This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
Anthropic pushed Claude Code v2.1.257 on September 1, 2026 at 17:53 UTC. The release has two stories stapled together: Claude Fable 5.1 is now the default Fable model (1M context, $10/$50 per Mtok, $0.25/Mtok cache reads), and the bottom half is a long security sweep — a Containment-Escape auto-mode rule, a one-time prompt before outside-reads, plugin symlink escape blocked, MCP allow/deny list after-startup bypass closed, OAuth credential cleanup gap fixed, worktree isolation over-refusal relaxed, Bash deny-rule redirect bypass closed, and multi-cloud auth-header hardening on Bedrock / Mantle / Vertex / WIF / Foundry. Six distinct sandbox/auth/escape fixes in one release.
OpenAI shipped an Aug 29 changelog entry putting mTLS and X.509 workload identity federation into general availability on the OpenAI API. Both ride on the same per-project trust-anchor machinery; both are gated by RBAC (`api.mtls.read` / `api.mtls.write`). mTLS does not replace API keys — it gates them. X.509 WIF does not (yet) work in Codex. The path forward for regulated agent deployments is now production-real, not preview.
OpenAI documents the Assistants sunset on August 26, 2026 and four transcription-model shutdowns on February 26, 2027. Inventory affected workflows, preserve state and test the output capabilities the successors haven't established as equivalent.
OpenClaw pre-release 2026.9.1-beta.1 (Aug 28, 2026) ships six operational fixes: Gateway restart recovery (preserve admitted turns across restarts), Gateway config-write reliability (committed writes stay pending through watcher handoff), Worker recovery (re-arm admission deadlines, terminalize dead-worker turns), Linux Node 24 LTS installer fix, Control UI file safety, and Codex managed runtime bump to v0.150.1. Beta is a preview of stable, not a production recommendation.
OpenAI Codex v0.151.0 (Aug 28, 2026) ships a configurable MCP server grace period, extensions that can inspect or replace MCP tool results before they reach the model (the missing hook for PII redaction, audit logging, rate limiting, and format conversion), and per-repository plugin catalog combination. Plus a stack of sandbox and Guardian fixes.
Cursor Cloud Agents now boot into a blank canvas, land work in a Cursor-managed Origin repo, port-forward a live browser preview, and publish to Vercel in one click. The agent-to-production loop that previously required GitHub, a hosting service, and a deployment tool is now closed inside one product. GitHub is optional for the first time in any mainstream AI coding tool.
Claude Code 2.1.251 (Aug 28, 2026) ships PreModelSwitch/PostModelSwitch hooks, foreground subagent live streaming, and a spend limit bar. It also fixes eight distinct security bugs in the same release: TOCTOU path-traversal via in-place symlink swap, plugin command path-traversal, project-settings telemetry bypass, Workflow scriptPath escape, Grep/Glob Read-deny bypass, ANTHROPIC_CUSTOM_HEADERS approval for credential/routing/org headers, server-managed sandbox/credential/proxy guard, and sandboxed Bash output isolation. Read both halves of this release.
Qwen3.8-Flash-Next dropped as Qwen4 architectural preview: 125B/6B MoE with Gated DeltaNet + QSA, native multimodal, 1/9 training compute of Qwen3.7-Plus, open weights, runs in 78GB.
Apple shipped the M5 Ultra Mac Studio with 512GB unified memory on Tuesday. 4.3x the AI compute of M3 Ultra. 9.8x M1 Ultra. 120+ tokens per second on DeepSeek-V4-Flash. This changes the local AI agent deployment economics permanently.
Z.ai dropped GLM-5.3-Flash on Tuesday: 320B total / 18B active MoE, the first open-source frontier to ship hybrid sparse-plus-linear attention, native multimodal in the GLM-5 series, and Artificial Analysis Intelligence Index 57 at roughly one-tenth the cost of the previous tier. Here is what changed, what it beats, and what I'm migrating before my next billing cycle.
Eleven stacks. Same 4,200 agent calls per day. Two weeks. The models crossed a threshold in July; the inference stacks did not all cross with them. Here is what survived, what broke, and the hardware math nobody publishes.
Alibaba flipped Wan 3.0 to GA on August 24 with two features nobody else has shipped: 30-second one-take video generation and the industry's first document-to-video input. Here's why that combo changes the workflow math.
