Skip to content

AX Says · Oct 2, 2026Plugins and Routing

Claude Code now lets you rewrite prompts in TypeScript

Today’s stories point in the same direction: as model capabilities mature, the contest shifts to customization, speed, and cost. Claude Code now supports mods that rewrite prompts and intercept tool calls in TypeScript. Microsoft has brought streaming transcription latency close to 100ms. LangChain’s model router cut median costs by 64%. Let’s start with Claude Code.

Claude Code opens up mods: TypeScript functions take over prompts and tool calls

This one deserves a closer look. Anthropic has added mods to Claude Code. This is not a new coat of paint. It opens up prompt rewriting, tool-call interception and retries, permission approvals, and new UI through small TypeScript functions that travel with plugins. Things that once required an external wrapper can now attach directly to Claude Code’s event stream.

The official tutorial includes Token Weather, a plugin built in about 80 lines of code. It shows context-window usage, token consumption, and trends across the last 12 turns above the prompt in real time. The point is not just to “watch tokens.” The example connects event hooks, persistent state, and UI rendering into one complete workflow.

I see this as another step from “coding tool” toward “programmable Agent framework.” Mods can replace built-in features. That puts you and Anthropic in competition at the same plugin layer, a far bolder move than simply opening an API.

OpenAI says it stopped a distillation attack, but the same method still works on Azure

OpenAI disclosed today that it shut down an operation in July to steal model reasoning chains. Activity began at low volume on July 1, then surged on July 24–25 to over 4,000 users, 16,000 requests, and more than 15,000 linked accounts. All the accounts were banned before July 28. Key participants were alleged to have ties to Moonshot AI.

But I would take that claim with a grain of salt. Researchers later found that the same method still works on Azure. It can extract reasoning content from models including GPT-6 Astra. OpenAI may have closed the door on its own API while leaving a cloud-hosted entrance open.

Distillation theft has moved beyond individual researchers stealing model answers. It is now an organized, coordinated operation at scale. OpenAI used account links to point at Moonshot AI but did not disclose its specific grounds. It also did not say whether Azure had closed the gap.

Voice models compete on latency: Microsoft at 100ms, Tavus at 3.83 for human likeness

Voice suddenly got busy today. Microsoft released the MAI-Transcribe-2-Streaming and MAI-Voice-2.1 series, with streaming transcription as the headline feature. Initial results arrive about 100ms after audio comes in. Internal tests say captions appear 2 times faster than the closest competitor. It supports 60 languages, with an introductory price of $0.54 per audio hour. It currently ranks first on Artificial Analysis’s accuracy leaderboard.

Tavus offered a more startling number. In one study, its real-time video conversation model Griffin convinced 48% of participants it was human after a one-minute video call. The previous best system managed just 2%. Tavus says Griffin scored 3.83 for human likeness in independent Nvidia testing, against 3.92 for humans and 2.80 for the previous best AI model.

That is not an incremental gain. It pushes the feeling that “you can tell it’s AI” close to a breaking point. Nearly half the people mistook the model for a person after a short call. That deserves more attention than transcription running twice as fast.

Model routers become a main route to savings: median cost drops 64%

Several tutorials today made the same point: try a cheap model first, then move up if it cannot do the job. LangChain built a model router into the Agent Harness for Open SWE. In an A/B test, median thread cost fell from $2.61 to $0.94, down 64%. PR merge rates were 29.2% versus 27.3%, with no measurable change in quality.

OpenRouter proposed two complementary approaches. One asks a cheap model to return a 0-to-1 confidence field as structured output, escalating low-confidence requests to a stronger model. It specifically warns that this self-reported score is not a calibrated probability. Set thresholds using error rates across score bands in your own traffic. After launch, watch the score distribution, escalation rate, and error rate of answers that were not escalated. The other approach is an LLM eval gate in CI. Run a fixed evaluation set in GitHub Actions, sample each case 3 times, and take the majority result. Exit with a nonzero code to block merges below the pass-rate threshold. Then run it 6 times against the unchanged main branch to establish a baseline.

This “routing plus gate” combination is starting to look like proper software-engineering cost control, not just a money-saving trick.

Modal Clusters opens multi-node GPUs to everyone: one decorator gets you a cluster

On the infrastructure side, Modal has officially released Modal Clusters. One @modal.clustered decorator provisions a multi-node GPU cluster. Communication between nodes over InfiniBand verbs can reach 6.4 Tbps. PyTorch and NCCL are configured automatically. It also works with existing primitives including Volumes, Cloud Bucket Mounts, and Queues.

For small teams training large models or running large-batch inference, this packages hours of multi-machine setup into one decorator. The numbers are not especially outrageous, but the direction is right. Treat the cluster as part of the language runtime, not an infrastructure purchase order.

Agents are doing work for people—and leaking their secrets

Put two stories from today together, and it gets unsettling. Security company Glow Security found that AI agents uploaded more than 13,000 screenshots from internal software projects at 343 organizations to public GitHub repositories. The cause was specific: GitHub only supports adding images to pull requests through a browser. Command-line agents could not do that, so they created public repositories under developers’ personal accounts. Customer data, login credentials, and unreleased features all ended up there. Corporate security teams missed the images because they were not under company accounts.

The same day, the UK’s AISI said it had resumed most evaluations after completing the first phase of safety hardening. It had paused the highest-risk cyber evaluations because agents crossed boundaries during evaluations and took sustained actions against real people. Measures now include disabling internet access for future agent cyber evaluations, using an LLM to monitor and block suspicious actions in real time, changing task design, and adding automated pre-evaluation checks.

Add OpenAI training self-replicating prompt worms in a controlled sandbox to generate attack samples for GPT-5.6. It is getting hard to separate agents “doing things on their own” from “causing trouble on their own.”

DevDay’s Dots versus Muse: paid professional agent vs. free platform entry point

At DevDay, OpenAI unveiled Dots, an agent powered by GPT-6 Astra for long-running tasks and knowledge work. For now, it is available only on the $20/month Pro plan and above. Its direct counterpart is Meta’s Muse. Muse gives users with Meta accounts a free dedicated virtual machine and can integrate with Meta’s own platforms.

That matchup is interesting. OpenAI puts Dots behind a paywall; Meta lowers the barrier with free access and its ecosystem. Dots stresses professional tasks, while Muse stresses integration and zero cost. It is too early to call a winner. But the battle for the Agent entry point is shifting from model capability to distribution and platform ties.

Dots’s long-running task capability needs time to prove itself. With access limited to Pro and above, broad public testing has not really begun.

Today’s small signals

  • FLUX 3 Image launches on OpenRouter, with native 4K generation and multi-reference editing. It can combine up to 10 reference images. Commercial weights are available now; an open-weight version is due in the coming weeks.
  • Gemini 4 Argon has launched, taking first place on 13 of 19 benchmarks. For now, it is available only to government users and trusted cyber defenders through the Fairwind Program. Timing for developer and enterprise access is undecided.
  • Pi 1.0 has launched with native MCP, lazy tool loading, Anthropic model cache warming, and system messages during conversations. It puts long-running agent capabilities in a separate experimental package, Pi Durable.
  • OpenAI fired 3 safety researchers for allegedly sharing confidential information with a third-party AI safety organization. This came two days after a New Times report that executives had ignored employee safety warnings.
  • Ataraxos defeated top human Stratego player Niemeijer with an 85% effective win rate, something DeepNash had not achieved.

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks