[AINews] not much happened today

OpenAI released GPT-6.1 Sol, priced at $2/$10 per million input/output tokens versus $10/$50 for Astra, according to a pricing summary. OpenAI staff described Sol as "good, cheap AND fast," and it reportedly beats GPT-6 Sol by 6.4 points on DeepSWE v1.1 and Opus 5.5 by 2.2 points on AutomationBench. Sol [Max] entered Agent Arena at #5 (+11.23%) with a $0.56 median cost per task, 39% cheaper than GPT-6 Sol while scoring 1.52 points higher, and 81% cheaper than Astra while landing within 1.04 points. A global Codex usage reset was set for Oct 2 at 10AM PT.

Sonnet 5.5 [Max] debuted at #3 (+12.5%) and ranked #1 in the Chat category, at $2.74 per task versus $1.58 for #2 Opus 5.5, which keeps it off the Pareto frontier. Anthropic models now hold the top three Agent Arena spots. Gemini 4 Argon [High] took #1 in Text Arena. MiMo-V2.6-Pro and Flash entered Agent Arena at #5 and #9 among open models. StepFun's Step 5 Preview ranks #7 among open-weight models on Vals at $2.54 per task, averages nearly two hours per task, and has a 1M-token context window.

Unconfirmed rumors include Claude Fable 5.5 for next week, said to outperform a reportedly delayed "Astra 6.1," and a "GPT-6 Astra Lite" listing speculated to be the same model as Sol.

Meta open-sourced ESP32 firmware and a Linux SDK for hardware working with Muse, and made 5,000 units of its smart-home bridge free for subscribers while supplies last. DeepSeek Harness shipped macOS and Windows desktop builds. Pi now runs on Cloudflare Durable Objects via agents SDK v0.26.0 alongside its v1.0 release, and T3 Code's orchestrator rewrite merged after passing 400K users. OpenAI's Agents API added one-call browser computer use, Bedrock Managed Agents and portable environments, claiming 99.97% turn reliability and 20% faster tool calls.

Meta released six papers on open math problems produced with Muse Spark 1.1 and 1.2 through plain meta.ai chat. Google's Gemini multi-agent system Cogentic produced new results on five open theory problems. Nathan Lambert and Tom Zick launched Trillium Labs, a non-profit for open post-training recipes and infrastructure, with initial support from Halcyon Futures and Schmidt Sciences.