What we're building, what broke, and what we learned automating real operations.
Mistral Large 4's weights landed on 29 October under a modified MIT licence. Internal automation is fine. The catch is who hosts it, and on what.
Deepgram is acquiring Daily, maker of the open-source Pipecat voice agent framework. What changes for businesses running voice automation, and what does not.
Akamai launched a sandbox for agent-written code and an open-model inference service. Who needs the sandbox, and what it does not fix.
OpenAI's new mid-tier model drives screens for invoice entry and portals at half Sol's price. Its own safety tests show why it needs a fence.
A config mistake and an SGLang bug sent other customers' data back as normal 200 responses. What it says about how AI pipelines get tested.
Meta's 816B open-weight Muse Spark Open buys you control, not a cheaper API. What it changes for business automation, and what the license requires.
Google opened Gemini 4 Argon to paid API users. It leads on long documents and automation tests, but price any pilot at the December rate of $4/$20.
Muse Wonder matches Opus 5.5 on price but not on business-workflow scores. The stronger Wonder Full is subscription-only. Where it might still fit.
Google announced coworker agents with their own Workspace identities, scoped access, audit trails and hard spend caps. Access is limited for now.
Haiku 5.5 cuts the sticker price 90% and jumps on agent benchmarks. Independent tests show the effort level, not the price list, sets your real cost.
Microsoft's new Hooks preview lets a workflow block an agent's tool call before it runs. Useful, but a failed hook lets the action through.
A proposed standard for how AI agents sign in to businesses, announced as Muse quietly reaches half the sites one security firm studied.
Mistral's 1.05T-parameter model undercuts Claude and GPT on price and targets business workflows. The open weights and licence are still to come.
Anthropic and OpenAI shipped mid-tier models a day apart at identical list prices. Token use, caching and effort settings decide what you actually pay.
OpenAI Dots run on GPT-6 Astra, work while you are offline and take tasks in Slack and Teams. Where they fit, and where you should not rely on them yet.
OpenAI cut Sol to $2/$10 and Luna to $0.10/$0.50 per million tokens. Cheaper agents, but Sol still fails two in three AutomationBench tasks.
xAI's Grok 4.7 keeps Grok 4.6's price but trails Claude Fable 5.1 and GPT-6 Astra on agentic coding. Where a cheaper model fits, and where not.
AIR disclosed a zero-click RCE that bypasses plugin SHA pinning in four AI coding agents. Two are patched, two are not. Here is what to do.
Perplexity's new local AI agent for Windows keeps data on-device, but it only runs on RTX GPUs with 24GB+ VRAM - hardware most businesses don't own.
A new full-duplex voice model cuts turn-taking latency in half and unbundles voice from reasoning cost — here is what it means for phone-based automation.
OpenAI packaged its Codex agent runtime into an API anyone can call. The infrastructure is now a checkbox - the billing math is not.
Salesforce shipped seven named AI agents with real jobs, not chat windows. Six are live now - here is what the packaging shift and pricing maze mean for buyers.
US agencies named DeepSeek, Alibaba, Moonshot and three others for large-scale distillation. What it changes if your automation runs on their models.
Meta's Muse sends emails, books travel and makes payments for one person. Why that makes it the wrong tool for a business, and what to take from it anyway.
GPT-6 Astra is built to run a computer on its own, doing real desktop tasks in about 40 minutes instead of 75. What it changes for automation, and what it does not.
Anthropic's new top coding model keeps its $10/$50 base pricing but cuts cache reads 75%. That mostly helps one kind of automation: long-running agents.
Google's fourth Flash model in four months does near-frontier agent work cheaply. The catch is cost per task, an expiring intro price, and the treadmill itself.
OpenAI's Jalapeno chip beat Nvidia's GB300 on throughput and latency in its first benchmarks. Here's what that does, and doesn't, mean for your AI costs.
Nvidia has reportedly agreed to buy Hugging Face, the default home for open-weight models, for about $12.9 billion. What the concentration means for your pipeline.
Salesforce made Claude its default model across Agentforce and shipped a 37-skill Claude plugin. What Claudeforce changes for CRM automation, and the lock-in risk.
Z.ai released GLM-5.3-Flash on August 26: open weights, 1M context, image and video input, priced near $0.10 per million tokens. Where it fits, where it won't.
Anthropic merged Claude's memory across chat and Cowork on Aug 25 - but it ships off by default for the Team and Enterprise plans most businesses run on.
Thomson Reuters launched a proprietary LLM trained on its own legal and tax data. The lesson for smaller businesses is about owning domain data, not the $40M price tag.
Salesforce moved agentic coding out of solo terminals into shared Slack channels with a mandatory approval gate before production pushes.
Twin1 exited stealth with $20M and live deployments at Linklaters and Dechert — a preview of AI agents tied to one employee, not a chat window.
OpenAI's Ultrafast tier runs GPT-5.6 Sol up to 14x faster on Cerebras chips. No price or ship date yet, but here's what it means for real-time agents.
DeepSeek's flagship coding model went GA at $0.87 per million output tokens. Three days later, peak pricing hit $3.96. Here's what that means for anyone budgeting around it.
Google's new workhorse model is half the price and nearly twice as reliable on agent tasks — and still fails 7 of 10 real business workflows.
Grok Bot skips API integrations and drives your apps like a human would. The catch: every bot on your account shares one login pool.
A $1.2B funding round comes with real customer numbers: 70% autonomous resolution, 28,000 hours automated monthly. Here is what it signals.
In a UK government cyber test, Mythos 5 built fake identities to social-engineer a real maintainer. Here's what it means for the agents you've already deployed.
OpenAI slashed Luna and Terra pricing as Anthropic and Google also cut costs. Cheaper tokens do not mean cheaper finished tasks.
Intro pricing ends Aug 31. A new tokenizer means the real cost increase compounds well past 50% for many workloads.
MIT found 95% of generative AI pilots show no measurable ROI. The cause is almost never the model — it is what happens to the pilot after launch.
Hours saved is the easiest number to measure and one of the weakest reasons to automate. A better test, and three cases where we say don't.
Automation you cannot inspect is just a black box with better marketing.
Want this kind of system in your business? Book a free scoping call.