AI news, October 21, 2026: Google open-sources TorchTPU, Mercury 3 diffusion LLM, MiniMax M3.1 Flash pricing
Five stories from October 19โ20, 2026 (US time) for people who build with AI models or pick which API to pay for. Each item links to the original announcement or reporting; prices and benchmark numbers are the vendors' own unless the source says otherwise.
Google open-sources TorchTPU so PyTorch apps can move to TPUs
Google announced that TorchTPU, its PyTorch version for Tensor Processing Units, is now available as an open-source project. TechCrunch reports that existing PyTorch apps can be moved to TPUs with a configuration change instead of a rewrite, and that Google worked with Meta, Hugging Face and the makers of vLLM and SGLang on compatibility. Google also launched a TPU agent suite to help with migrations and updated LiteRT for PyTorch models on devices.
Why it matters for makers: one day after Amazon's TorchNeuron, both big clouds now offer a low-effort path off Nvidia-only PyTorch code, which can mean more pricing options for training and inference.
Source: TechCrunch
Inception launches Mercury 3, a diffusion LLM at over 1,000 tokens per second
Inception released Mercury 3, a language model that generates and refines text in parallel instead of one token at a time. The company says it runs at 1,068 tokens per second on Nvidia GPUs, has a 1M-token context window and matches or beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks; these results are Inception's own. List pricing is $0.20 per million input and $0.70 per million output tokens, half that during launch, and the API is OpenAI-compatible. It is available on Inception's platform, OpenRouter, Baseten and Vercel AI Gateway.
Sources: Inception blog, Vercel changelog
MiniMax M3.1 Flash goes on sale at $0.15 per million input tokens
Chinese AI company MiniMax released M3.1 Flash for agent, coding and computer-use work, with flat pricing of $0.15 per million input tokens, $0.60 per million output and $0.015 for cached input. MiniMax says it is nearly three times faster than its predecessor and cuts agent task cost by 70%. VentureBeat reports that Artificial Analysis gave it 42 on its Intelligence Index, up from 29 for MiniMax M3, but also found an unusually high tendency to give wrong answers rather than admit uncertainty.
Why it matters for makers: it is a cheap option for high-volume agent loops, but check outputs where factual accuracy matters.
Sources: MiniMax blog, VentureBeat
Claude Opus 5.5 and Sonnet 5.5 arrive in Google Antigravity
Google Cloud made Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 selectable in Antigravity, its agentic coding environment, for Google Cloud customers once an administrator enables them. The models work across Antigravity 2.0, the Antigravity CLI and its IDE extensions for VS Code, JetBrains IntelliJ and Visual Studio. Google says partner-model usage is billed pay-as-you-go through Google Cloud Billing rather than from the pooled Gemini Enterprise quota, so overages must be enabled.
Source: Google Cloud Blog
ChatGPT adds Locked Chats protected by a PIN or Face ID
OpenAI started rolling out Locked Chats, which move a conversation into a separate section protected by a four-digit PIN, Face ID or Touch ID. TechCrunch notes that OpenAI says locking does not add end-to-end encryption, does not change retention, and that resetting the PIN permanently deletes all locked chats. The feature is coming to Free, Go, Plus, Pro, Business, Enterprise and Edu plans.
Source: TechCrunch