GPT-6 adds interactive responses as Mistral Large 4 enters preview

OpenAI introduced GPT-6 with “Intelligent UI” in ChatGPT on October 7. Responses can combine text with graphics, charts, buttons, forms, and interactive tools. The model chooses a format based on the request and can start displaying an answer while it continues reasoning or using tools.¹² The global rollout began for Plus, Pro, Business, and Enterprise users, with Free and Go users scheduled to follow starting October 8.²¹

The feature lets users handle some tasks within a response, using tools such as calculators and bill splitters.¹³ It is available in the Chat tab, not the Work tab, and Enterprise access may depend on workspace admin settings.³² The Pro reasoning option uses GPT-6 Astra and does not support Intelligent UI. Access to Astra is initially limited to selected organizations.³²

Mistral AI introduced Mistral Large 4 as a public preview on October 6.¹ The mixture-of-experts model has one trillion parameters, accepts text and images, and produces text.¹ It has been available through Mistral Studio and the API since the announcement; the company plans to publish its weights by the end of October.¹

Mistral is positioning the model as a European alternative that companies could run on their own infrastructure once the weights are available.²³ Meanwhile, the company is continuing reinforcement learning, fine-tuning, and testing with development and cybersecurity specialists.¹⁴ Verified benchmark results are still pending, while definitive license terms and confirmed API pricing have not been announced.¹³

Reflection AI introduced Beam on October 5 as a text-only model with planned open weights for coding, reasoning, and agent tasks.¹² The mixture-of-experts model has 501 billion total parameters, according to the company, but its weights were not publicly available during the reporting week.³² An early version was available to selected users, with a public weight release planned for later in October.³

Reflection positions Beam as an alternative to large open models such as GLM-5.2 and Qwen.¹³ The company claims Beam needs one-third to one-quarter as much hardware as GLM-5.2 on some tasks.³ That is a comparison for particular workloads, not a general efficiency figure. Public weights and independent testing will be needed to assess whether the claimed advantage holds in production.

Aleph Alpha released Kolibri-1 on October 3, 2026. The weights for the German-English text model are available on Hugging Face under the Apache 2.0 license.¹² Kolibri-1 uses a sparse mixture-of-experts architecture and targets organizations that want to run a model on their own infrastructure. The release focuses on downloadable weights rather than a hosted API service.²

Aleph Alpha advertises a context window of up to 1,048,576 tokens. Technical information, however, identifies 262,144 tokens as the native context and recommends that shorter length for more efficient serving.² Deployment also takes preparation: Kolibri-1 requires a dedicated vLLM plugin.² The published benchmark results come from the vendor and had not been independently verified at launch. Organizations considering the model should test performance and infrastructure needs against their own workloads.²

On October 6, Anthropic expanded its Cyber Verification Program to incorporate Project Glasswing. Vetted security organizations can access Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 under different verification and safety requirements.¹ Glasswing partners reported at least 129,000 verified software vulnerabilities found between April and July 2026 while using Claude Mythos models.¹

The figure illustrates the potential scale of AI-assisted security reviews, but it cannot be attributed to Mythos 5.1 alone. Anthropic provides neither a model-by-model breakdown nor a complete dataset of the findings.¹ Finding a vulnerability also does not mean it has been fixed. Access to Mythos 5.1 remains limited to vetted cyber defenders and life scientists.²

Google Cloud introduced Gemini agent on October 8. Users are meant to give it an objective rather than describe every step; the agent then plans and carries out the task through an interface or API.¹ Google says it can answer questions, generate media, write and run code, and create documents and presentations.¹ The product is initially available only to selected customers in private preview.²

The agent can use tools and internal systems and delegate parts of a task to subagents.¹ Google is positioning it to handle workflows across applications rather than respond to isolated requests.¹ Its reliability on complex tasks has not yet been independently established. Google has not given a general availability date.²

DeepSeek and Huawei released open-source software for Huawei’s Ascend accelerators on September 30. It includes an Ascend version of DeepGEMM, a matrix multiplication library, alongside TileLang and other compute and communication libraries.¹² DeepSeek describes TileLang as a simpler alternative to CUDA for programming AI-chip operators.¹ The release concerns software for Ascend, not a change to API prices.¹²

A reported figure of about 98% for Ascend 950 refers to performance relative to the theoretical FP8 peak in a specific MegaMoE operation, not savings on token generation.³ It does not show whether the software makes inference cheaper overall: memory access, communication between chips, and model-serving overhead also affect costs.³ For companies using Ascend hardware, the released code provides additional optimization tools, but any savings need to be measured under their own operating conditions.¹³

P-Zero Research introduced TasteVal on October 6. The benchmark tests how AI models repeatedly design experiments and interpret results for fixed research problems. In the accompanying study, Opus 5.5 reached the best human experts’ target score after about 17 GPU-hours of experiments, compared with roughly 40 GPU-hours for the experts. The authors put its compute efficiency at 2.30 times the expert baseline.¹²

For the strongest models, the researchers estimate that this efficiency has doubled every three months since December 2025. That is a preprint estimate, not an independently confirmed trend. TasteVal also measures work on fixed problems, not the ability to identify new research questions. Most of the human experts had not worked at a frontier AI lab. On some tasks, hiding the scoring function substantially reduced the measured efficiency of earlier models.³²