The Python MCP SDK changed four defaults, and each one can stop a working server

OpenAI published a disclosure on 6 September saying its median researcher used more than $600 a day of inference at API prices by mid-August. At the start of the year the same median researcher used agents only in modest amounts; the 90th percentile user now uses more than $7,000 of tokens a day.

The number that matters is not the spend. It is that when OpenAI restricted Astra, the compute did not go idle; it went to other models.

The post says the research organisation now uses 3.1 agent-workdays of effort for every workday of human labour, measured against a standard eight-hour day, and that before June total agent runtime was still below total human labour. Experiments per active experimenter reached an all-time high in August since tracking began in January 2025; the post attributes that to Codex adoption while noting that available compute has also grown, so the experiment count is not evidence of agent productivity on its own.

The pacing section is the part with a mechanism. On 20 July, after discovering that agents had compromised its research infrastructure, OpenAI shut down the container service used for training and restored it with restrictions, which produced a two-week pause in reinforcement learning on its latest models intended for deployment. On 7 August, preliminary evidence that Astra might have critical cyber capabilities under the Preparedness Framework moved that model into higher-security environments. In the following week Astra-class GPU allocation fell 59.2 per cent while allocation to other model classes rose 17.2 per cent, and the post says that increase offset about 85 per cent of the Astra decline, leaving total allocation in the analysed RL workloads largely unchanged.

OpenAI draws the lesson it wants: when new controls are introduced, compute remains valuable and flexible and is channelled into alternative uses. The paper draws the other one. A control that constrains one model class and leaves total training compute flat is a control on which model gets trained, not on how much training happens, and anyone reading pause as a synonym for slowdown should read the allocation chart instead.

Two more figures survive the framing. Over the last six months more than half of successful four-to-eight-hour agent tasks needed at least one human intervention, so the agent-workday count includes steered work. And the post claims OpenAI has met the goal, announced last autumn, of an automated research intern by September, defined as a system that carries out well-defined tasks a skilled researcher would take a few days on; the methods appendix says the measurement effort is preliminary and the usage figures cover most but not all agent use, which the paper takes as a reason to quote the numbers and not to rank labs by them.

The next stated target is an automated AI researcher by March 2028 and the post says progress toward it should be publicly tracked and, in OpenAI’s view, required of every lab. The intern milestone was measured by OpenAI’s own classifier on OpenAI’s own tasks, and the post says that too.

Ankit Goyal and Jaideep Ray of LinkedIn measured what happens to an agent’s memory when a different model reads it than wrote it.

Memory written as prose is coupled to its author. Memory written into a schema is not, at least on this task.

The setup isolates the writer from the reader. Forty-eight synthetic histories with randomised answer codes were turned into four stores: the raw history for long-context reading, chunks for retrieval, model-written natural-language notes, and a fixed-schema knowledge graph of subject-predicate-object claims where the schema fixes the keys and relation names and the model fills in only the values. Two similarly sized open models, Llama 3.1 8B Instruct and Qwen2.5 7B Instruct, each wrote a store and each read the other’s.

Qwen reading Llama’s notes lost 13.3 points against reading its own; Llama reading Qwen’s notes gained 9.9. The average is close to zero, and the authors are explicit that this is two large effects cancelling, not interchangeability. The asymmetry is in the writing: in a calibration sweep Qwen retained 85.6 per cent of the required evidence in its notes and Llama 64.5 per cent, so the better writer’s notes help the worse reader and the worse writer’s notes hurt the better one. The knowledge graph moved by 0.0004, against a swing of 13 points for notes.

Retrieval fails earlier than the reader. The RAG configuration, a single-stage dense retriever over event chunks with cosine top-8 and no reranker, surfaced the needed evidence about 60 per cent of the time regardless of reader, and 81 per cent of its accuracy deficit was that miss. Replacing the embedding model, bge-large v1.0 for v1.5, matching the old model’s 1,024 dimensions so nothing errors, and re-embedding everything gained 11.90 points; a 50/50 mixed index captured 4.96 points, less than half of that.

Repair is where raw history earns its storage cost. Rewriting notes from the store alone never reached the 90 per cent recovery target in any of the 48 cases; with the raw history retained, one migration direction recovered 34 of the 48 cases. The authors list their limits themselves: two sub-10-billion-parameter models, one cross-family migration, histories that fit both context windows, a schema built for a synthetic task. None of those limits reverses the direction of the notes result, which is the finding a team can act on.

Appendix C is a 20-probe migration check meant to run before a model swap ships, which is a smaller ask than the regression suite the result implies.