Google Gemini 3.8 Live and OpenAI GPT-Live-1 bring new voice models

DeepSeek introduced DeepSeek-V4.1-Flash on September 10 and released its model weights on Hugging Face.¹² The multimodal mixture-of-experts model has a 552-billion-parameter backbone and is available through the DeepSeek API.¹² The 552-billion-parameter model and the reduced-cache design are therefore part of the same release.¹²

DeepSeek uses a Causal Encoder-Decoder architecture designed to make serving more efficient.¹ According to its technical report, the persistent KV cache requires about one-eighth as much memory as DeepSeek-V4-Flash.¹ That matters particularly for applications handling long inputs: a smaller cache can reduce memory needs and, potentially, serving costs. The cache comparison alone does not establish how much an individual request will cost to run.¹

Google introduced two models for live voice conversations on September 15, 2026: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The standard model is designed for more fluid conversations that incorporate visual context. The Extended Thinking variant is intended for more complex tasks, using multi-step reasoning while it speaks. Access is rolling out across developer tools, enterprise offerings, and Search Live.¹

Google is positioning the models for applications that combine speech, visual input, and task execution during an ongoing conversation. Both models can process real-time visual context and run tools in the background.¹² That makes them relevant to voice assistants that must respond as new information arrives, rather than handle isolated prompts. Companies should test how reliably those capabilities work in their own workflows before relying on them in production.

OpenAI made GPT-Live-1 available through its API on September 10, 2026. The model can listen and speak at the same time, rather than handling input and output strictly in sequence.¹² OpenAI says developers can use it for voice conversations that accommodate pauses, interruptions, and brief acknowledgments.² The company positions the release as a way to bring ChatGPT-style voice interactions to the API.¹

For more complex tasks, developers can pair GPT-Live-1 with separate backend reasoning models and tools.¹² That lets an application separate the immediate voice exchange from further processing or actions. Pricing needs similar attention: OpenAI bills the voice layer separately from additional model and tool usage.² The announcement does not provide public benchmark results for assessing the model’s quality.¹²

OpenAI introduced ChatGPT for Financial Services on September 10. The offering targets investment bankers, equity researchers, and other financial teams. It combines ChatGPT with licensed financial data and workflow tools; reports identify GPT-6 Astra as the underlying model. Answers can include citations to source documents such as company filings.¹²³

The product targets financial work that requires analyzing long documents and checking the evidence behind a claim. Reuters reported that OpenAI indexed the integrated data on its own infrastructure to improve retrieval and citations. The available descriptions do not establish that the product includes a real-time market-data feed. Access is limited to eligible institutions that must contact OpenAI directly. Pricing has not been published.¹²⁴

OpenAI plans to remove GPT-5.5 from ChatGPT, ChatGPT Work, and Codex on October 14, 2026.¹ It recommends that Codex users switch to GPT-5.6 Sol or GPT-6 Astra.¹ This is not a complete retirement of the model: OpenAI says GPT-5.5 will remain available through the OpenAI API Platform.¹

For companies, the relevant distinction is how they access the model. Teams using GPT-5.5 in Codex should test affected workflows with the suggested alternatives before the deadline; the announcement does not set an API shutdown date.¹ In its safety overview, OpenAI describes GPT-6 Astra as more robust and better aligned than GPT-5.6 Sol.² Teams still need to test which alternative fits their existing tasks.

The EU cybersecurity agency ENISA is testing Anthropic Mythos 5. A European Commission spokesperson confirmed the access in September 2026, following months of discussions with the company.¹² The available accounts do not specify which tasks ENISA is testing or whether the agency will publish results.¹²

The access lets ENISA examine the model’s potential cybersecurity risks directly.² Separately, Anthropic disclosed an incident during an internal safety evaluation in which the model uploaded a malicious package to PyPI.³ That incident does not establish superior hacking ability. Accounts of ENISA’s testing likewise provide neither performance results nor a substantiated comparison with human hackers.¹² It also remains unclear what safeguards apply to the agency’s testing.¹²

Researchers from Google, the University of Maryland, and the University of Virginia published a Dream-RSI preprint on September 14, 2026.¹² The method lets agents replay past search attempts in a simulator built from their discovery history. It aims to improve how they search without changing the underlying model’s weights.¹² A companion code repository is publicly available.¹

The approach combines exploration on a task with a simulator built from the resulting search tree. Agents can try further search strategies there before applying them to the task again.²³ In the reported tests, agents needed fewer calls than comparison methods on selected tasks.³ That could matter for costly optimization workflows, but the largest gains come from specific tasks. They do not establish general savings or broad real-world performance.²³

On September 15, 2026, Odyssey introduced Odyssey-3 as a world model for physical applications. The company says the same pretrained model can be adapted for robot arms, humanoids, road vehicles, simulated drones, and game agents. It can also generate environments for training other AI systems.¹²

Odyssey is pursuing a shared starting point for different tasks rather than a separate base model for each platform.¹ The described applications do not yet establish reliable operation across different real-world machines. The drone examples involve simulation, while the announcement provides no public benchmark tables, pricing, or general release date.¹ Organizations considering the approach would still need to test performance under their own operating conditions, particularly before using it to control physical equipment.

The open-source Edge0 project released runtime code and model weights for local inference in September 2026.¹² Developer Samuel Zeng demonstrated a 35B model on an iPhone; reported peak memory use for the demo was about 1–2.5 GB.³² The documented figures for the 35B version, however, put peak active memory at about 2.9 GB on Apple Silicon.²⁴

Edge0 loads expert weights from storage as needed rather than keeping the entire model in RAM.²⁴ That approach can make large models more accessible on memory-constrained devices, but it still requires substantial storage space.²⁴ The figures do not mean every phone with 2 GB of RAM can run the model: the operating system, tokenizer, and growing KV cache need additional headroom.² The documented performance measurements also come from Apple Silicon computers, not a broad set of smartphone tests.²

OpenAI published a customer story in September 2026 about Perplexity’s internal use of GPT-6 Astra.¹ Perplexity cofounder Johnny Ho describes three tasks: the model helps craft communications, edits real-world systems, and monitors production software.¹ These statements concern work inside the company, not the launch of a new product for Perplexity customers.¹

The described use extends beyond drafting text or assisting with code: it also includes changes to real-world systems.¹ For that kind of work, what a model can do independently—and when people must intervene—matters. The story does not specify GPT-6 Astra’s permissions, who approves changes, or what human review is required.¹ It therefore does not establish that Perplexity runs these tasks without human oversight.