Codex CLI 0.164 adds presets and keeps your cache, Block's speech-only Dessa, and the Mercury 3 diffusion LLM

Codex CLI is OpenAI's coding agent for the terminal. Version 0.164.0 lets you save frequently used prompts and slash commands into ten numbered /presets slots and pull them up from the keyboard. /model and /permissions now take arguments, so you can set model, reasoning effort, speed, and permission profile in one line, and enterprise MCP (the protocol for connecting external tools) servers can authenticate with Codex account grants or a dedicated sign-in. On the same day, OpenAI changed Codex and Work so that switching reasoning effort mid-conversation no longer invalidates the cache.

npm install -g @openai/codex@0.164.0

A common pattern is low reasoning effort for quick tasks and high effort for hard ones, but until now each switch meant re-reading a long context, which cost money and time. Now you can change effort freely within one conversation. Credential masking for remote Bash commands and the Amazon Bedrock setup wizard (now on by default) are immediately useful for teams with strict security rules or those running Codex on AWS accounts.

Dessa is a speech-to-speech conversation model from Block's AI research lab. It was trained from scratch on human spoken conversation only, with no text and no pretrained models. Because it's full-duplex (it listens while it talks), it can drop in a 'yeah' or 'got it' while you're speaking, and it expresses laughter, pauses, and emotion in its voice. Audio goes in and audio comes out, with no separate LLM in between. Block says it trained the model on a single NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or betterVendor claim.

Most voice agents today chain speech recognition, an LLM, and speech synthesis, or bolt voice onto a text LLM. Dessa takes a different route and learns language from speech alone. A model small enough to run on a laptop matters for on-device assistants and privacy-sensitive support calls. That said, its knowledge comes only from conversational data, and the voice is hard to tell from a human; Block itself says it wants deception risks tested first, so it's too early to build products on it.

Mercury 3 is Inception's diffusion LLM. Instead of writing one token at a time, a diffusion LLM drafts the whole answer and refines it over several passes, which makes it fast. Context grows from 260K tokens in Mercury 2.5 to 1M, and the API is OpenAI-compatible, so existing clients only need a new base URL and model name. It's available on the Inception API, OpenRouter, Baseten, and Vercel.

curl https://api.inceptionlabs.ai/v1/chat/completions -H "Authorization: Bearer $INCEPTION_API_KEY" -H "Content-Type: application/json" -d '{"model":"mercury-3","messages":[{"role":"user","content":"Hello"}]}'

Inception reports 1,068 tokens per second on NVIDIA GPUsVendor claim and says it's 4–8x fasterVendor claim than GPT-6 Luna and Claude Haiku 5.5. Pricing is $0.20 per million input tokens and $0.70 per million outputVendor claim, half that during the launch promotion. For latency-sensitive work like classification, extraction, and autocomplete, it's a candidate to replace small autoregressive models. Inception hasn't published a named benchmark table, though, so test quality yourself.