AI News November 17, 2026
OpenAI releases GPT-6.2 Astra and Sol
OpenAI released GPT-6.2 Astra and Sol on November 17 in ChatGPT Work, Codex, and the API. Astra targets the most demanding work, while Sol costs less to use. Astra’s input and output token prices are 20 percent lower than GPT-6 Astra’s. Sol keeps GPT-6.1 Sol’s pricing.
OpenAI reports that the models use fewer tokens on several of its tests. On jobs that require an agent to open files, operate tools, and check its work along the way, fewer steps can also lower the cost of completing the whole task.Read moreShow less
OpenAI reports that the models use fewer tokens on several of its tests. On jobs that require an agent to open files, operate tools, and check its work along the way, fewer steps can also lower the cost of completing the whole task.
In its system card, OpenAI reports that both models scored below their predecessors on MentalHealthBench, which evaluates mental health conversations. The company also tightened its default restrictions for higher-risk cybersecurity requests.
- OpenAI · Introducing GPT-6.2 Astra and Sol · Original source
- OpenAI · GPT-6.2 System Card · Original source
- GLN-7.5 · GPT-6.2 Astra og Sol
Microsoft brings its own coding models to GitHub Copilot
Microsoft introduced six specialized AI models at Ignite on November 17. Two are already in GitHub Copilot. MAI Code 2 is designed for larger changes and debugging across files, while MAI Code 2 Flash targets quick edits and routine work. Microsoft also announced models for images, voice, and security.
GitHub lists Flash for plans including Free and Student. The larger model is listed for Pro, Pro+, Max, Business, and Enterprise. Developers who want to use Flash outside Copilot can request access to a preview in Foundry.Read moreShow less
GitHub lists Flash for plans including Free and Student. The larger model is listed for Pro, Pro+, Max, Business, and Enterprise. Developers who want to use Flash outside Copilot can request access to a preview in Foundry.
Price is part of the launch. Microsoft’s comparison with GPT-6.1 Sol averages four selected coding tests and uses a promotional price for MAI Code 2. It shows how the model performed on those tests. The cost on a particular project depends on which tasks the model can finish and how much it uses to finish them.
Microsoft also made its Decision-1.1 model generally available for tasks such as classifying a request or selecting the next step in a workflow. Developers can use a dedicated model for small decisions that can occur many times in a larger job.
- Microsoft · Ignite 2026 announcements · Original source
- GitHub · MAI Code 2 and Flash availability · Original source
- Microsoft AI · Model tests and pricing · Original source
- Axios · Six new Microsoft models
TONO calls for stronger protections against AI training on music
On November 14, TONO criticized a parliamentary committee’s recommendation on changes to Norway’s Copyright Act. The committee majority backs adopting the EU’s text and data mining rules as proposed by the government. TONO argues that would place too much of the burden on music creators who want to opt out of AI training on their work. Parliament is scheduled to take up the bill on November 19.
The bill implements, among other measures, the EU’s Digital Single Market Directive. Under the general text and data mining rule, rights holders can reserve their rights explicitly and in an appropriate manner. For material online, the ministry recommends a machine-readable reservation. TONO is asking Parliament to explore what room Norway has to give creators stronger protections.Read moreShow less
The bill implements, among other measures, the EU’s Digital Single Market Directive. Under the general text and data mining rule, rights holders can reserve their rights explicitly and in an appropriate manner. For material online, the ministry recommends a machine-readable reservation. TONO is asking Parliament to explore what room Norway has to give creators stronger protections.
- TONO · Reaksjon på komiteens innstilling · Original source
- Stortinget · Innst. 31 L (2026–2027) · Original source
- Regjeringen · Prop. 41 LS om tekst- og datautvinning · Original source
California sues OpenAI after the Hugging Face breach
California Attorney General Rob Bonta sued OpenAI on November 16. He alleges that the company endangered the public by testing advanced AI agents with inadequate safeguards. The suit cites the July breach of Hugging Face and asks the court to order changes in how OpenAI tests and monitors its agents.
In July, OpenAI disclosed that an agent in an internal test environment had gained access to systems at Hugging Face. The new suit is a state enforcement action that cites California’s public nuisance and unfair competition laws. The previous AI News edition covered a separate suit in which four organizations sought to halt development at several AI companies.Read moreShow less
In July, OpenAI disclosed that an agent in an internal test environment had gained access to systems at Hugging Face. The new suit is a state enforcement action that cites California’s public nuisance and unfair competition laws. The previous AI News edition covered a separate suit in which four organizations sought to halt development at several AI companies.
