Frontier signal stream

In a company engineering publication dated 25 Aug 2026, OpenAI released the first measured performance results for Jalapeño, described as "OpenAI's first custom inference chip". Tested on InferenceX, "a public benchmark from SemiAnalysis", across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, OpenAI reports Jalapeño "delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems", and 2.1 to 4.1 times higher performance on highly interactive workloads. OpenAI states it plans "to begin deploying Jalapeño within OpenAI's compute infrastructure by the end of the year" and that it is "the first generation of a multigenerational roadmap" with Gen 2 "deep in development". The chip was co-developed with Broadcom per OpenAI's earlier unveil publication. Material because it is working first-party silicon with measured third-party-benchmark results from the largest buyer of merchant AI accelerators, landing the day before Nvidia's quarterly print.

Anthropic published a study in which Claude Opus 4.8 and a Mythos Preview model autonomously ran de novo protein-binder design campaigns, producing 1,320 designs of which 354 binders were confirmed against 14 of 15 targets, independently produced and tested by Adaptyv Bio and Twist Bioscience. Anthropic reported overall hit rates of 22.6% (Opus 4.8) and 26.7% (Mythos Preview), rising to 35.1% in single-target mode, versus the 10-15% it says is typical today.

OpenAI officially named its next major model family 'Astra' via a math report stating an internal version solved ten open problems in mathematics and theoretical computer science that had seen no progress for at least a decade (spanning high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics — including establishing the existence of non-sofic groups). Each proof was formalized in Lean for machine-checkable correctness; OpenAI says the tokens for all ten solutions would cost ~$2,000 at API rates. Astra is designed to coordinate multiple agents on problems for hours or days, and is expected to be the first model submitted under the US administration's planned pre-release AI review framework.

OpenAI launched its GPT-5.6 family in three tiers (Sol for hardest reasoning/security, Terra for high-volume business, Luna for fast low-cost work), setting new state-of-the-art on TerminalBench 2.1 (Sol 91.91%). Notably the models are restricted to ~20 trusted partners initially after OpenAI shared them with the US government, following Trump's June 2 executive order mandating federal benchmarking of new AI models before wide release; all three tiers are classified 'High' risk for cyber and bio/chem capability.