AI Prompt
BlogAbout
← Back

GPT-6.1 Terra, Cursor's Agent SDK, and More

4 min read

Last week's GPT-6.1 Terra sighting is now an official launch. OpenAI released a cheaper model built for fast computer use: it reads the screen, acts across apps, and is pitched for all-day agent work. That moves it from a partner-config leak to a product they want people running continuously.

The pitch is practical more than flashy. A mid-tier model that can watch the screen and drive apps is the kind of thing builders will actually leave on all day, if the cost holds. OpenAI announcement.

What Else Happened

OpenAI says its models produced three research results in math, CS, and physics. They include a possible refutation of the Exponential Time Hypothesis via a new k-SAT algorithm, a proof of the local smoothing conjecture, and an existence proof for quantum field theories like QCD. Math and CS details are on GitHub with Lean proofs, and outside experts are verifying them. Meanwhile, more than 1,000 mathematicians have joined a new Association for Human Mathematics that commits to human-only work. (Dan Roberts' announcement, X story on the AHM)

Cursor shipped its Agent SDK. You define agents as code — prompts in Markdown, tools in TypeScript — get built-in evals, and deploy on Cursor's hosted infra or your own. Cursor also added DeepSeek V4.1 Flash, which it says sits on the CursorBench cost frontier at $0.60 to $0.98 per task. (Agent SDK post, DeepSeek V4.1 Flash in Cursor)

Former GitHub CEO Thomas Dohmke launched Entire, a Git forge built for agents and humans working together. It swaps classic pull requests for "Trails" with AI scoring on security, quality, and risk, and keeps agent history in the repo. It's reportedly backed by $60M. (Dohmke's launch post)

New lab Mirendil released Fern, a from-scratch model with a 10M-token context window. That's about half a million lines of code or days of agent traces in one pass, aimed at AI R&D chores like auditing training data and debugging runs. (Mirendil's announcement)

Gemini 4 Argon is rolling out to Antigravity IDE users over the next day. It's been climbing coding and research benchmarks, though some developers are reporting false policy-violation warnings. (Rody Davis on the rollout, X story on benchmarks and warnings)

Anthropic formally filed its S-1 for an IPO. The earlier news was about timing; this is the filing itself. Separately, Dario Amodei talked up value-based pricing, from pennies for routine help to tens of millions for drug-design breakthroughs. (Anthropic's S-1 post, X story on Amodei's pricing comments)

Launches and Demos

Accounts Worth Following

  • Mirendil (@mirendil) – a brand-new research lab shipping from-scratch long-context models, worth tracking from day one.
  • David Im (@davidim) – CEO of Sume Labs, who posts sharp teardowns of AI video pricing and builds cheaper alternatives.
  • Emily (@IamEmily2050) – hands-on tests of new models, with Three.js builds and image-pipeline experiments you can learn from.

What I'm Reading

  • Daryl Ginn on tools – kicked off a big debate about designers moving from Figma to Claude and Cursor, while many still keep Figma for team alignment
  • X story on the designer tool shift – broader conversation around that Figma-to-agent move
  • Addy Osmani on subagents – before adding another subagent, check whether a bloated system prompt is the real problem; skills and tool search often fix it

That's it for today. 👋