New models. Useful releases. The technology moving things forward.
DeepSeek has updated V4.1 Flash to its 1017 version across the API, website, and app, while opening Search and Fetch APIs for multilingual web retrieval. Existing model integrations retain the deepseek-flash identifier, calling method, and pricing. The retrieval APIs require a DeepSeek API key and each have an account limit of five requests per second, constraining concurrent application workloads.
OpenAI has released prompt-to-voice for all API customers, allowing developers to create reusable voices from text descriptions through the API or voice Playground. Voice creation is free and speech uses regular GPT-Live session billing; prompt-generated voices are unsupported in the Text to Speech, Realtime, and Chat Completions APIs. The update also adds nine natural voices covering German, French, Hindi, Japanese, and Korean.
Anthropic reports that Claude Mythos 5 produced improved quantum circuit designs across eight tasks in chemistry, optimization, simulation, and number theory. Researchers can examine the published method, which checks proposed code and scores its qubit and gate requirements through automated evaluators. The experiments included occasional human guidance, and practical applications still depend on suitable quantum hardware.
Reflection has released Beam’s weights under Apache 2.0, moving beyond its earlier select-user preview to downloadable deployment and customization. Developers can self-host the 501-billion-parameter model or access hosted services, with BF16, FP8, and NVFP4 formats available. Reflection’s own API remains in public beta, so its access status differs from the downloadable weights.
Claude Code 2.1.302 enables fork subagents by default in the desktop app, VS Code, Agent SDK, and noninteractive sessions, matching terminal behavior. Developers also gain an idle-compaction setting and a model selector for the client-side auto-mode classifier. Fork subagents can be disabled with an environment setting, and the classifier selector does not change server-side auto mode.
AWS Transform now generally supports converting Azure SDK code in .NET applications to AWS equivalents, including Functions, Blob Storage, and Queue Storage. Teams can review and adjust conversion options before running the modernization agent, helping automate service integration changes during migration. AWS Transform is available in eight regions across North America, Europe, and Asia Pacific; Cosmos DB conversion to DynamoDB remains in preview.
NVIDIA has released Dory 1.0 4B, an English reasoning model with downloadable weights and support for up to 256K tokens of context. Developers can adjust how often its shared layers run, trading computation for accuracy without retraining. The documented vLLM and llama.cpp integrations currently require Dory-specific builds, adding setup work for teams deploying it locally.
OpenAI is rolling out Slides to eligible Pro, Business, and Enterprise users with Space access on web and desktop, enabling collaborative editing and presentation exports. GPT-6 Pro also gains parallel agents, faster responses, and improved plugin and tool use for eligible Pro users. Managed workspace access depends on settings and rollout eligibility, while Enterprise sharing in Space remains an opt-in preview.
Anthropic says an internal Claude research model traced several standardized elliptic-curve seeds to a common pseudorandom generator initialized with timestamps. The publication includes code that cryptography researchers can use to reproduce the recovered seeds and selected base points. The findings clarify how those parameters were generated; Anthropic reports no new attack or cryptographic weakness, and some origins remain unresolved.
Cursor cloud agents can now build interactive canvases, while cloud environments support up to 300 repositories and the Environments API enters beta. Developers can share canvas snapshots through browser links that give teammates read-only access without rerunning the agent. Sharing requires an eligible paid plan, team membership, and a privacy mode permitting data storage; administrators can disable it.
NVIDIA has released Nemotron 3.5 Super VL for agents working with images, webpages, and documents, alongside Bolt 74B-A9B for local workloads in 24GB of memory. Both are available now through downloadable model options. A smaller Bolt E4B model is scheduled for October 23, so that version remains an announcement.
OpenAI is rolling out GPT-Live-1.1 and the smaller GPT-Live-1.1 Quip in ChatGPT Voice, bringing more natural accents across languages. New voice options cover French, German, Hindi, Japanese, Korean, and Mexican Spanish, alongside a new English voice. Available choices vary by region and account, so users may receive different options during the rollout.
Mistral has released Timestral, four time-series forecasting models ranging from 5 million to 209 million parameters. Developers can download the weights and run forecasts on CPUs, while Mistral Forge supports adaptation to proprietary data. The models handle multiple related series and known future inputs; their weights use a Modified MIT license.
GitHub's agentic autofix is generally available to Code Security and Advanced Security customers on Enterprise Cloud. Assigned default-branch alerts prompt Copilot to explore files, propose changes, rerun CodeQL, and open a draft pull request for review. Administrators must enable access and configure billing; usage consumes Code Security AI Credits and GitHub Actions minutes.
Meta has released Muse Spark Open, a multimodal model for coding and multi-step agents with a million-token context window. Developers can download its weights or use hosted inference partners, while Meta describes local deployment using two M5 Ultra Mac Studios. Its custom Muse Model License includes commercial conditions, so access to weights does not mean unrestricted use.
No updates in this category on this page.
Earlier updates