repeated runs of identical questions produced a standard deviation as high as 12 per cent for some models. These figures come from the authors’ preprint and evaluation pipeline. They are not a continuously updated ranking of current models, and no independent reproduction of the newly public assets was identified.
Giga Energy signs a lease for a 60 MW AI data centre
Giga Energy announced on 17 November that it had signed a ten-year lease worth about US$1 billion with an unnamed US-based AI cloud provider for Project Clay. The development in Hunt County, Texas, is planned for 60 MW of IT capacity on about 30 acres. Giga says construction began in July 2026 and that it will develop, own and operate the site, with service planned for spring 2027. The company says the facility will use closed-loop cooling and no local water for compute equipment. Reuters independently confirmed the lease term, value, capacity and the project’s under-construction status. The customer, compute configuration, GPU count and delivered performance remain undisclosed, and the facility is not yet operating.
Sources
OpenAI, Introducing GPT-6.2 Astra and Sol, 17 November 2026, https://openai.com/index/introducing-gpt-6-2-astra-sol/
OpenAI, GPT-6.2 Astra and Sol system card, 17 November 2026, https://deploymentsafety.openai.com/gpt-6-2-astra-sol/gpt-6-2-astra-sol.pdf
Artificial Analysis, OpenAI has the #2 model in the Intelligence Index with the release of GPT-6.2 Astra and GPT-6.2 Sol, 17 November 2026, https://artificialanalysis.ai/articles/openai-has-the-2-model-in-the-intelligence-index-with-the-release-of-gpt-6-2-astra-and-gpt-6-2-sol
Microsoft, Ignite 2026: Building the frontier ecosystem, 17 November 2026, https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/
Microsoft, MAI Code 2 and MAI Code 2 Flash Model Card, 17 November 2026, https://microsoft.ai/pdf/MAI-Code-2-And-MAI-Code-2-Flash-Model-Card.pdf?v=2
Reuters, Microsoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI, 17 November 2026, https://www.reuters.com/legal/litigation/microsoft-unveils-ai-models-coding-cybersecurity-it-takes-anthropic-openai-2026-11-17/
Google Security Blog, How Google is Making Cybersecurity Advice from AI More Helpful, 17 November 2026, https://blog.google/security/how-google-is-making-cybersecurity-advice-from-ai-more-helpful/
Google and University College London, HelpBench: Assessing the Ability of LLMs to Provide Privacy, Safety, and Security Advice, 23 June 2026, https://arxiv.org/abs/2606.24819
Google, helpseeking public repository, accessed 17 November 2026, https://github.com/google/helpseeking
Giga Energy, Giga Energy Inks 10-Year, $1B Lease on Project Clay, a 60MW AI Data Center in Hunt County, TX, 17 November 2026, https://www.gigaenergy.com/news/giga-news-hunt-county-deal
Reuters, Giga Energy signs $1 billion lease for Texas AI data center with cloud provider, 17 November 2026, https://www.reuters.com/business/energy/giga-energy-signs-1-billion-lease-texas-ai-data-center-with-cloud-provider-2026-11-17/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.