Permanent copy of issue 38. Today's issue is always the front page.
OpenAI said on Sunday that its research organisation now uses 3.1 agent-workdays for every human workday, a mark it calls an automated research intern. Chief scientist Jakub Pachocki wrote the same day that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed. OpenVDN released open weights for a MiniMax H3 video model that still wants a rack of B200s to beat real time.
Date: 2026-09-07
Updated: 2026-09-07 06:20
Status: LIVE
Broadcast
OpenAI published internal token bills for its own researchers. The median spent more than $600 a day on inference at API prices by mid-August; the 90th percentile spent more than $7,000. After a July security halt, Astra-class GPU allocation fell 59 percent in a week, then other model classes absorbed most of the slack. Pachocki asked for voluntary slowdowns until shared safety bars exist. OpenVDN put VDN-H3 weights and training code on Hugging Face.
Watch
Stories
OpenAI said on Sunday that its research organisation now uses 3.1 agent-workdays for every human workday, a threshold it calls an automated research intern. The median researcher was spending more than $600 a day on inference at API prices by mid-August; the 90th percentile used more than $7,000 a day.
Cluster inference at flagship list rates is now the daily tool of the lab that trains the next model.
2026-09-06 / Help Net Security / chip / central
OpenAI chief scientist Jakub Pachocki wrote on Sunday that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed. He said he expects voluntary slowdowns until shared safety bars exist, three days after the company shipped GPT-6 Astra.
The same lab that is burning tokens on agents is asking the industry to ease off. The compute has not left the cluster.
2026-09-06 / The Next Web / chip / mixed
OpenVDN on Saturday released VDN-MiniMax-H3, a hybrid-attention derivative of MiniMax H3, with weights, training code, and an inference stack. The team says the 8-step model generates a 14.4-second 768p clip in 11.23 seconds on eight B200 GPUs, or about 51 seconds on a single B200.
The weights are on the hub. Real-time still wants a rack of B200s. Desk silicon can run it slowly.
2026-09-06 / Hugging Face / efficiency / local
Axis
local [####################----------] central
Score: 65
clusters still winning
up 1
OpenAI is metering intern-scale agent work at API list rates inside the lab. Pachocki asked other labs to slow down. Open video weights still need dense GPUs for real-time. The spend still sits in the cluster.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
83 | central | OpenAI said the median researcher was spending more than $600 a day on inference at API prices by mid-August, and the 90th percentile more than $7,000. That is cluster tokens, billed at list. | ||
| ENERGY |
|
86 | central | No new grid filing overnight. The intern-scale token burn still runs on the same dense racks Flex is buying power electronics to feed. | ||
| EFFICIENCY |
|
44 | local | OpenVDN released VDN-MiniMax-H3 weights and training code. Real-time 768p still wants eight B200s; a single B200 takes about 51 seconds for a 14-second clip. |
Line charts (0=local 100=central 2026-08-22 to 2026-09-07)
Market overlay (2026-08-09 to 2026-09-07)
Latest: BTC $79,357 / hashrate ~939.45 EH/s / token markers avg $5.3033/M / axis 65. These lines are different units. We compare timing, not dollars-to-hashrate.
Not shown: GPU cloud $/hour, hub electricity, miner hashprice, or lab cost per token -- none of those are a clean public feed.
Power and list prices
BTC $79,357 / hashrate ~939 EH/s. List-price gap (not a subsidy %): 44.4x list gap (blended) GPT-5.6 Luna -> Claude Fable 5 ($0.45/M to $20.00/M). (mempool.space + Coinbase/CoinGecko)
Marker prices (USD per 1M tokens; verified 2026-08-19; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)
| # | Model | Org | CC | Weights | Tier | In | Out | Blend |
| 1 | Claude Sonnet 5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| 2 | GPT-5.5 | OpenAI | US | closed undisclosed |
flagship | $5.00 | $30.00 | $11.25 |
| 3 | DeepSeek V4 Flash | DeepSeek | CN | open ~284B total / ~13B active (MoE) |
commodity | $0.44 | $1.32 | $0.66 |
If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n
Previous issue | This week | Next issue | archive
home | rss | archive | models | data | board | mcp | llms.txt