Google Research has released Harness², an open-source framework that lets AI agents refine their own operating 'harness'—prompts, tools, memory and workflow—without changing the underlying model's weights. The system works through two feedback loops: one that tweaks the harness for the task at hand by comparing different execution attempts, and another that keeps a persistent record of lessons learned to guide future tasks. Tests on GLM-5 and Gemini 3.5 Flash showed performance gains of up to 11.5 points on one benchmark and 11.8 percentage points on another after a single refinement step.
github.com
· 2026-10-19
Neuralwatt CEO Chad Gibson analyzed a month of cloud traffic and found that different agent harnesses produce wildly different prompt-cache hit rates, from 1.6% to 11.1% uncached input tokens. The study found that a cache miss roughly doubles a request's energy use, and that deep sessions past 128k tokens with major prompt re-reads can cost as much compute as roughly 600 normal turns at that depth.
neuralwatt.com
· 2026-10-16
Ultramarine 1 Flash is a new model priced at $0.13 per task on internal benchmark PI Knowledge Work, which the company says correlates strongly with the Artificial Analysis Intelligence Index. It is built on GLM-5.3 Flash and generates 6.5% fewer tokens than that base model while outscoring Claude Fable 5 (max) and GPT-5.6 Sol (max), which cost $8.19 and $2.09 per task respectively. Pricing on the Ultramarine API is $0.15 per million input tokens, $0.03 for cached input, and $0.50 for output.
paperinstruments.com
· 2026-10-15
A new study reports that AI models in the 9B-30B parameter range can already perform many individual steps needed for autonomous replication across consumer IoT hardware like smart ovens and budget Android TV boxes, though they still fail at chaining these steps together reliably. The researchers say a 9B model successfully ran inference on 8 of 9 tested hardware types, and that stronger models like GLM-5.2 can already complete full end-to-end replication in simplified tests.
anthropic.com
· 2026-10-12
A new open-source benchmark called the Telegraph Test shows that instructing large language models to answer in 'cablese' — the clipped, article-free style once used by telegraph operators — reduces billed output tokens by 40-49% on several models' own API meters, while preserving or even slightly improving factual recall. The effect held across four model families tested with roughly 1,300 questions over 50 passages, with one exception: gpt-5-mini's reasoning overhead made cablese prompting roughly double its cost instead of cutting it.
fiveminutesforward.com
· 2026-10-07
A development team set out to run an entire month of coding work on the GLM 5.3 Flash model, keeping within a $68 budget for the first two weeks. The second half of the month saw usage spike to 1 billion tokens across other models, driven by an expensive vibe-coded Wagtail MCP server prototype that alone cost $150 and 5kWh, plus infrastructure capacity issues that forced switches to DeepSeek V4.1 Flash and Qwen 3.8 Flash.
wagtail.org
· 2026-10-02
Anthropic published benchmark results claiming Zhipu AI's open-weight GLM-5.3 model can autonomously generate working cyberattack exploits, including chained exploits, with safeguards that can reportedly be bypassed. In its Exploitbench tests, GLM-5.3 produced 50 successful Chrome exploits out of 410 attempts, close to the 56 achieved by Anthropic's own unreleased Claude Mythos model, while rivals Kimi K3 and DeepSeek V4.1 Flash scored 0% on a related control-flow hijack test.
tomshardware.com
· 2026-09-30
Anthropic reports that GLM-5.3, a new model from China's Zhipu AI (Z.ai), can autonomously build end-to-end cyber exploits similar to Anthropic's own Claude Mythos Preview. Anthropic says simple jailbreak techniques bypassed GLM-5.3's safeguards between 64% and 100% of the time in simulated tests, whereas the same attacks failed against safeguarded Claude models. NIST's Center for AI Standards and Innovation separately assessed GLM-5.3 as the most cyber-capable open-weight model released so far, trailing the US frontier by about four months.
anthropic.com
· 2026-09-29
A technical post describes a method for turning an off-the-shelf large language model, GLM-5.3-Flash running on Privatemode, into a decision-making model that outputs a typed choice with confidence scores in a single forward pass, rather than generating a full JSON response. Benchmarked against public datasets, the setup reportedly matches TypeSafe's specialized Jev model in decision accuracy and speed, while also supporting typed decisions on images, which Jev cannot do.
privatemode.ai
· 2026-09-26
An analysis argues that the cost of running machine learning models is dropping by orders of magnitude annually, driven by GPU efficiency gains that double roughly every two years—a pace not seen since early Moore's Law. The piece distinguishes proprietary models like GPT-6 Astra from open-weight models such as GLM-5.3-flash, noting that hosted and locally-run versions improve at different rates, with per-token pricing for frontier models not falling as consistently as costs for smaller models.
jyn.dev
· 2026-09-23
Z.ai's newly released GLM 5.3-flash is an open-weight AI model cheap enough to run on roughly $5,000-$15,000 of hardware, and third parties have already stripped its safety refusals via 'abliteration,' producing versions that comply with requests for hacking and other harmful content. The author frames this as a deadline pressuring security initiatives like Project Glasswing and Daybreak, which use fast frontier AI to find and patch software vulnerabilities before dangerous capabilities become widely accessible.
jyn.dev
· 2026-09-08
A developer ran the same prompt—building a self-contained sci-fi hangar scene with Three.js, including hovering drones, animated lights, and camera paths—across 10 combinations of AI models (including GLM, Luna, SOL, Astra, and Qwen variants) and coding harnesses like Codex, OMP, OpenCode, and DSH. The test tracked metrics such as completion time, token usage, tool calls, error rates, and whether the model verified its own output by opening the file in a browser and checking screenshots.
alvins82.github.io
· 2026-09-08
GoKawiil