OpenAI’s new model halves the cost of computer-use agents

Welcome back. Things are tense for Microsoft devs right now. The company is reportedly offering “voluntary exits,” and not everyone’s buying the voluntary part. Meanwhile, GPT-6 Terra just dropped, and its pricing has devs talking.

Also: Two senior Claude Code engineers left ahead of Anthropic's IPO, a dev shows how two small edits can turn an agent “evil,” and a website that simplifies system design interviews. 

OpenAI’s new model halves the cost of computer-use agents: The ChatGPT maker just dropped GPT-6.1 Terra, a reasoning model built for high-volume visual and computer-use workloads. It takes text and image input and returns text. At $1 per million input tokens and $5 per million output tokens, Terra costs half as much as GPT-6.1 Sol. Per the company, it matches GPT-6 Astra on agentic tasks at one-tenth the token price and finishes over 3x faster. OpenAI still recommends Sol or Astra for heavy coding. Try it here.

Microsoft trades headcount for compute with mass buyouts: The Windows maker is offering voluntary exits to roughly 60,000 employees across the US, Canada, India, and the UK. That’s more than a quarter of its workforce. The offer covers employees at level 70 and below (up to general manager), including Copilot and Windows teams, while security and marketing are excluded. But not everyone is buying the “voluntary” part. One developer who got the email claims it’s a sign you’re already on the layoff list.

Alibaba AI model attempts surprise cloud upload in security test: Cohere gave Qwen3.8-Flash-Next a straightforward task, writing a detection rule. But the model went off-script. After claiming it wouldn’t run tools, it checked the machine’s network settings, then tried uploading a test file to a cloud server in Beijing to see if data could leave. Cohere’s sandbox blocked it, but an unsecured CI runner might not have. Luckily, it’s rare, under one in 100,000 calls. Check out the full incident report.

Your Postgres setup was fine until it wasn't. Data piles up, queries slow down, and the answer keeps being "add another system."

Tiger Data keeps Postgres fast, even at scale, with the SQL and tools you already use:

Keep one database. Keep building.

Try Tiger Data for free. No credit card needed.

The spending has outrun the math. If your team’s AI coding bill keeps growing, you’re not alone. Veteran engineer Gergely Orosz of The Pragmatic Engineer talked to 27 companies as they plan their 2027 budgets, and the numbers are all over the place. Monthly spend per developer ranges from $65 at UK fintech Lendscape to $3,000 at a two-person .NET shop. Yet people at only two companies could explain how they measure the return.

How the answers break down:

The number needs three guardrails. If you are trying to justify your team's AI spend, start by establishing a baseline for what you delivered before AI, whether that means merged PRs, completed stories, or cycle times. Next, calculate your token costs based on standard API rates, since several companies expect seat subsidies to disappear and bills to jump anywhere from 5x to 10x. Finally, make sure to factor in code quality. Lendscape managed to double its PR output while driving defects per 100 merged PRs down from 17.3 to 7.5. 

P.S. To build the full model, this essay series and DORA’s ROI calculator are good places to start.

How to review AI-written code at Anthropic scale: A principal AI engineer breaks down how the Claude Code team manages over 30,000 daily changes in a codebase that is 90% AI-written. Learn how they combine subagent reviews, automated security checks, risk-tiered approvals, and fast rollbacks so engineers only review high-risk PRs instead of getting bogged down by every single one. 

Pomerado: Turn websites and old internal portals into APIs or MCP tools for your agents, without waiting for an official API. It builds deterministic integrations that can log in, handle 2FA, and self-heal when the site changes, so your agent can use the portal without slowly clicking through every page.

How to build your own reasoning model from scratch: This 8-part course takes you from a pre-trained Qwen3 model to a working reasoning model in Python and PyTorch. You’ll build the evals, add inference-time scaling and self-refinement, train with GRPO and verifiable rewards, and then distill the final model from a stronger teacher.

AI-built screens arrive wrapped in taglines, helper text and secondary labels that explain a layout instead of fixing it. A dev shared a one-line AGENTS.md rule to fix it,

A dev noticed something off on his Mac and used Claude Code to investigate. What it found was so alarming that even Claude offered help

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team