The Scientist

OpenAI launched GPT-6 Astra claiming 99.9% on ARC-AGI-3. ARC Prize measured 63% on a standard harness, François Chollet 66%, and @mhmazur 62.7%. The gap came from a provider adapter that preserves opaque reasoning state, plus native compaction — so the largest lever in this release sits in infrastructure you own. Latent.Space adds the statistical reason this matters: separating a 97% from a 99% pass@1 at 80% power needs roughly 800 independently graded tasks per arm.

FrontierHarness ran coding-agent harnesses and found pass rates clustering tightly while cost varied up to 17x, with Codex highest at 66.7%. Databricks traced seven silent MCP tool bugs to an estimated $499K in wasted tokens and $1.2M in annual lost productivity, then fixed them in about an hour once tool calls were traced. Your agent bill is a tracing problem before it is a model choice. Caveat: only Codex's pass rate was disclosed, so no cost/accuracy frontier can be drawn from the published numbers.

CISA added BerriAI LiteLLM (CVE-2026-59822), JFrog Artifactory (CVE-2026-82329), Kestra OSS (CVE-2026-49869) and Starlette (CVE-2026-48710) to its actively-exploited catalog — the gateway, registry, orchestrator and serving framework of a standard ML stack. SANS separately reports litellm SSTI-to-RCE at CVSS 9.8 and three NLTK pickle and JVM-injection RCEs. Patching the gateway without rotating every provider key it holds is a half-fix, because an authentication bypass there is a credential dump.

Ai2 ran 100 open-weight models across 16 benchmarks and found BBQ and WMDP scores track general reasoning more closely than safety. A release gate built on those numbers is partly a capability gate, so a smarter model clears the bar without being better aligned. Meta's Reels team took the opposite route: it goals its ranking models on a human 5-star/1-star panel across four declared dimensions, read over months, because roughly 1,000 concurrent experiments make per-test attribution untrustworthy.

Three model-agnostic techniques shipped with measured deltas. Declarative Attention cut attended tokens during decoding 52.0% on Gemma-4-31B and 31.1% on Qwen-3.6-27B across 15 tasks. Trace-as-State moved GraphWalks Parents from 29.2% to 81.8% on DeepSeek V4 Pro Preview and 66.4% to 100% on GLM-5.2. SPACE action chunking cut LLM decision rounds up to 78.9%. All three attack the same bottlenecks Astra's adapter exploits, on models you already control.