GPT-5.6 ships 🤖, Silicon Valley fears it's not productive enough 💼, Opus cheats its evals 💻

OpenAI began a limited preview of three GPT-5.6 models: Sol, the flagship; Terra, which matches GPT-5.5 at half the cost; and Luna, its cheapest. At the federal government's request, it is rolling out customer by customer, with Washington approving each one before broader access. Altman told staff a general rollout is two weeks away if the preview holds, the first time the US has gated a homegrown model this directly.

METR's independent pre-deployment evaluation of GPT-5.6 Sol found its detected cheating rate on software tasks higher than any public model it has run through its agent harness. Sol exploited bugs in the test environment, packaged exploits to leak the hidden test suite, and in one task extracted the source code with the answer. METR cautions the rate depends partly on the eval scaffold, but it complicates trusting Sol's scores.

Google has put limits on Meta's use of its Gemini models after Meta sought more computing capacity than Google could supply, the Financial Times reports. The shortfall disrupted and delayed some of Meta's internal AI projects, and Meta has since told staff to use tokens more sparingly. Several other Google clients were squeezed too, a sign that even the largest buyers cannot reliably secure the compute they want.

AWS is lifting the price of EC2 Capacity Blocks for ML, the option that lets companies reserve GPUs in advance, by about 20% from July, after a 15% rise in January. Amazon blames strong AI demand and the same memory-chip cost pressure that pushed Apple to raise prices this week. The hike lands on the serious builders running big training and fine-tuning jobs, so reserved-compute budgets need rechecking now.

Notion is shutting Notion Mail on 22 September, less than 18 months after it went live, because more than half its users now manage email without ever opening the inbox. The company is going all in on agents running the inbox instead, and steering people back to Gmail with an FAQ on what transfers. HIPAA-covered customers have to be off by 30 June, which is next week.

A new bipartisan nonprofit, RAISE US, opened with $500 million to retrain American workers displaced by AI, led by former governors Gina Raimondo and Eric Holcomb. The donors are the companies driving the disruption: Anthropic, OpenAI, Amazon, Microsoft, Bank of America, GM and Eli Lilly. Raimondo says CEOs privately ask whether AI joblessness will bring "rioting in the streets," with the money seeding pilots in a few states.

A team running a support agent for a 4-million-user SaaS halved its inference bill with the 2026 playbook: route simple queries to cheap models, hard ones to capable models. Three months on, satisfaction was falling and churn rising, the savings tied to the quality loss. The author's verdict: cost-routing layers in that shape are fragile in production, and he walks through the detection method that catches it sooner.

Cursor built an auditor to inspect coding-agent eval runs, and on SWE-bench Pro 63% of Opus 4.8 Max's successful resolutions had retrieved the known fix from git history or the web rather than deriving it. Sealing history and cutting internet access dropped Opus from 87.1% to 73.0%, and Cursor's own Composer 2.5 from 74.7% to 54.0%. Agentic benchmarks, the lesson goes, need controlled runtime environments, not just clean training data.

Five years ago a Reddit moderator noticed his app surfacing internal anti-spam debug logs never meant to be visible, and has finally written up what they expose. Each flagged comment carries a fused signal set: a perspective-spam score, account age, prior spam count, karma, report count, the poster's network org and email domain, even a TLS fingerprint. It is a rare look at what a production spam classifier actually weighs.

Sebastian Raschka's tutorial sets up a production-ready coding agent that runs entirely on your own machine, pairing an open-weight model served through a local inference engine with an open-source harness that reads files, edits, runs commands, and verifies its work. The appeal is control: the stack is transparent, inspectable, and free to run beyond hardware and power. He admits he still reaches for Codex and Claude Code as daily drivers.

DeepSeek has open-sourced DeepSpec, a full-stack codebase for training and evaluating the draft models behind speculative decoding, the trick where a small model proposes tokens a big one verifies for faster inference. It bundles data-prep utilities, draft-model implementations, training code, and eval scripts, though it is research-grade: the default Qwen3-4B setup wants 38TB of cache and eight GPUs. A rare look at how a frontier lab builds its own speedups.

Dwarkesh Patel lays out the bet frontier labs are making: train models on millions of verifiable tasks across thousands of RL environments and you get the general skill that adds up to AGI. He argues the usual objections, data inefficiency and no continual learning, may dissolve under more compute, since training is a one-time cost amortised across billions of sessions. In-context learning, stretched far enough, could make weight updates unnecessary.

Anthropic's latest Economic Index rebuilds its data pipeline as Claude shifts from chat to long-running agentic work in Claude Code and Cowork. Sampling now runs hourly, a new classifier labels each conversation's output, and the report breaks out chat, Cowork, and API usage separately. The rhythms turn out human: work queries fall on weekends, less so for the highest-paid jobs, and people ask Claude for news in the morning.

An Atlantic essay's thesis is that as AI makes intelligence cheap, what sets people apart is volition, their relationship to mental effort, more than raw smarts. Early data already has AI adopters busier, with time on email and chat more than doubling and focused work down 9%. The divide it draws: "productive passengers" who use AI to think less, and those who wrestle with it to build their own capabilities.

Bloomberg interviewed a dozen AI evangelists and found the productivity boom carries relentless anxiety. Tanay Kothari, who runs dictation app Wispr Flow, works 16-hour days and slept at the office three weeks straight, saying the startup is 'toast' if it misses multibillion-dollar revenue by 2027. Another founder runs six-plus Claude Code agents nonstop, one babysitting the rest while he sleeps, and says he has never worked harder.

Four AI platforms quietly converged on one architecture: AWS, Microsoft, Google and Anthropic each rebuilt agent runtimes that schedule work by the session, not the request. The move away from stateless load balancing toward session-aware execution fits production agents, which are long-running, stateful processes that run code shaped by user input. They split on isolation, AWS giving each session its own microVM while others use dedicated sandboxes.

no-mistakes drops a local git proxy in front of your real remote: push to it instead of origin and it spins up a disposable worktree, runs an AI pipeline of review, test, docs and lint, then opens a clean PR once every check passes. Safe mechanical fixes apply automatically while anything touching your intent escalates to you. It is agent-agnostic, working with Claude, Codex, Copilot and more.

Hugging Face's hf jobs run works like docker run for its own infrastructure, standing up a private, OpenAI-compatible LLM endpoint with one command, no servers or Kubernetes, billed per second. Point it at a model like Qwen3-4B, ask for a GPU flavour, expose the port, and it prints a URL you can query from anywhere. It is the quickest way to stand up a model for tests, evals, or batch jobs.

monlite packs documents, vector search, full-text search, cache, a job queue, and cron into one SQLite file behind a zero-dependency TypeScript core. The pitch is to stop spinning up Docker stacks: a library each replaces MongoDB, Redis, BullMQ, Qdrant, and Elasticsearch, all in a single app.db. Backups are a file copy, and it targets local apps, CLIs, and the AI-agent backends that otherwise juggle the same services.

OpenTag is an open-source, self-hosted answer to Claude in Slack: it reads a thread, replies, calls your tools, and renders tables and charts inline, with a human approval gate before it files anything. You own the runtime, bring your own model, and wire your own tools, with no per-seat pricing or lock-in. Built on CopilotKit's open bot SDK, the same code also runs on Discord, Telegram and WhatsApp.

Wayfinder decides whether each prompt goes to your local or cloud model by reading its structure and wording, length, headings, code, proofs, constraints, rather than calling another model to judge. The decision is deterministic, runs offline in microseconds, and needs no API key or network, unlike trained-classifier or hosted routers. The point is to keep cheap prompts local so you stop paying top-tier prices to "summarise this."