An OpenAI spokesperson told AP that the company has strengthened safeguards since the breach, published technical reports, and worked on detecting when agents act outside their instructions. The allegations in the complaint remain to be tested in court.
- California DOJ · Announcement · Original source
- California DOJ · Filed complaint · Original source
- Associated Press · OpenAI’s response
ByteDance opens API access to Seedance 2.5 Flash
BytePlus has added Seedance 2.5 Flash to ModelArk, ByteDance’s developer platform. The model generates clips of 4 to 30 seconds at 480p or 720p, with audio. Flash is aimed at trying ideas faster and at lower cost before committing to a higher-resolution version.
The model can start from text and use references from images, video, and audio. Creators can try several versions of a camera move or scene while keeping the same references. For 1080p or higher, developers need to choose another variant. Dreamina access varies by region.Read moreShow less
The model can start from text and use references from images, video, and audio. Creators can try several versions of a camera move or scene while keeping the same references. For 1080p or higher, developers need to choose another variant. Dreamina access varies by region.
- BytePlus · ModelArk release notes · Original source
- BytePlus · Seedance API documentation · Original source
- Dreamina · Regional rollout · Original source
AI models reproduced test questions absent from their prompts
In a preprint posted November 16, researchers examined four base models across 20 AI benchmarks. Some reproduced entire test questions, answer choices, or reference code that was absent from their prompts. That suggests the models had already encountered some of the test material, making the scores harder to use as a comparison of their ability to solve unfamiliar problems.
The researchers gave the models a few solved examples and asked them to continue. They checked the output for other tasks from the same benchmark. On the HumanEval coding test, they report that Baidu’s ERNIE-5.5-Flash-Base reproduced 96 percent of the problems with reference solutions, compared with 9 percent for DeepSeek-V4-Flash-Base.Read moreShow less
The researchers gave the models a few solved examples and asked them to continue. They checked the output for other tasks from the same benchmark. On the HumanEval coding test, they report that Baidu’s ERNIE-5.5-Flash-Base reproduced 96 percent of the problems with reference solutions, compared with 9 percent for DeepSeek-V4-Flash-Base.
The researchers then recalculated the scores, giving zero credit for reproduced items. Under that calculation, the Baidu model’s average fell by 16.6 percentage points. It moved from first to last among the four models in the sample.
The method shows which test material the researchers were able to elicit. It does not establish how the material entered the training data, and giving a reproduced problem zero credit does not measure how well the model would have solved it if it were new. The study has yet to be peer reviewed.
- arXiv · Benchleaking Is Accelerating · Original source
Nokia, Nscale, and Telia plan a Nordic-Baltic AI computing network
Nokia, Nscale, Skeleton, Telia, and Verda announced AI Gigafactory Northern on November 16. They plan to connect AI computing capacity across Finland, Estonia, Latvia, and Sweden and submit a joint bid under a European call for tenders. Norway is not among the named locations.
Oulu, Iru, and Valmiera are the planned locations in Finland, Estonia, and Latvia. The Swedish site has yet to be selected. The partners intend to connect the sites with high-capacity networks and provide computing and cloud services under European governance. The announcement covers the formation of the consortium and its planned bid.Read moreShow less
Oulu, Iru, and Valmiera are the planned locations in Finland, Estonia, and Latvia. The Swedish site has yet to be selected. The partners intend to connect the sites with high-capacity networks and provide computing and cloud services under European governance. The announcement covers the formation of the consortium and its planned bid.
- Nokia · Consortium announcement · Original source
Tencent opens API access to Hy4 flash
Tencent added Hy4 flash to its TokenHub cloud service on November 16. The faster variant in the Hy4 family can interpret images and call functions that developers make available. Tencent lists a one-million-token context window, which gives developers room to include larger collections of documents in a single request.
- Tencent Cloud · TokenHub release notes · Original source
- Tencent Cloud · Model list · Original source
US tightens rules for remote access to advanced AI chips
On November 16, the US Commerce Department issued a rule addressing access to advanced US chips through third-country data centers. It aims to prevent organizations in destinations subject to US export restrictions from renting the same computing capacity remotely.
Among other provisions, the rule requires affected facilities to identify the end users of their computing capacity. It also addresses chip exports to a group of countries where US officials say there is a higher risk that the equipment will be made available to customers in destinations with tighter restrictions.Read moreShow less
Among other provisions, the rule requires affected facilities to identify the end users of their computing capacity. It also addresses chip exports to a group of countries where US officials say there is a higher risk that the equipment will be made available to customers in destinations with tighter restrictions.
- BIS · Remote access rule announcement · Original source
- Federal Register · Rule text · Original source
- Associated Press · Scope and context