Mercury 3 makes diffusion fast and cheap, MiniMax undercuts on agent cost, and Claude Code learns your setup

1. Mercury 3: diffusion LLMs get serious

Inception launched Mercury 3 on October 19. Instead of writing one token at a time, it generates text by refining many tokens in parallel, and Inception reports 1,068 tokens per second on NVIDIA GPUs, 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5, which it says it matches or beats on quality. Context jumps to 1M tokens (from 260K), and it supports tool calling and structured outputs. List price is $0.20 in and $0.70 out per million tokens, half that at launch, with cached input 90% off. It's OpenAI API compatible and available on Inception's API (100M free tokens to start), OpenRouter, Baseten, and Vercel AI Gateway as inception/mercury-3. Weights aren't released.

Our take: The benchmarks are Inception's own, and its co-founder says plainly it won't replace Opus-class models as the main driver of a coding agent. Where it shines is the stuff that has to feel instant: voice bots, browser agents, and coding subagents.

2. MiniMax M3.1 Flash

MiniMax released M3.1 Flash on October 19, an everyday model for coding, research, and computer-use work, available on the MiniMax API and in MiniMax Code. It has a 1M-token context and flat pricing of $0.15 in, $0.60 out, and $0.015 cached per million tokens, and MiniMax says agent tasks cost 70% less than on its predecessor. Independent testing from Artificial Analysis scored it 42 on its Intelligence Index (up from 29 for M3), near GLM-5.3 Flash and just behind DeepSeek V4.1 Flash-1017 at 45, at about $0.22 per task. The catch from the same testing: it's unusually likely to give a wrong answer instead of admitting it doesn't know.

Our take: Great price for agent loops where tests or tools check the work. Don't let it answer factual questions unsupervised.

3. Claude Code v2.1.304

Anthropic's October 19 Claude Code release adds /auto-mode-setup, a guided setup that reads your project and recent sessions to teach auto mode about your environment. It also adds /claude-api managed-agents-secure, which compares a Managed Agents agent's settings against Anthropic's published security guide without changing anything, plus a decider: "person" option so a mod can require a human to approve a tool call. On the team side, Code Review gets Quick, Standard, and Deep levels and an option to cover every repo in a GitHub org. More than 40 fixes include one for worktrees and branches being deleted on exit.

Our take: If you've avoided auto mode because it kept asking about things it should know, run /auto-mode-setup once. And update today if you use worktrees.

4. Cloudflare Clairic: answers as values, not text

Cloudflare's Workers AI team previewed Clairic (@cf/cloudflare/clairic) on October 19. You send text (and up to four images) plus a schema of typed fields, and it returns one value per field: an exact text span, a number, a date, a duration, a list, an estimate with a p10 to p90 range, or a short forecast. Every answer comes with a confidence score and a p_null probability that the input doesn't contain it. It reads the input in one pass rather than generating tokens, costs $0.10 per million input tokens, and Cloudflare says open weights are coming. Cloudflare's own docs warn that long chains of arithmetic aren't reliable yet, so do the math in code.

Our take: For invoice totals, dates, and form fields, a model that can say "probably not in here" is more useful than one that confidently writes JSON. Use the confidence scores to decide what goes to human review.

5. Headroom: give your agent less to read

Headroom topped GitHub's daily trending list this week. It's an Apache-2.0 tool that compresses what an agent reads (tool outputs, logs, files, search results, and history) on your machine before it reaches the model. You can use it as a Python or TypeScript library, a local proxy (headroom proxy --port 8787), an MCP server, or a one-line wrapper like headroom wrap claude (Codex, Cursor, OpenCode, Cline, and others are supported). Originals are cached locally so the model can fetch the full text if it needs it. The project claims about 20% fewer tokens for coding agents and 60 to 95% fewer for JSON.

Our take: Those savings are the maintainers' numbers, so check headroom dashboard on your own sessions. Long log dumps and giant JSON responses are where it should pay off first.

Try this today

Put a spending cap on every AI provider key you own. This week OpenCode disclosed a breach of a legacy Console system that exposed provider API keys for 1.4% of users who had linked OpenAI, Anthropic, or Google accounts, and some people saw surprise bills. Take ten minutes: list every key you've pasted into a tool, revoke any you don't recognize or no longer use, set a monthly budget limit in each provider's billing settings, and give each tool its own key so you can kill one without breaking the rest.