Issue 33: OpenAI Marks Astra Critical
Permanent copy of issue 33. Today's issue is always the front page.
OpenAI said Astra is the first model it has designated Critical for cybersecurity under its Preparedness Framework, and that it plans a limited release with extra safeguards. Anthropic launched Claude Fable 5.1 for general use and Claude Mythos 5.1 for vetted cyber and biology work. Hugging Face published 207 Apache-licensed WebGPU kernels so browsers can run local inference without a cluster. Clusters still take the spend.
Date: 2026-09-02
Updated: 2026-09-02 06:00
Status: LIVE
Broadcast
OpenAI said Astra meets its Critical cybersecurity bar and that it will limit the sharpest cyber tools at launch. Anthropic put Fable 5.1 on general sale and kept Mythos 5.1 behind a trust gate. Hugging Face shipped 207 WebGPU kernels for in-browser inference. Nvidia and CrowdStrike said post-trained Nemotron models can write detections cheaper than a closed frontier model.
Watch
Stories
OpenAI says Astra meets Critical cyber threshold
OpenAI said Astra is the first model it has designated at the Critical cybersecurity level under its Preparedness Framework, meaning it can find unknown flaws and build exploits across many hardened systems without step-by-step human guidance. The lab said it delayed parts of training and release to add safeguards, restarted a large reinforcement-learning run on August 28, and plans to ship Astra soon with the most advanced cyber tools limited to testers and later Daybreak Blue.
A frontier lab is now saying its next model can write working exploits on its own. Access will be gated. The capability still lives in the cluster.
2026-09-01 / OpenAI / chip / central
Anthropic launches Claude Fable 5.1
Anthropic said Claude Fable 5.1 and Claude Mythos 5.1 are the same model with different safeguards: Fable is generally available, Mythos is limited to vetted cyber and life-science programs. The company said Fable 5.1 should cost about 25 percent less than Fable 5 on typical token-billed work because cache-read prices fall, and that it can now be used to find software bugs but not to write exploits.
The new flagship is cheaper on cache reads, not a desk-side weight drop. Mythos stays behind a trust gate.
2026-09-01 / Anthropic / chip / central
Hugging Face ships WebGPU kernels for local AI
Hugging Face published @huggingface/kernels and 207 Apache-2.0 WebGPU kernels on the Hub so browsers can load versioned GPU ops without a server. The company said the kernels were 2.57 times faster by geometric mean than ONNX Runtime WebGPU on an Apple M4 across 809 matched cases, and launched Fleet so users can benchmark on their own devices.
This is a real local pole: optimized GPU math that runs in the browser, not another closed API tier.
2026-09-01 / Hugging Face / efficiency / local
Nvidia and CrowdStrike ship Nemotron cyber agents
Nvidia said it and CrowdStrike evaluated an isolated offense-defense loop that uses Nemotron open models, Falcon telemetry, and specialized agent harnesses. CrowdStrike said a Blue Solano defender based on post-trained Nemotron beat a leading closed frontier model in its own tests at 99 percent lower cost, and Nvidia reported a 2.5 times detection-rate gain for the optimized open-model pipeline in backtests.
Open weights did the specialist work. The product still sits inside CrowdStrike and Nvidia racks.
2026-09-01 / NVIDIA Developer Blog / chip, efficiency / mixed
Axis
local [##################------------] central
Score: 61
clusters still winning
up 1
OpenAI and Anthropic both shipped frontier model news. Hugging Face put optimized GPU ops in the browser. Open Nemotron is inside a CrowdStrike product. The spend still sits in the cluster.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
79 | central | OpenAI said Astra meets its Critical cyber bar. Anthropic launched Fable 5.1 and Mythos 5.1. Nvidia and CrowdStrike said they post-trained Nemotron open models for SafeMind. | ||
| ENERGY |
|
84 | central | No new grid filing on this morning's wire. The constraint is unchanged: large AI loads still sit on interconnection queues and behind-the-meter plans. | ||
| EFFICIENCY |
|
39 | local | Hugging Face published 207 WebGPU kernels and said they were 2.57x faster than ONNX Runtime WebGPU on an Apple M4 in matched cases. CrowdStrike said a Nemotron-based defender beat a closed frontier model in its own tests at far lower list cost. |
Line charts (0=local 100=central 2026-08-18 to 2026-09-02)
Market overlay (2026-08-04 to 2026-09-02)
Latest: BTC $76,594 / hashrate ~919.16 EH/s / token markers avg $5.3033/M / axis 61. 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 $76,594 / hashrate ~919 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
home | rss | archive | models | data | board | mcp | llms.txt