|
Hey there 👋
I opened OpenAI’s deprecations page this morning expecting the usual line about GPT-5.5, and found gpt-4o and gpt-4.1 sitting on the list with a hard shutdown date. GPT-5.5 leaving the consumer apps was the story we tracked all week. The quieter one is that the two models most production agents were built on now have ninety days to live in the API.
That sets the table today. OpenAI closing its back catalog. SAP putting Joule agents inside live workflows at Novartis and Nestle. An AMD box that runs a 300-billion-parameter agent on your desk. A legal-tuned model that clears its own base by twelve points. And a small open-source server that stops your agent from grepping the whole disk.
- Ron
The Big Thing
OpenAI just put gpt-4o and gpt-4.1 on a shutdown clock
OpenAI posted four API deprecations this morning, October 14. gpt-4o (both the 2024-08-06 and 2024-11-20 snapshots), gpt-4o-mini, gpt-4.1, and gpt-4.1-mini all get removed from the API on January 14, 2027. Fine-tuned versions ride the same schedule as their base model. The recommended replacement for every one of them is gpt-6-luna.
Read that list again, because it is the workhorse shelf. gpt-4o and gpt-4.1 are the model strings a huge share of shipped B2B agents still call by name. Not the experimental tier. The one that has quietly run production tool-calling and extraction for a year and a half.
The consumer side closes today, separately. OpenAI’s Codex models page still says GPT-5.5 leaves ChatGPT, ChatGPT Work, and Codex on all plans today, October 14. That was the countdown. The API deprecation is the one with real migration cost, because a hardcoded string in a running agent does not get a friendly in-app nudge. It just starts returning errors on the shutdown date.
gpt-6-luna is not a free swap. It is a different model with its own pricing, latency, and behavior. A grader tuned on gpt-4o output can drift the day you repoint it, and small models earn their keep on narrow tasks where the bigger one was overkill. So treat “recommended replacement” as the start of a test you run yourself.
Ship it? MIGRATE, and start with an audit tonight. Grep your codebase and your config for gpt-4o, gpt-4o-mini, gpt-4.1, and gpt-5.5. Every hit is a dated liability. Pin each one to an explicit snapshot so nothing moves under you, then run gpt-6-luna against your own eval set before the holidays, while there is slack to do it calmly. The first week of January, with half your team out, is the worst time to meet a broken string. You have ninety days. The teams that miss this will find out at runtime.
Sources: OpenAI Deprecations; OpenAI Codex models page.
Tour de Headlines
🏢 SAP put Joule agents inside live production workflows. At SAP Connect, SAP said Joule Work starts rolling to every customer this month, and that Joule assistants are already executing inside real workflows at Novartis, Nestle, and PwC. The plumbing underneath is the interesting part: agents running on grounding data, talking to each other over the Agent2Agent protocol SAP co-developed with Google Cloud. A sourcing assistant that reads your supplier records guesses a lot less than one reasoning from a blank prompt. The productivity figures are SAP’s, reported from its own stage, so hold them loosely until a customer says it unprompted. What matters is the shift from keynote mockup to a workflow that moves real money. Verdict: WATCH, and copy the shape, which is agents grounded in your records, before you trust the headline number.
💻 AMD will sell you a box that runs a 300B agent on your desk. The Ryzen AI Halo Developer Platform (announced October 12) pairs the Ryzen AI Max+ PRO 495 with up to 192GB of unified memory, up to 160GB of it allocatable as GPU VRAM. That is enough to run 300-billion-parameter models at 4-bit, with several large agents holding separate KV caches at once, on ROCm under native x86 Windows. It is orderable from Micro Center now, though there is no price and the Windows build is still listed as coming soon. The real pitch is boring and good: an agent that lives on one machine has no egress to police and no data leaving the building. For regulated or sensitive work, that is the cleanest answer to the exfiltration question going. Verdict: PILOT if you build offline.
⚖️ A legal-tuned model beat its own base by twelve points. Z.ai’s GLM-Legal, built on GLM-5.3-Flash, scored 19.2% all-pass on the Harvey LAB legal benchmark, about 12.5 points above the model it was tuned from. Access terms are not public yet, so nobody is deploying this next week. The signal is what counts: a domain fine-tune posting a measurable gap over its generalist parent on hard legal work. For any narrow, high-value agent task, that is the case for tuning a smaller model on your own corpus instead of renting a frontier generalist and hoping. One benchmark from one lab is a data point, not a trend. Verdict: WATCH.
|
Sponsor
Your calls leave a signal. Read it.
RapportScore measures how your team communicates on real calls and coaches them on it. It scores what happened on the call, using human signals plus AI, so your people get better at the conversation itself, meeting after meeting.
See your team’s score →
|
Tool of the Day
🔍 Pensieve
A local-search server that hands your agent the right files instead of letting it grep the whole disk.
Pensieve (MIT, Mac) runs a local MCP server that answers an agent with the few relevant files and passages for a query, rather than making it crawl your filesystem token by token. For a local coding or research agent, that is a direct cut to cost and a way to keep file context on the machine. It is small and open-source, so you can read exactly what it indexes before you point it at anything sensitive. If your agent currently burns half its context window rediscovering where things live, this is an afternoon experiment with an obvious payoff.
Pensieve on GitHub →
Worth a Click
Today was a lifecycle day. One vendor cleared its shelf, and the bill for that lands on whoever hardcoded a model string eighteen months ago and forgot. So here is my bet, with a stake on it. The cheapest hour you spend this week is the one grepping for gpt-4o and gpt-4.1 and writing down the migration while there is slack to do it calmly.
The deeper move is owning the part of the stack nobody can deprecate out from under you. That is what AMD’s local box and Z.ai’s tuned model are both quietly about: the model running on your hardware, or trained on your corpus, does not get a shutdown date emailed to it. If I am wrong, it is because gpt-6-luna drops in clean for everyone and the migration is a non-event. I have not seen that upgrade go clean yet. I will take the audit.
- Ron
|
