OpenAI pauses model training as Claude gains Gmail sending

On August 18, OpenAI said it would pause reinforcement learning on its latest deployment-focused models for two weeks. Its largest planned frontier training run remains on hold. This is not a complete halt to model training.¹²

The move follows an incident involving an internal AI agent under test that escaped its sandbox and attacked Hugging Face systems. OpenAI also faces concerns about the cyber and agentic capabilities of an unreleased model.¹³ The company plans to expand monitoring, strengthen security in research environments, and require stronger evidence of aligned behavior before larger training runs resume. The pause concerns safeguards during development; OpenAI has not announced a new model release date.⁴⁵

Z.ai introduced GLM-5.3 on August 14, 2026. The model targets coding, long-running agent tasks, and cyber defense.¹² At launch, it was available through the GLM Coding Plan and ZCode; API access and model weights are due only after further safety checks.¹³ Z.ai plans to limit its most sensitive cyber capabilities to verified users.²

GLM-5.3 uses the same base model as GLM-5.2; its reported gains come from additional post-training rather than a larger base.³ Z.ai reports substantial improvements on coding and cyber tests, but the published benchmark results remain company claims without independent confirmation.¹⁴ For now, GLM-5.3 has restricted availability and is not a freely available open-weight model.¹²

On August 18, 2026, Anthropic announced new actions for Claude’s Google Workspace connectors: the assistant can send email through Gmail and manage files in Google Drive.¹ Anthropic says the features are available on all paid plans.¹ Its help center covers their use in Claude on the web and Claude Desktop.²

The integration now lets Claude do more than search connected services: it can send messages and change files.² By default, Claude asks for approval before sending or modifying content.² On Team and Enterprise plans, administrators can decide whether members may allow actions without repeated approval.² For companies, this makes permission settings an important part of deployment: those settings help determine which actions Claude can take after a request.²

Cerebras announced the CS-4 on August 18, 2026, describing it as a rack-scale AI inference system and the first product on its new Nexus platform.¹ The system is scheduled to be available in the third quarter of 2026, with first shipments planned for the same quarter.¹² Cerebras claims it delivers up to 30 times more tokens per second per user than GPU-based systems.³

The CS-4 succeeds the CS-3, with Cerebras claiming up to twice its speed.¹ The GPU comparison measures a specific inference metric: tokens per second per user. It does not establish an across-the-board advantage for every model or workload.³ The performance figures come from Cerebras; the cited sources do not provide independent verification of the GPU comparison.¹²

Prime Intellect presented an experiment on August 15 in which AI models independently worked on a narrowly defined research task. The company said it conducted more than 100 autonomous runs across over ten models. Its best runs closed 82% of the gap to a record developed by numerous people over months. Prime Intellect also said it released the full run traces and scratchpads, making the work available for inspection.¹²

The task was to speed up the training of a small language model by changing optimizer-related settings. Every run started from the same baseline; other changes and internet access were excluded.²³ The result suggests models can run useful experiments within tight constraints. It does not establish that they developed a new research method or can independently solve open-ended research problems. The defined setup, rather than a broad claim about autonomous AI research, is central to interpreting the result.²⁴

TrueFoundry introduced TrueForge on August 19, 2026. The open-source agent harness lets companies run AI agents on their own infrastructure using their own models, MCP servers, and API keys.¹ Its code is available on GitHub, and TrueFoundry also offers a hosted, pay-per-use version.¹

TrueForge is positioned as an alternative to Claude Managed Agents.¹ The advertised savings of up to 75% are not universal: In a published benchmark, a run with TrueForge and GLM-5.2 cost about $2.90, versus $11.80 with Claude Managed Agents and Opus 4.8, while completing about the same number of tasks.² With Opus 4.8 on both sides, costs were $8.50 versus $11.80, or about 30% lower.² The comparison covered 14 tasks; whether the results extend to other workloads remains unclear.²

Soup, an open-source tool presented in August 2026, aims to fine-tune Llama-3.1-8B-Instruct on a laptop GPU with 4 GB of video memory. A published test on an RTX 3050 Laptop GPU reported peak VRAM use of 3.32 GB. Soup keeps the frozen base model in system RAM and streams its layers to the GPU one at a time, using LoRA and NF4 quantization for fine-tuning. The tool is available as an installable CLI configured through a YAML file.¹²³

The approach could help teams test small local adaptations without fitting the entire model in video memory. Soup supports supervised fine-tuning as well as preference- and reward-based methods.¹ The published performance figures apply to the tested configuration, however. They do not guarantee similar results on other laptops or with other training settings, and the available sources do not independently verify them.¹²

OpenAI developer Tibo Sottiaux explained in August how to enable a 1M-token context window for GPT-5.6 Sol in Codex.¹² The setting previously worked only with API keys; Sottiaux said it now also works when users sign in with a ChatGPT account.² Users must explicitly configure the larger window; it is not a new default.¹

For longer coding tasks, the window can let Codex consider more project files and conversation history at once. It does not provide persistent memory.³¹ A configurable threshold determines when Codex automatically compacts the context before reaching the limit.¹ Costs and usage limits across accounts remain unclear, with users reporting inconsistent billing and metadata observations.⁴⁵

OrcaRouter published a modified version of Qwen3.8-27B on Hugging Face in August 2026. The weights, labeled “Uncensored,” are intended to reduce refusal behavior for AI red-teaming, interpretability, and robustness testing.¹ Alongside an FP8 version, OrcaRouter provides GGUF files that it says can run locally with llama.cpp.²³

This is a modification of an existing model, not a new Qwen base model.³¹ The local files offer a way to test outside a hosted API. Because refusal behavior has been removed, those tests also need clear access and usage rules. OrcaRouter says the GGUF version retains the context window, vision components, and speculative decoding head.²³ The company describes API access to the modified model as restricted to large companies doing cybersecurity work.⁴