Descript's router switch: buy the testing loop, the model choice follows

*A week in the engineering queue shrank to two hours; the dependence did not vanish, it moved to the router's catalog.*

Opening read

Descript's account of moving its Underlord agent onto OpenRouter is the most useful item today. The useful part is a clock. Adding a model used to mean a week in an engineering queue. Now it means an evaluation run inside two hours. My read is that the testing loop, not the router, is what let the company cut its model spend in half. Its dependence on outside vendors didn't shrink so much as relocate. If you build with AI, that clock is the number to measure in your own shop before you buy any gateway.

Model routing: the week of waiting was the real switching cost

This brief has been arguing that cheap open-weight models (models whose trained parameters anyone can download) belong in the fallback slot behind the frontier. Descript is the first shop here to show what it costs to get one into production. Its account of running Underlord, its video-editing agent, is mostly about plumbing. I think the plumbing matters more than the fallback.

The consensus read is fair. A router such as OpenRouter is a middleman. It's a service that lets one codebase reach many AI models through a single connection, and a serious product should keep direct lines to the labs it depends on. Descript had that, with direct integrations to OpenAI, Anthropic, and Google. Its problem was never the price of a call. Adding any new model needed engineering time. The queue often meant a week of waiting before anyone could learn whether the model was worth having. OpenRouter Blog (https://openrouter.ai/blog/insights/descript-model-evaluation-queue/) The lines to the labs were open. The line to a decision wasn't.

That week is the number to care about. Aleks Mistratov, Descript's head of AI product, says the team can now evaluate a new model inside an hour or two. The post describes the loop as running through Slack. There, Claude Tag, Anthropic's Slack integration, kicks off the evaluation runs and opens the pull requests (proposed code changes) that follow. The earlier case study supplies the consequences. Descript had 13 models in production as of August. Its total model spend fell to under half its old level once open-weight models took share. OpenRouter Blog (https://openrouter.ai/blog/insights/descript-model-evaluation-queue/) OpenRouter Case Study (https://openrouter.ai/blog/case-studies/descript-case-study/) I read the spend cut as downstream of the speed. OpenRouter itself warns that a low sticker price can turn expensive once a model retries, runs long, or needs a stronger model to catch its failures. That's exactly why the testing loop, and not the unit price, is what banked the saving. OpenRouter Tutorial (https://openrouter.ai/blog/tutorials/choose-best-ai-model/) A cheap model you can't test cheaply never gets used.

Where I depart from the router-as-freedom story is lock-in. It didn't disappear. It moved. Mistratov's own rule is that a provider missing from OpenRouter has to make an offer he can't refuse before he will wire it up directly. So the router's catalog now decides which models Descript sees at all. OpenRouter Case Study (https://openrouter.ai/blog/case-studies/descript-case-study/) Catalogs lag. I called that Together, Fireworks, or OpenRouter would host DeepSeek-V4-Flash by September 15, and that call is provisionally a miss. Descript hedges the dependence in one sensible way. It keeps its own open-weight deployments on Baseten, which the router tries before anyone else, with the rest of the market held in reserve. OpenRouter Blog (https://openrouter.ai/blog/insights/descript-model-evaluation-queue/)

None of this tells you whether to buy. For an engineer deciding whether a gateway is worth another layer, the check is simple. Take the last model you added. Count the calendar days from its public release to its first evaluation run in your product. Then split those days into engineering work and queue waiting. If waiting dominates, Descript's fix targets your problem. If the suite itself is the slow part, a router changes nothing, and the budget belongs there first.

Agent testing: replay the incident, hold the model still

In June, Microsoft engineers presented Chronicle as a way to record an agent run at every boundary and freeze the full state there. A boundary means each point where the agent calls a model or a tool. Freezing the state lets the untouched parts be stubbed, meaning replayed from the saved answer instead of calling the model again. That way one code change can be tested on its own. AI Engineer (https://www.youtube.com/watch?v=Lc8zRh9muoY) The published paper turns that into a method. Cut-point replay serves some recorded model and tool calls from the incident and runs the others live against new code. A production failure becomes a regression test (a check that a change did not break what worked). The authors report that recording adds 23 microseconds per boundary and that a full replay makes zero model calls. Across their six recorded incidents, they say the tests failed on faulty code and passed on benign changes. They also report catching every mutation that let a recorded unsafe action through. A mutation here means a deliberately broken version of the guard code. A baseline that stubbed every boundary caught none. Hugging Face Daily Papers (https://tldr.takara.ai/p/2609.20625) Six recorded failures replayed through simulated model boundaries is a small proof. Still, it answers what Descript's loop leaves open for an engineer deciding how to test agent code without paying for live calls. It tests the code while the model holds still.

