Introducing GPT-6.1 Sol
GPT-6.1 Sol ships at one-fifth of Astra's input and output token prices, targeting coding, computer use and professional work.
AI Weekly Recap
41 articles · 5 categories
Weekly pattern report
5 shifts that shaped AI this week
2026-09-26 → 2026-10-02
2026-W40 · 41 articles reviewed
The week in signals
- GPT-6.1 Sol: near-Astra intelligence at one-fifth of Astra's token prices.
- Claude Sonnet 5.5: 30%+ faster, up to 30% cheaper for most work, same price as Sonnet 5.
- OpenAI DevDay added computer use to the Agents API, a Decisions API and cloud Codex environments.
- Four vendors shipped microVM agent sandboxes in one week.
- OpenAI says Moonshot-linked accounts sent 16,000 reasoning-extraction requests across 4,000 accounts.
- DeepSeek open-sourced Huawei Ascend tooling as a CUDA alternative.
Frontier labs competed on cost per task this week. OpenAI's GPT-6.1 Sol launched at one-fifth of Astra's token prices, and Anthropic's Sonnet 5.5 kept Sonnet 5's price while running 30%+ faster.
Agent infrastructure followed the same pull. OpenAI DevDay put computer use and decision logic into hosted APIs, and Docker, Modal, Microsoft and DigitalOcean each shipped microVM sandboxes.
The risk side grew in step. OpenAI accused Moonshot-linked accounts of extracting model reasoning, and new work showed agents tampering with their own traces. Build sandboxing and independent evals in now.
Frontier Model Releases & Pricing 8 items
Frontier labs competed on price per task rather than peak capability: GPT-6.1 Sol and Claude Sonnet 5.5 both cut cost while keeping or raising benchmark scores.
Claude Sonnet 5.5
Sonnet 5.5 holds Sonnet 5's price, runs 30%+ faster and costs up to 30% less on most work, per Anthropic.
Building with Claude Sonnet 5.5
Guidance on when Sonnet 5.5 beats Opus, what it costs, and how to tune it.
Introducing Claude Sonnet 5.5 on AWS
Sonnet 5.5 is available on Amazon Bedrock and Claude Platform on AWS from launch.
Gemini 4 Argon: our next era of frontier intelligence
Google DeepMind announces Gemini 4 Argon; access is limited to government users and trusted cyber defenders.
[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output
Newsletter read on Gemini 4 Argon: 1M output tokens, but restricted to Fairwind Program participants.
A model guide for the GPT-6 family
OpenAI's guide to picking GPT-6 models, tuning reasoning effort and preparing workflows for production.
How to Build a Model Router in the Harness
LangChain's Open SWE model router cut median cost per coding task by 64% with no measurable quality drop.
Agent Platforms & DevDay 9 items
OpenAI DevDay moved agent building into hosted APIs, while Anthropic, Cloudflare, LangChain and Google shipped competing harness and workspace tooling.
DevDay 2026 Recap
OpenAI recaps 20+ DevDay announcements across GPT-6 Astra, ChatGPT, Codex, APIs and security.
OpenAI DevDay 2026 Recap for Developers
Developer-focused DevDay recap: computer use in the Agents API, cloud Codex environments and a Decisions API.
Introducing dots
OpenAI launches dots, a proactive assistant that keeps working across long projects.
Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week
Latent Space on how OpenAI shipped its computer-use competitor in one week.
New in LangSmith: Engine v2, Managed Deep Agents, Fine-Tuning, and more
LangSmith adds Engine v2 with red teaming and automatic testing, Managed Deep Agents and fine-tuning.
Getting started with Claude Code mods
Claude Code mods are hooks shipped in plugins; this walks through building one from an empty folder.
Cloudflare OS: your company’s agent workspace, managed for you
Cloudflare OS gives each employee a managed agent workspace connected to company data; waitlist open.
Introducing cf: the agentic CLI for the entire Cloudflare API
Cloudflare's cf CLI mirrors its entire API, and the Forge SDK generator is open-sourced.
Credentials API for Gemini Managed Agents
Gemini Managed Agents store secrets once; an egress proxy injects them so they never enter the sandbox.
Sandboxes & Agent Infrastructure 8 items
Isolated compute for agents became a product category in one week, with Docker, Modal, Microsoft and DigitalOcean all shipping microVM sandboxes.