Six stories: GitHub had another bad week and Cursor shipped Origin to compete with it, Anthropic's watermark on Claude writing drew the sharpest critique of the year, GPT-5.6 Sol went 50% cheaper on OpenRouter and is now the best vision model OpenAI ships, Google stopped pushing some Android source tags and GrapheneOS called it out, an AI-generated GitHub Copilot "Autofix" patch let attackers compromise Snowflake's Jira, and the hottest opinion piece of the week argues models are getting dumber on purpose.
I spent last week pointing the same 4,200 agent runs per day at eight production runtimes — AWS Bedrock AgentCore, Cloudflare OS, Cursor Composer 2, the Linux Foundation Agent Stack, Stripe-routed OpenRouter, Claude Code Auto Mode, OpenAI Agent Mode + Operator, and Modal Sandboxes. The agent runtime wars are not coming. They are already here. Here is the score card, the layered architecture I am shipping Monday, and the trap nobody wants to talk about.
Aug 16, ~11:14pm Pacific — Anthropic quietly shipped Claude Opus 5.5 (codenamed 'Mythos Refresh') on api.anthropic.com, AWS Bedrock (anthropic.mantle), and Google Vertex AI. The Intelligence Index v4.1.1 hit 58 — six points ahead of Qwen 3.8 27B Apache, one ahead of DeepSeek V4 Pro (1.7T), the first closed-weights model to crack 90 on GPQA Diamond, and +14.4 on HLE over the Apache drop. Native 1M-token context ships at 1.7× the 512K tier. The closed-weights tier isn't dying. It is sharpening.
Friday the 15th of August, 2026, Alibaba pushed Qwen 3.8 27B to Hugging Face under Apache 2.0. Dense. 27B parameters. Native vision. 262K context, extensible to one million. By Sunday morning Artificial Analysis had it at 52 on the Intelligence Index v4.1.1 — the same number as GPT-5.6 Luna max, one point behind GLM-5.2 (753B) and DeepSeek-V4-Pro (1.7T). On five benchmarks it beats Claude Opus 4.6 Max. The closed/open gap just collapsed from months to benchmark points.
I spent last week running the same 12,847 production agent traces through three eval platforms — an open-source framework, a well-funded grading SaaS, and a smaller reliability-focused vendor. Two of them graded 28% and 41% of my real failures as passes. Here is what I found, the eval stack I now run on Mr. Technology, the code I wired, and the single most important lesson about agent evals nobody is talking about.
Six stories: Gemini 3.7 Flash lands a million-token context, GPT-5.6 Sol goes Ultrafast on Cerebras, Claude Code moves to default Auto mode, a published extraction attack makes closed-API reasoning traces a real security surface, Docker enters the agent sandbox wars, and Cursor joins SpaceX (probably).
Stripe finalized its acquisition of OpenRouter for more than $7 billion over the weekend, according to Bloomberg. The number that matters isn't the price — it's the structure: a complete vertical stack (measure, route, bill, pay) with two structural contradictions buried underneath.
Voice agents are now production infrastructure, not demos. The four-model pipeline (STT + turn detection + LLM + TTS) with a 960ms p95 latency budget is the stack that actually works in 2026 — and it is not what Twilio or your chat-agent playbook wants you to build. Here is the architecture, the models, the code, and the failure modes that kill deployments.
DeepSeek shipped V4-Pro-0813 on August 12 with 1.7T parameters, 893 GB of MIT-licensed weights, DSpark speculative decoding, and benchmark parity with Opus 4.8 across agentic coding. The Pro tier is now broadly competitive with the strongest proprietary models — and it is downloadable.
Z.ai shipped GLM-5.3 today: same base as 5.2, pure post-training, +50% on the company's coding bench, SOTA on CyberGym at 84.5% (Mythos 5: 83.8%, GPT-5.6 Sol: 83.6%), ExploitBench doubled from 24% to 54%. Weights land in two weeks. The interesting story is not the model. It is what the post-training data looks like.
Three weeks after 3.6 Flash, Google ships a workhorse model that quietly leapfrogs its predecessor on coding and web-dev benchmarks — and cuts the per-token price in half. The interesting story isn't the model. It's the cadence.
Google shipped Gemini 3.7 Flash today at $0.75/M input and $3.75/M output — the same price as 3.6 Flash. Across the board it tops Claude Sonnet 5 and lands within a hair of GPT-5.6 Terra on agentic benchmarks: Terminal-Bench 2.1 at 85.8%, DeepSWE v1.1 at 65.3%, Code Arena Elo 1588, OSWorld-2.0 at 47.9%. The Flash-vs-frontier gap just closed for real.