Semantic review is behind discovery. Evidence and Pulse may be incomplete until the backlog is cleared.
2 material first-party/frontier/open-problem candidate(s) have waited more than 6 hours for semantic review.
EVIDENCE REGISTER
A chronological register of published signals. Source class, verification state, evidence quality and relevance remain visible before interpretation.
OpenAI reduced GPT-5.6 Sol developer pricing from $5 to $4 per million input tokens and from $30 to $20 per million output tokens, a 20% input and 33.3% output reduction, with promotional pricing available at least through November 21, 2026. This lowers the cost of sustained frontier-model workloads but does not change model capability.
Anthropic introduces CHIVE, an agentic counterfactual-evaluation pipeline for explaining naturally occurring LLM behavior. On its checkable proxy evaluation, activation oracles, natural-language autoencoders and sparse autoencoders do not improve prediction of counterfactual behavior over a transcript-only baseline. The authors stress that the proxy differs from harder system-card use cases, so the result is evidence of an interpretability/control bottleneck rather than proof that activation-reading methods are useless.
DeepSeek released an experimental multimodal version of V4 Flash on its API, adding image understanding while retaining the text model’s reasoning and agent capabilities. DeepSeek reports large gains on vision-dependent agent benchmarks and says multimodal agent performance approaches Claude Opus 4.8; independent benchmark confirmation is still pending.
Anthropic made Claude Mythos 5 available for repository vulnerability scans in Claude Security, returning structured findings and suggested fixes without exposing raw model access; it also announced partner integrations, a $35M Defender Advantage Fund, and planned expansion of trusted cyber access.
Researchers introduce three behavioral probes showing that several high-performing open-source speech-recognition models reproduce benchmark-specific reference text even when the audio contradicts, masks, or underdetermines that text. The result indicates that public benchmark scores can overstate general-purpose transcription capability and strengthens the case for held-out, temporally separated evaluations.
Google-affiliated researchers introduce EnvHarness and EnvRigger, a programmable layer that reshapes existing agent environments from observed failure trajectories while preserving the original verifier. Across five benchmarks in four domains, the paper reports up to a 9.0 percentage-point improvement on held-out instances and 9.8% fewer execution steps.
A new Microsoft-linked arXiv benchmark evaluates 507 stateful, policy-conditioned business workflows with executable checks on final backend state and side effects. The strongest tested model reaches 65.36% pass@1 but succeeds in all 20 repeated attempts on only 25.25% of tasks, showing that occasional agent success substantially overstates dependable execution.
OpenRouter released an anonymous stealth preview called Ox Alpha on August 20. Its official listing confirms a 1,048,576-token context window, text/image/video input, tool use, structured outputs and positioning for coding and sustained agentic work. Community claims that it beats leading frontier models are not independently validated, and the developer identity remains undisclosed.
Meta published new Muse Spark 1.2 evaluations and demonstrations spanning visual coding, audio-visual understanding, tool-augmented multimodal reasoning and a specialized robotics stack. The robotics setup uses Muse Spark as a high-level planner with a lower-level VLA policy for manipulation; Meta also previewed WildArtifactBench for real-world multimodal agent utility.
Anthropic moved computer use, the Skills API and the Files API to general availability and added a new browser-use tool that combines screenshots with page structure. Computer use can now execute several actions per turn; Anthropic reports a customer claims workflow falling from 32 to 13 minutes with 100% completion in its tests.
Guidelight AI Standards assessed public safety and control practices at OpenAI, Anthropic, Google, xAI and Meta; Reuters reports the best grades were only C+ and highlights insufficient preventive controls and containment across the sector. This is a governance/control bottleneck signal, not a new capability result.
OpenAI disclosed a broader operational slowdown around frontier-model work: two weeks of deployment-focused RL/model testing were paused, its largest planned frontier RL run remains on hold, stronger sandboxing and AI-based monitoring are being added, and the Preparedness Framework is being rewritten as models approach its critical cyber thresholds.
Z.ai says its upcoming open-weight GLM-5.3 reached 84.5% on CyberGym and will be released publicly in about two weeks after additional security work; the benchmark claims are not yet independently verified.
Anthropic’s August 2026 first-party Risk Report says its current models are providing significant speedups to AI research and engineering, while explicitly concluding that they do not yet fully substitute for Anthropic researchers and have not clearly doubled the overall pace of progress. The report also names internally deployed Mythos-class systems used in the assessment.
Google released Gemini 3.7 Flash on 13 August 2026 for software coding, multi-step agentic workflows and automated business tasks. Independent reporting describes meaningful gains in planning, instruction following and code generation, with lower pricing than the preceding Flash release.
DeepSeek released the GA version of V4 Pro on 13 August 2026 across app, web and API. DeepSeek reports substantial agent gains, including Terminal Bench 2.1 at 87.9, HLE at 42.7 without tools and 60.0 with tools, Toolathlon-Verified at 74.1 and DSBench-FullStack at 71.1. The API also adds native Responses API support and low/high/max thinking effort. This is a material frontier-model release, but independent evaluation is still needed before treating it as a new global capability ceiling.
SpaceXAI released Grok 4.6 on 12 August 2026. Independent reporting citing Artificial Analysis says the model improves roughly five points over Grok 4.5 on the Artificial Analysis Intelligence Index, lands around GPT-5.6 Sol overall, and is especially strong on agentic work while remaining materially cheaper than the top Anthropic models. This strengthens SpaceXAI's frontier position, but it does not currently establish a new absolute capability ceiling.
Pathway reports that BDH-CQ, a proprietary 150-million-parameter post-Transformer system, combines in-context learning with recurrent latent reasoning and reaches 29.5% pass@2 on the public ARC-AGI-1 evaluation set. The reported operating point uses about 0.85 H200 GPU-seconds per task, corresponding to a computed inference cost of $0.00070 per task at the paper's hardware-price assumption. A documented black-box audit by external-affiliation co-authors reproduced the deployed system's 29.5% score without access to model weights.
The 4 August Pulse recorded Qwen3.8-Max at 2.4T parameters and 1M context, relevance 7/10, with no ASI forecast change.