The Whistle: MTurk Shuts Off, OpenAI Goes Agentic, and the Cloud Giants Learn to Share (a Little)
Welcome to a new recurring feature. Each week I’ll round up the announcements that actually mattered across AWS, Google Cloud, open source, and frontier AI – with opinions attached, because news is cheap and judgment isn’t.
AWS: the week Mechanical Turk died
This one barely made a ripple in the tech press, and it shouldn’t have: as of September 29, Amazon Mechanical Turk reached end of support and is no longer available.
MTurk is the most quietly influential service AWS ever shipped. “Artificial artificial intelligence” – asking humans to do the work machines couldn’t handle yet – was the pipeline that fed a generation of ML models before scraping became free, and it kept an entire gig economy afloat on pennies and HITs. Nothing has fully replaced it. I have old scripts somewhere that still reference it, and I have no idea what to do with them now. The crowd-work era got its official end date and almost nobody wrote an obituary.
Meanwhile, AWS keeps bolting AI onto its biggest billing surfaces:
CloudWatch Omni launched September 17 as “collaborative AI-powered observability”: auto-discovered topology, natural-language queries, and an AI-guided investigation mode powered by DevOps Agent, plus a second flavor purpose-built for gen AI and agentic workloads with built-in evaluators. I’ll believe topology auto-discovery the moment it doesn’t also auto-discover a new line item in my bill. The SRE job isn’t disappearing – it’s getting renamed, and the rename comes with an inference charge.
EventBridge got enhanced custom event buses September 24: a single event bus shared across every account in your organization, with ordering guarantees and a simplified subscriber resource. Every enterprise event-driven migration I’ve lived through hit the same wall eventually – cross-account buses and ordering discipline. “Improved economics at scale” is doing a lot of load-bearing work in that announcement; check the per-publish price before you re-plan anyone’s bus architecture.
T8i instances hit GA on custom sixth-generation Granite Rapids Xeons, AWS-only, claiming up to 30% better price-performance than T3 for the burstable crowd. Custom Intel silicon exclusive to AWS is a rare sentence; a decade ago it was exclusively the other way around.
Google Cloud: an open-source agent that moves you off AWS
The item I found most interesting this week comes from Google: GKE Agentic Migration, announced September 24 and in public preview a day later, is an open-source agent plugin that translates AWS EKS infrastructure and Kubernetes manifests into GKE landing zones – delivered as reviewable pull requests. Under the hood it’s a bundle of agent skills plus a local MCP server, so instead of throwing one giant prompt at an LLM and hoping it reasons correctly about IAM, ingress, and autoscaling all at once, the guardrails are deterministic and the output is a diff a human approves.
This is the shape migration tooling should take: not a black-box “migrate” button, but an agent with a ledger of what it verified, emitting PRs I can actually review. It’s also, not coincidentally, Google putting an open-source weapon in the knife fight for AWS estates at a moment of maximum cloud-budget sensitivity. If a company’s spending this much on Kubernetes, someone in a budget review is already holding an EKS line item with a highlighter. The MCP server is where I’d be looking twice – that’s the load-bearing piece, and it’s free to poke at.
The timing alongside AWS’s own agentic push isn’t subtle. Neither vendor’s announcement mentioned the other’s, obviously, but if you’re architecting on either platform your migration question just changed from “someday” to “at what price.”
Frontier AI: DevDay goes agentic, and Anthropic ships a discount
OpenAI’s DevDay (September 29, San Francisco) landed 20+ launches, and the keynote had a theme you could smell from the parking lot: agents that don’t wait for you to type.
Dots is the headline: always-on agents powered by Astra that “take important work off your plate,” available on Pro and Business Premium, off by default for Enterprise and Edu until an admin flips the switch. It’s OpenAI’s answer to Meta’s Muse, and the consumer-agent land grab is now officially a three-plus-frontier war. What nobody ships alongside these things is the bill cap for the always-on part.
GPT-6.1 Sol brings near-Astra performance on agentic coding and computer use at one-fifth of Astra’s standard input and output prices. The race-to-the-bottom on price isn’t just a DeepSeek hobby anymore.
ChatGPT Space is a persistent workspace where your team, ChatGPT, and your Dots share context – Slack, Notion, and Drive with agents embedded. Codex runs in the cloud now, startable from a phone, with shared team environments. There’s a Decisions API for handing narrow, bounded choices (approve/reject, pick one of N) to a model in real time. And a new Pro 500 tier claims 25x the Plus usage allowance with an Ultrafast tier at up to 300 tokens/second.
Here’s the tension, though: Axios reported that per the New York Times, OpenAI will not release its newest flagship, GPT-6.1 Astra, because of security concerns – while simultaneously announcing Ultrafast access to Astra for Pro 500 and Enterprise customers. I get the distinction (API tiers of the same model with safety gates) but it’s a thin membrane. “The frontier model is too dangerous to ship as-is” and “here are twenty products built on it” is already a strange annual report, and this is the third straight OpenAI release cycle with a variation of it.
Anthropic, for balance, shipped Claude Sonnet 5.5 on September 28: 30%+ faster than Sonnet 5 and up to 30% less cost per task, at the same $2/$10 per-million token price (Reuters confirmed – the model simply needs fewer tokens for the same work). That’s a real bill cut, not a marketing cut. Haiku 5.5 arrives “in the coming weeks” for the high-volume tiers. It’s the cheaper complement to Opus 5.5 from a week earlier, and worth noting Anthropic is reportedly building toward an IPO; shipping efficiency gains on schedule is what a pre-IPO lab’s product calendar looks like.
And the safety thread keeps thickening. Amodei’s essay this month called for the industry to pace the frontier as recursive self-improvement kicks in; Altman and Hassabis publicly endorsed, TechCrunch confirmed the three labs have been quietly talking for weeks about an industry standards body, Altman endorsed letting independent verification organizations into the company, and somehow the conversation still has an antitrust problem and a White House calling the risk discussion a hoax. Meanwhile Axios reported OpenAI and Anthropic are jointly probing tens of thousands of security incidents at the same time everyone is shipping always-on consumer agents. I don’t know what the correct regulatory theory is here. I know “the labs are coordinating safety” and “the labs are coordinating” getting confused in court later is a very 2027 problem.
Open source: the funding talk got less naive
September 24’s Google open-source roundup surfaced some of the more honest writing on OSS money in a while. Laurie Voss’ “Nobody pays for open source. We can force them to.” argues voluntary charity and license-change attempts have both failed, and the fix is boring infrastructure: commercial package registries and mirror vendors paying automatic royalties down the dependency tree to the maintainers, funded by companies already spending $1B+/year on OSS (currently flowing to mirror and security vendors, not the humans writing the code).
That’s the first OSS funding pitch in years I didn’t immediately shrug at. Not charity, not a new license, but a metered payout through the exact middlemen already taking a cut of everything. The failure mode is every registry slowly becoming Stripe-with-a-tollbooth, but I’d take that over the current model of maintaining critical infrastructure in spare evenings while a trillion-dollar economy runs on it. An LF Energy study the same week made the parallel case for utility grids (2-5x net value over proprietary procurement), and separately, reporting on AI’s influx into open source quantified the strain as AI-generated code and vulnerability floods hit maintainers. We’re automating PRs about twice as fast as we’re funding reviewers.
That’s the week. If something in here was wrong or you’ve got an opinion stronger than mine, info@stackref.com always works.
This post was created with a Hermes AI agent (hermesagent.ai) from a weekly news gathering workflow, then reviewed and approved by Keith before publication.