- Towards AGI
- Posts
- Goldman Just Took AI Coding Agents From Its Bankers. Who Gets Yours?
Goldman Just Took AI Coding Agents From Its Bankers. Who Gets Yours?
who gets your ai, what it can finish, and what it really costs
Today, we're diving into:
Hot Tea: Goldman Sachs is pulling AI coding tools like Claude Code from its bankers and traders and keeping them for engineers
Gen AI: A tough new benchmark finds the best AI model completes just 12% of high-value agentic work, and finance and legal tasks score essentially zero
OpenAI: ChatGPT and Codex will draw from one shared, token-based allowance from December
Goldman Just Took AI Coding Agents From Its Bankers. Who Gets Yours?
For months, Goldman Sachs let its bankers use tools like Anthropic's Claude Code and Microsoft's GitHub Copilot on the job. Now those tools will be limited to the firm's engineers only, according to internal guidance published last month and seen by Bloomberg, PYMNTS reported on Monday.
This is the same bank whose CIO, Marco Argenti, was praising "vibe coding" by non-technical staff in July. Four months later, the experiment has a new owner.
Access Without A Plan Is Not An AI Strategy.
Find out which of your teams are running AI on data and permissions nobody has mapped.
What Goldman Actually Changed
Bankers and traders lose access. Goldman's engineers, a group of around 12,000 people, keep it. The tools themselves are not banned from the building. They have simply been reassigned.
"We make certain AI tools available to our people to determine the best use cases across different parts of the firm." - Nick Carcaterra, Goldman Sachs spokesperson
The firm also said that after a period of experimentation, assessing factors including productivity and risk management, it put the tools "in the hands of those who are best suited to leverage them for the long term."
Not Everyone Is Happy
A person familiar with the matter said the change frustrated some traders. They had been using AI for "monitoring client trading flows or creating written and graphic materials."
Read that list again. None of it is software engineering. It is exactly the kind of work that tends to migrate to whatever powerful tool is lying around.
This Is Not A Ban. It Is An Allocation Decision.
The easy headline is that a big bank got cold feet on AI. That misses the point. Goldman did not reduce its AI ambition. It decided that a tool which can write and run code belongs with the people trained to review, test and own that code.
The wrong reaction is to copy the move wholesale and lock everyone out. You would lose the experimentation that tells you where AI actually helps.
The better alternative is to match the tool to the job. A trader who needs to summarize client flows does not need an agent that can push code. An engineer does. Each role should get the most capable tool it can safely govern, and no more.
Questions You Should Answer This Quarter
Who is using agentic tools today? Build the list from licenses and logs, not from a survey.
What can each tool reach? Map the data, systems and credentials sitting behind every seat.
Who reviews the output? If nobody owns the code or analysis an agent produced, it should not be in production.
What replaces the tool for everyone else? Removing access without a safer option just pushes people to shadow AI.
Answering those questions starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer. It maps lineage across the modern data stack and identifies downstream dependencies, so you can see what an AI tool touches before you decide who gets it.
The organizations moving first will not just hand out the most powerful AI tools. They will know exactly who holds each one, and why.
The Best AI Model Finished 12% Of Real Work. Finance And Legal Scored Zero.
SemiAnalysis has launched GDPx, a benchmark built to measure general agentic capability in the domains where it argues the first trillion dollars of AI revenue will come from. Its verdict is blunt: even the best model scores just 12%.
That is not a typo, and it is not a weak model. It is what happens when frontier AI is asked to do long, messy work inside a realistic environment instead of answering a tidy prompt.
What GDPx Tests
11 domains. Chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering.
Real environments. Finance tasks, for example, sit inside a deal room of roughly 4,000 files.
Long runs. The default time limit is eight hours, and models run in their own native agent harnesses.
SemiAnalysis contrasts this with earlier benchmarks such as GDPval, which it says were impressive for their time but gave the model all the context it needed directly in the prompt.
Who Won What
According to SemiAnalysis, Anthropic's Claude Opus 5.5 is the best all-around model, winning in quant trading, hardware design and chip design. OpenAI's GPT-6 Astra wins on recursive self-improvement and materials science. Google's Gemini 4 Carbon is the best biology model.
The authors say their finance and legal tasks are so challenging that "basically every model scores exactly 0."
The Real Bottleneck Is Not Intelligence. It Is Context.
Look at where the models failed. Finance and legal are the domains where the answer lives in scattered documents, unwritten conventions and definitions that change from one deal to the next.
A model can be brilliant at reasoning and still fail if it cannot find the right file, trust the right version or understand what your firm means by a term. That is the gap between a demo and a deployment.
You should treat this as a planning signal, not a reason to stop. The better approach is to scope agents to well-defined steps inside these workflows, give them structured and governed context, and keep a human accountable for the final judgment.
Test on your own work. A public leaderboard will not tell you how an agent handles your deal room.
Break long tasks into checkpoints. Eight-hour autonomy is where errors compound quietly.
Pick the model per domain. GDPx shows no single model wins everywhere.
The organizations moving first will not just wait for a smarter model. They will build the context that lets today's models finish the job.
Starting in December, the usage allowance on ChatGPT's Plus, Pro and Business plans "will be shared across ChatGPT and Codex," according to OpenAI's help center, Notebookcheck reports. OpenAI says the change comes "as we bring the agentic capabilities of Work into the main ChatGPT experience."
On the surface, it is a billing tweak. Underneath, it is a preview of how every AI vendor will meter agentic work.
How The New Meter Works
Counted in tokens. Usage is measured in tokens, the units of text a model reads and writes.
Heavier work costs more. "More advanced models and complex tasks generally use more of your allowance," the help article says.
Instant is exempt. "Instant chat will not draw from your usage."
Limits still apply. At the limit, users can wait for a reset, upgrade or buy credits for eligible features.
OpenAI says most users won't need to wait for a reset more often than they do now. What it has not published is the size of the allowance itself.
This Fundamentally Changes How You Budget AI
When chat and coding drew from separate buckets, a heavy coding week did not touch anyone's chat. In a shared pool, a long agent run and a quick research question compete for the same tokens.
The wrong reaction is to buy bigger plans and hope the limits never bite. The better alternative is to know, by workload, which model you are using, how many tokens a task consumes and what that task is worth.
Seats tell you who has AI. Tokens tell you what AI is actually costing you.
Know Your Inference (KYI) is built for exactly this shift. When you understand what each AI workload costs to run, how fast it performs and how accurate it is on your own cases, you can route simple work to lighter models, save premium capacity for tasks that need it, and stop being surprised by the meter.
The organizations moving first will not just buy more AI capacity. They will know what every token is buying them.
Journey Towards AGI
Research and advisory firm guiding on the journey to Artificial General Intelligence
Know Your Inference Maximising GenAI impact on performance and Efficiency. | Model Context Protocol Connect with us, and get end-to-end guidance on AI implementation. |
Your opinion matters!
Hope you loved reading our piece of newsletter as much as we had fun writing it.
Share your experience and feedback with us below 'cause we take your critique very critically.
How's your experience? |
Thank you for reading
-Shen & Towards AGI team