Databricks's catalog exit is a preview, so test before you unbundle

*AI agents make it tempting to cut software bills, but your data is only yours once a rival system can take it over cleanly.*

Opening read

Keep paying for the part of your software that holds your data, and use the next renewal to prove you can leave it. Databricks published an exit path for its data catalog today, and it is a preview. My read is that the unbundling of business software is real and early, and buyers who test exits now will set the terms later.

Software buying: prove the exit before you unbundle

The strongest case for rethinking software bills this week comes from Nate Jones. He argues that a subscription has always sold three things as one package: the data it stores, the screens people work in, and the workflows, meaning the built-in steps that move work forward. His point is that agents (programs that act on their own) are getting good enough to do the looking-up, the sorting, and the doing themselves, so each part of the bundle now has to justify its own price. Nate Jones (https://natesnewsletter.substack.com/p/ai-agents-software-buying) I think he is right about the direction.

Where I part ways is on timing. You can only stop paying for one piece of a bundle if the pieces you keep can move, and data has been the hardest to move. In July, Databricks made its Unity Catalog the official record-keeper for the tables it manages, a catalog being the system that tracks which version of each data table is current. Databricks pitched the change as what makes writes from outside tools safe, but that safety runs through its own catalog. Snowflake, a rival, noted that month that Databricks's own engines got full read and write access through an outside catalog only with Unity Catalog switched off. Databricks Blog (https://www.databricks.com/blog/how-unity-catalog-managed-tables-bring-interoperability-performance-and-unified-governance) Snowflake Blog (https://www.snowflake.com/en/blog/bidirectional-interoperability-snowflake-horizon-databricks/) Whoever keeps the record controls the table.

Today Databricks offered a way out. It added a step called UNREGISTER to the catalog standard for Apache Iceberg, a shared format for storing large tables. The step lets one catalog give up a table without deleting anything. It then hands the next catalog the pointer it needs to take over, so no files are copied or rewritten. The reason it was needed is telling. A table could be added to a second catalog, but the first had no way to let go, so both could keep acting as the owner. Queries then start to disagree, and data can go missing without any error. So the safe hand-off is new to the standard, and Databricks's own version of it in Unity Catalog is a private preview you request through your account team. Databricks Blog (https://www.databricks.com/blog/unlocking-data-portability-preventing-catalog-lock-register-and-unregister-apis)

The workflow piece is at the same stage. OpenRouter, a service that routes requests to many AI models, has two tools that let any model reached through its two main request formats run commands in a sandbox, a walled-off computer kept apart from everything else. One follows OpenAI's format and the other Anthropic's, so the doing part of software can be swapped across model makers. That is the action layer Jones means only if your workflows run through AI models and not a vendor's built-in steps, and even then both tools are still in beta, an early test version. OpenRouter Blog (https://openrouter.ai/blog/insights/server-side-code-execution-tools-for-ai-agents-compared/) OpenRouter Blog (https://openrouter.ai/blog/announcements/shell-tool/) Portable, in other words, but not yet proven.

Here is a check a business leader, or the engineer who owns the data, can run before the next renewal. Split your largest subscription on paper into Jones's three parts. For the data part, ask the vendor one question: can another system take over as the official record without an export, and will yours stop accepting changes once it does? Count the products where the answer is a documented yes. Those are the contracts you can renegotiate from strength; everywhere else, cutting the screens or workflows first would strand your data. I do not know how many vendors will follow Databricks, and the evidence cannot show that its preview holds up under real workloads.

Local models: the tokenizer fix is the one to retest

Engineers running models on Macs should rerun their test prompts when Ollama's next version ships, because a correctness fix is inside it. Ollama, software for running AI models on your own machine, has refreshed MLX, the engine it uses on Apple chips, in release after release since July. The 0.40.0 test builds go further. They change how text is cut into tokens, the small pieces a model reads, to match what each model's publisher intended. They also switch supported models onto MLX by default on Apple chips, so the new behavior reaches Mac users who never chose it. The notes I have do not say how answers shift, but a tokenizer mismatch is the kind of fault that can make a local model answer differently than its maker's tests suggest. Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.32.3) Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc1) Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc3) That is reason enough to retest.