Voice models: a 9B model beat a model Google already replaced

Realtime-Venus is a pair of separately trained 9B models (nine billion parameters, small by frontier standards). One handles speech. The other handles speech with video. Full-Duplex-Bench v1.5 scores a model that listens and talks at the same time. On it, the paper reports the speech model answered 75% of interruptions and kept going through background speech 86% of the time. On the authors' numbers, that puts it ahead of Gemini 3.1 Live and GPT-4o. Hugging Face Daily Papers (https://tldr.takara.ai/p/2609.13814) Those targets are already a generation old. Google's announcement dates Gemini 3.8 Live's release to September 15, and OpenAI put GPT-Live-1 into its API on September 10, both before this brief. Google DeepMind Blog (https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/) OpenAI News (https://openai.com/index/introducing-gpt-live-1-in-the-api) An engineer picking a voice model should rerun the interruption test against those two before a small model earns a slot. The paper can't say how it fares against what shipped this month.

AI spend: the customer should be able to tell

Salesforce executive Rohan Kumar relayed a customer who couldn't square AI's supposed potential with products that look much like last year's. Nate Jones's reply is a test. If the technology matters, the customer should eventually feel it. They'd feel it in work that couldn't be done before, a price that didn't pencil out before, or a level of care no firm could afford to hand out. Nate Jones (https://natesnewsletter.substack.com/p/ai-cost-to-serve-customers) Cost is the reason they can't feel it yet. Working out what a customer needs and organizing an answer used to cost a lot. So firms rationed who got a person, who got a form, and who got a tailored reply, by what each customer was worth. Jones's own time study in July found that pulling a customer's records together from separate systems ate the time on a support ticket. Writing the answer barely registered. Nate Jones (https://natesnewsletter.substack.com/p/first-ai-agent-use-case) For a business leader placing next quarter's AI budget, that is the audit. Spend that lands on drafting replies buys a faster typist. Spend on the record-gathering step is what could change who gets a human.

Augur calls

• Call: ScrollEd will have a publicly downloadable consumer app listed on the Apple App Store or Google Play by 2026-10-31.
• Criteria: Hit if a ScrollEd-branded app from the company is publicly listed and installable on either store by the due date, shown by the store listing or a company announcement linking to it; miss otherwise.
• Due: 2026-10-31
• Odds: 55%

What to watch

• ScrollEd says it plans to launch its consumer feed at TechCrunch's October conference. If that launch happens, it will show whether turning textbooks into short video has an institutional buyer or only scrollers. TechCrunch AI (https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/)
• Google's Gemini 3.8 Live has been public since September 15, per Google's announcement. The next full-duplex benchmark round should include it. Google DeepMind Blog (https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/)
• The Chronicle code and its six-incident benchmark are public. Any team that replays its own agent traces through it will tell us whether zero-model-call replay survives real boundaries. Hugging Face Daily Papers (https://tldr.takara.ai/p/2609.20625)
• vLLM's 0.30.0 release lands a build fix for DeepGEMM on CUDA 12.9. Teams that serve open models on their own chips should confirm their pipelines pick it up before the next release candidate. vLLM Releases (https://github.com/vllm-project/vllm/releases/tag/v0.30.0)

Verification notes

Descript's before-and-after figures come from OpenRouter's own blog and aren't audited. The evidence can't show what its evaluation suite measures or whether output quality held after the spend cut. OpenRouter Blog (https://openrouter.ai/blog/insights/descript-model-evaluation-queue/) Chronicle's results rest on six recorded failures with simulated model boundaries. Realtime-Venus's comparisons are against Gemini 3.1 Live rather than the 3.8 Live released September 15. The evidence also says nothing about whether Realtime-Venus's weights are available to download. Hugging Face Daily Papers (https://tldr.takara.ai/p/2609.20625)

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-21?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-21)