Mistral releases improved Large 4 weights and halves API pricing
Mistral released downloadable FP8 and NVFP4 weights for its improved 1T-parameter, roughly 52B-active multimodal MoE, adding low, medium and high reasoning modes; Mistral reports a 13-point Terminal-Bench 4.0 gain to 41%, but these headline evaluations are vendor-reported. It can power self-hosted coding, defensive-security and document agents; the license requires a separate agreement for model-as-a-service businesses. Access: Mistral API, hosted services and local/self-hosted weights. API pricing: $0.70 input and $2.10 output per 1M tokens; cached input $0.07/M (Mistral, checked 2026-10-30). Local: official FP8 and Blackwell-targeted NVFP4 checkpoints exist, but exact RAM/VRAM requirements were not publicly verified; the model card demonstrates eight-way tensor parallelism without naming a minimum GPU configuration. [1] [2] [3]
Droid 1.0 batches agent tool calls into sandboxed programs
Factory made programmatic tool calling the default in Droid 1.0, allowing supported models to write sandboxed JavaScript that coordinates hundreds of tools while keeping raw results out of model context. It is useful for large codebase searches, ticket migrations and CI sweeps; Factory's own Terminal-Bench 4 runs report better cost-adjusted results than several rival harnesses, but the comparisons are vendor-run and Droid's sandbox can still invoke whatever external capabilities its configured tools expose. Access: desktop app, CLI and Apache-2.0 TypeScript SDK; it uses hosted or user-supplied models rather than releasing a new model. [4] [5]
Gemini adds in-chat editing for PDFs, documents and code
Google now lets Gemini users generate, preview and directly edit PDFs, documents, spreadsheets, slide decks, Markdown, text and code inside a conversation, including highlighted-region requests. It can turn uploaded notes into an editable study guide or refine a spreadsheet without switching applications; PDF editing works on web and mobile, while the other formats are web-only, and Google provided no independent accuracy evaluation. Access: hosted Gemini app for all users in supported languages; no separate API capability or local package was announced. [6]
Particle-1 reduces reasoning-token use for coding agents
Particle post-trained DeepSeek V4.1 Flash with reinforcement learning to create Particle-1, reporting about 45% fewer tokens on average while maintaining similar pass rates across four software-engineering evaluations. It is aimed at lower-cost coding and data-engineering agents, but results are vendor-reported, the training focus is narrow, and stronger compression can omit useful tests or qualifications. Access: Particle API. API pricing: not publicly verified from Particle as of 2026-10-30. Local: no released weights verified. [7]
OpenBot packages persistent multi-agent workspaces on the desktop
Discovery: OpenBot provides separate local workspaces, queues, file transfers, browser control and agent-to-agent messaging for Codex, Claude, Grok, Gemini and other CLI agents. GitHub's daily Trending page displayed 709 stars today on 2026-10-30, signaling attention rather than quality or adoption. The repository has substantive documentation and packaged releases, but it remains a development preview whose agents run with danger-full-access and no per-action approval after initial consent; its noncommercial license and dependence on external model services are additional constraints. [8] [9]