Permanent copy of issue 49. Today's issue is always the front page.
OpenAI released Astra for Law on Thursday, a GPT-6 Astra configuration with a search index of more than 230 million U.S. legal URLs, and said selected firms will get it first through Trusted Access in ChatGPT. Google, Nvidia, and Anthropic joined Emerald AI in a grid alliance that says demand response could connect another 100 gigawatts of data centers. Mozilla.ai measured llama.cpp, Ollama, LM Studio, and llamafile on a Mac Studio, a Linux GPU box, and a Steam Deck, and found build flags moved speed more than the choice of server. Baseten, Hugging Face, and Goodfire said they will publish safety checks for open-weight models after Hugging Face listed more than 6,000 copies with the safeguards stripped out.
Date: 2026-09-18
Updated: 2026-09-18 06:05
Status: LIVE
Broadcast
A law firm that rents Astra for Law is still buying tokens from a cluster, not loading a model onto a partner's laptop. Google, Nvidia, and Anthropic want the grid to treat those clusters as loads that can pause, so more halls can plug in without waiting for new plants. On a Mac Studio the same GGUF files ran within a few percent of each other once the flags were set. Hugging Face can host the weights; it cannot keep 6,000 stripped copies from circulating.
Watch
Stories
OpenAI released Astra for Law on Thursday, wrapping GPT-6 Astra in a search index of more than 230 million U.S. legal URLs and offering it first to selected firms through Trusted Access in ChatGPT. SiliconANGLE reported that the setup passed 54 percent of 200 questions on Vals AI's Legal Research Bench, against 38.7 percent for GPT-6 Astra with web search alone; Harvey and Legora are named as later API customers.
-- A law firm's research still bills OpenAI's cluster, not a model that lives on the associate's laptop.
2026-09-17 / SiliconANGLE / efficiency / central / #openai-launches-astra-for-law
Google, Nvidia, and Anthropic joined Emerald AI on Thursday in the AI Energy Management Alliance, TechCrunch reported, with utilities including AES, Constellation, National Grid, and NRG Energy. The group said pausing noncritical jobs and shifting compute could let another 100 gigawatts of data centers connect to the grid.
-- A 100 gigawatt claim is more computer halls on the wires a household already pays for.
2026-09-17 / TechCrunch / energy / central / #google-nvidia-anthropic-seek-grid-capacity
Mozilla.ai on Thursday measured llama.cpp, llamafile, LM Studio, and Ollama on a Mac Studio M4, a Linux box with one Nvidia L40S, and a Steam Deck, using the same GGUF weights. The lab said configuration moved speed more than the choice of server, including a 63 percent jump in Steam Deck prompt processing from a newer Vulkan shader toolchain.
-- On a Mac Studio the same weights ran within a few percent once the flags were set.
2026-09-17 / Mozilla.ai / efficiency / local / #mozilla-measures-local-model-servers
Baseten said Thursday that its Base Labs research arm will work with Hugging Face and Goodfire AI on safety checks for open-weight models. TechCrunch reported that Hugging Face lists more than 6,000 abliterated models, copies whose safeguards have been stripped; the partners have not said how the new checks will work.
-- Six thousand stripped copies on a public hub leave the safety job to whoever runs the file.
2026-09-17 / TechCrunch / efficiency / mixed / #hugging-face-partners-on-open-model-safety
Axis
local [#####################---------] central
Score: 71
clusters still winning
up 1
OpenAI put GPT-6 Astra behind a 230 million URL legal index. Google, Nvidia, and Anthropic asked the grid for 100 gigawatts of extra room. Mozilla showed a Mac can serve the same weights as a rented GPU. Hugging Face will try to put safety back into open models that anyone can strip. The cluster still holds the training run and the power contract.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
87 | central | OpenAI wrapped GPT-6 Astra for law firms; no new desk chip shipped overnight. | ||
| ENERGY |
|
93 | central | Google, Nvidia, and Anthropic said demand response could connect another 100 gigawatts of data centers. | ||
| EFFICIENCY |
|
49 | local | Mozilla.ai found build flags, not the server brand, moved local token speed by as much as 63 percent on a Steam Deck. |
Line charts (0=local 100=central 2026-09-02 to 2026-09-18)
Market overlay (2026-08-20 to 2026-09-18)
Latest: BTC $78,101 / hashrate ~954.53 EH/s / token markers avg $5.3033/M / axis 71. 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 $78,101 / hashrate ~955 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
Subscribe
$ echo you@email.com | underclassic --subscribe
Previous issue | This week | Next issue | archive
One email per issue. No tracking pixels, no filler.
home | rss | page 2 | page 3 | archive | models | subscribe | data | board | mcp | llms.txt