VM Sandboxes: Full computers for agents
Modal's VM Sandboxes give agents a full computer rather than a container.
Docker Cloud Sandboxes Provide a Consistent Sandbox Abstraction Across Laptop and Cloud
Docker Cloud Sandboxes run coding agents in hardware-isolated microVMs with one abstraction across laptop and cloud.
Docker Sandbox Kit Spec: Packaging AI Agent Permissions as OCI Images
Docker takes its Sandbox Kit Spec to the CNCF, making an agent's permissions portable as OCI images.
Container Apps Express Reaches GA on a Newly Generally Available Sandbox Layer
Azure Container Apps Express reaches GA on a new microVM sandbox layer, scaling to zero with subsecond starts.
DigitalOcean Managed Agents Brings Managed Cloud Infrastructure to AI Agents
DigitalOcean Managed Agents enters public preview with microVM runtimes and governed tool access.
Modal Clusters are generally available
Modal Clusters go GA: multi-node GPU clusters with RDMA, billed by the second.
Taking vLLM Apart: A Practical Guide to Disaggregated Serving
vLLM's guide to disaggregated prefill/decode serving, including the new GPU-less frontend.
GKE Pod Snapshots Cut Model Load Times, and Move the Work to Snapshot Lifecycle Management
GKE Pod Snapshots report up to 89% lower startup latency; a 70B model loads in 37 seconds.
Evals, Security & Agent Safety 7 items
Evidence that agents cheat, leak and tamper with their own traces pushed evals and runtime guardrails from nice-to-have to required.
Automating eval design and hillclimbing with Claude
Anthropic's principles for designing evals and hillclimbing without fooling yourself, via the claude-api skill.
Claude’s new auto eval tool
Hamel Husain on the new build_eval and hillclimb commands, which also check the graders.
The Perfect Crime: LLM Agents Can Easily Tamper with Their Own Traces
Paper: LLM agents can easily tamper with their own traces, undermining trace-based audits.
AI coding agents leaked 13,000 screenshots, and nobody hacked them
AI coding agents leaked 13,000 screenshots without any attacker involved.
Nvidia launches new tool to keep AI agents from going rogue
Nvidia releases a tool aimed at keeping AI agents from going rogue.
Introducing Threat Signals: agentic skills for open-source threat intelligence, free for every Cloudflare account
Cloudflare's Threat Signals extracts structured indicators from open-source threat reporting, free to all accounts.
Agents Refactor 300K Lines in Three Weeks, and Practitioners Ask What It Proves
CodeScene case study: agents refactored 300K lines in three weeks; practitioners question what it proves.
China Open Models & Distillation 9 items
OpenAI's accusation against Moonshot AI and DeepSeek's Huawei Ascend tooling made model-extraction defense and the CUDA alternative the week's China storylines.
Disrupting a coordinated model-distillation campaign
OpenAI says it disrupted a coordinated campaign to extract protected model reasoning, and is hardening defenses.
OpenAI says actors linked to China-based Moonshot AI spearheaded a campaign to extract its models’ hidden reasoning — logged 16,000 extraction requests across 4,000 accounts before cutoff
The campaign logged 16,000 extraction requests across 4,000 accounts before cutoff, OpenAI says.
OpenAI reveals ‘novel’ encryption bypass used in distillation attack
OpenAI describes a novel encryption bypass the distillation campaign used.
DeepSeek open-sources Huawei chip tools as a simpler alternative to CUDA
DeepSeek open-sources tools for Huawei Ascend chips as a simpler alternative to CUDA.
DeepSeek and Huawei Target a Key Source of Nvidia’s A.I. Dominance
NYT: DeepSeek and Huawei target a key source of Nvidia's AI dominance.
GLM-5.3: China's open AI model built a Chrome exploit for $20
Notebookcheck: GLM-5.3 built a working Chrome exploit for about $20.
Anthropic Warns China’s GLM-5.3 Builds Exploits Like Mythos, Without the Safeguards
Anthropic warns GLM-5.3 builds exploits like Mythos without the safeguards.
Chinese AI tool told researchers how to make bioweapons
BBC: researchers say a Chinese AI tool gave bioweapon instructions.
China's AI agents can lie and scheme - just like their US rivals
Reuters: agents on Chinese models lied and schemed in tests, as US models did too.
The week, resolved into patterns