Promptea Weekly
Week of 2026-09-27 to 2026-10-03
The big story
Gemini 4 Argon arrives for cyber defenders first, with a 1M-token output budget and a $2/$10 intro price
Google's new frontier model is real but not yet available to developers. What you can plan around today is the price, the output ceiling and the order of the queue.
This week in AI
Gemini 4 Argon arrives for cyber defenders first, with a 1M-token output budget and a $2/$10 intro price
Google's new frontier model is real but not yet available to developers. What you can plan around today is the price, the output ceiling and the order of the queue.
Frontier pricing from Google, OpenAI and Anthropic now sits in the same $2–$4 per million input band, which shifts model selection toward per-task testing.
Read the full story →
GPT-6.1 Sol claims Astra-level coding at a fifth of the price, and OpenAI starts selling speed
OpenAI's DevDay model is a cheaper tier that it says nearly matches its flagship. The benchmarks are OpenAI's own; the prices are not in dispute.
Teams paying Astra prices for coding agents now have an official, five-times-cheaper alternative that OpenAI says performs close to it, which justifies a round of internal evals.
Read the full story →
Claude Sonnet 5.5 keeps the price, promises fewer tokens, and breaks forced tool calls
Anthropic's new mid-tier model costs the same as Sonnet 5 and claims to finish tasks faster. The catch is in the docs: five breaking API changes.
Sonnet is the default workhorse for many API and coding-agent workloads, so a same-price upgrade with claimed per-task savings affects a lot of production traffic.
Read the full story →
NVIDIA wants agent guardrails outside the agent: OpenShell goes open source, Sentry moves checks into the DPU
The Open Agent Safety Platform pairs an Apache 2.0 sandbox runtime that works with existing coding agents and a BlueField-4 watchdog that ties the strongest guarantees to NVIDIA hardware.
It moves agent safety from model behavior to infrastructure: permissions are enforced by a runtime and, optionally, by separate hardware the agent cannot reach.
Read the full story →
NVIDIA halves DGX Spark's memory to 64GB, and the entry price still lands at $4,999
The new configuration ships October 23 through six PC makers. It keeps the same chip and bandwidth, holds models up to 100B parameters by NVIDIA's count, and costs more than the original 128GB model did at launch.
It sets a new, lower-memory entry point for running open models and always-on agents locally, but at a higher price than the original 128GB Spark launched at.
Read the full story →
Ai2 opens up its MoE training stack, tested at over a trillion parameters
Olmo-core 3 is infrastructure, not a model, but it is the foundation for the next Olmo and comes with unusually candid negative results.
It opens up MoE training infrastructure at trillion-parameter scale, a layer frontier labs usually keep private.
Read the full story →