Model serving: GLM-5.3-Flash gets SGLang support

GLM-5.3-Flash now has a serving path, and that is all the evidence establishes. SGLang, a tool for serving models, meaning running them for many users on company hardware, added support for it in version 0.5.21. SGLang Releases (https://github.com/sgl-project/sglang/releases/tag/v0.5.21) My August call was that Z.ai would publish GLM-5.3 open weights (downloadable model files) by September 30, and it was graded a provisional hit, meaning it looks right pending final confirmation. The Flash name points to that same GLM-5.3 family, but that is my inference from the name. The release-note evidence I have says only that SGLang supports the model. It does not settle the call, and it is not a performance result. For engineers already serving models with SGLang, this is a candidate for the trial list once they have confirmed the model's availability and terms themselves, not a reason to switch.

Vendor risk: Anthropic's IPO timing is still unconfirmed

The open question on Anthropic's IPO (a first sale of shares to the public) is when it happens, not whether a process exists. In April the rumor, relayed by Nate Jones, was an offering in October 2026. By mid-September TechCrunch was reporting that the company was preparing its S-1, the disclosure document a company files before selling shares to the public, and Latent Space listed an IPO filing in its September 30 roundup. The AI Daily Brief now puts the target before Thanksgiving, a window wide enough to still include October. Nate Jones (https://www.youtube.com/watch?v=9N7qXkmntlU) TechCrunch AI (https://techcrunch.com/2026/09/13/whats-behind-the-ai-industrys-latest-warnings-of-doom/) Latent Space (https://www.latent.space/p/ainews-openai-devday-2026-dots-61) AI Daily Brief (https://podcasters.spotify.com/pod/show/nlw/episodes/What-the-Best-Business-AI-Users-Are-Doing-Different-e3pp6f5) So the schedule is imprecise, and I cannot call it delayed. The evidence I have contains no confirmation from Anthropic. Buyers weighing a long contract for Claude, Anthropic's AI model, should treat the date as soft and keep first terms short. I expect a public offering to bring more financial disclosure than customers get today, so my advice to a business leader is to wait for those numbers before locking in multiple years.

Augur calls

• Call: Anthropic's shares begin trading on a public stock exchange on or before 2026-11-25.
• Criteria: Hit if an exchange listing notice, an Anthropic announcement, or a regulatory filing shows a first day of public trading on or before 2026-11-25; miss if no Anthropic shares have traded publicly by that date.
• Due: 2026-11-25
• Odds: 35%

What to watch

• Databricks moving its exit APIs (the hooks software uses to talk to other software) out of private preview, or a second catalog vendor supporting the same hand-off, would make the exit testable by anyone. Databricks Blog (https://www.databricks.com/blog/unlocking-data-portability-preventing-catalog-lock-register-and-unregister-apis)
• OpenRouter's command tools leaving beta would signal that the cross-model action layer is ready for real workloads. OpenRouter Blog (https://openrouter.ai/blog/insights/server-side-code-execution-tools-for-ai-agents-compared/)
• Trump's new Super Intelligence Force, announced October 4, has a leader in National Intelligence Director Jay Clayton, and the Wall Street Journal reported that FTC Chair Andrew Ferguson will be a vice chair. The reporting I have does not spell out the force's authority, so a charter would tell business leaders if it touches buying or enforcement. TechCrunch AI (https://techcrunch.com/2026/10/04/trump-unveils-his-new-super-intelligence-force/)
• Ollama's 0.40.0 is still a test build. Its final version is the moment to rerun Mac test prompts. Ollama Releases (https://github.com/ollama/ollama/releases/tag/v0.40.0-rc1)

Verification notes

The exits in this lead are promised more than delivered: Databricks's REGISTER and UNREGISTER APIs for Unity Catalog are in private preview behind an account-team request, and OpenRouter's command tools are in beta. Databricks Blog (https://www.databricks.com/blog/unlocking-data-portability-preventing-catalog-lock-register-and-unregister-apis) OpenRouter Blog (https://openrouter.ai/blog/insights/server-side-code-execution-tools-for-ai-agents-compared/) Nate Jones's three-part split of software is his analysis, and the evidence contains no measured results from buyers who have unbundled. The GLM-5.3-Flash item rests on a single line of release-note evidence, and its link to my August call is an inference from the model's name. The IPO timing rests on press reports of a filing process and a podcast forecast, with no confirmation from Anthropic in the evidence.

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-05?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-05)