Cursor without OpenAI, Fable 5.2, Gemini 4, AlphaProof Code and /swarm

OpenAI models in Cursor: the date was pushed back
https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/
On November 12, 2026, the contract under which OpenAI supplied its models to Cursor was supposed to end. That same day, OpenAI added a short update to its post: access was “temporarily extended,” and no new date was given. Cursor publicly announced nothing, and OpenAI’s help page still describes the old date as proposed. So the shutdown did not happen, but there is still no final clarity. The important thing is that this is not about the API. The OpenAI API works as before; the models there are available for separate payment. Only the contract for supplying models inside Cursor is ending — the part included in a Cursor subscription.

Claude Fable 5.2 — the new top
https://www.anthropic.com/claude-fable-and-mythos-5-2
On November 9, Anthropic released Fable 5.2, which many rankings place first. Complex coding, multi-day agent sessions, architecture, science and finance. Anthropic’s own advice: Opus 5.5 for everyday tasks, and Fable 5.2 when the task is truly hard. It checks its own results, coordinates a huge number of subagents and can work unsupervised for hours and days. In Anthropic’s demo, the model took a prototype operating system from scratch to a working state.

What is very nice is that access was available immediately: Claude Pro, Max, Ultra, Team and Enterprise subscriptions, Claude Code, Claude API, AWS Bedrock, Google Cloud and Microsoft Foundry. There is a separate fast mode. On Hacker News, people note that on ordinary coding the difference from Opus is barely visible, while it costs like a premium model.

Google’s lineup: Carbon, Argon, Hydrogen and Decision
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/
https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
Google now has three Gemini 4 tiers, a separate decision model, AlphaProof Code for formal proofs and /swarm mode in Antigravity. All three Gemini 4 models have a large context window. Antigravity gained Orchestrator: Carbon/Argon plans and digs into the hard parts, Hydrogen does edits, tests and parallel tasks; switching happens automatically within a single session. Orchestrator with Carbon is available on Ultra, with Argon on Pro.

Carbon — the flagship for the deepest reasoning: complex architecture, root-cause bug hunting, science, security. It thinks longer, spends more tokens and is slower; for simple tasks it is overkill. For a long time they could not give people access — as soon as Fable 5.2 came out, they opened it immediately, but only on the Ultra subscription plan.

Argon — the middle model, the main workhorse. Google recommends it for most complex engineering and knowledge tasks. Available on Pro and Ultra.

Hydrogen — fast and cheap for everyday coding and agents; effectively the Flash replacement. It is adapted from Carbon, has four thinking levels and works well both standalone and as an executor or verifier under Carbon or Argon. Available on all tiers including the free one, but in Antigravity only on Pro and Ultra.

Gemini Decision 1 — not a chat model. It does not write text; in a single pass it scores options and returns probabilities: yes/no, picking from a list, or rubric scoring. This is for routing, policy checks, tool selection and sorting. It is very fast and cheap because it generates no output tokens.

AlphaProof Code: AI proves the security of real code
https://deepmind.google/blog/introducing-alphaproof-code/
On November 10, Google DeepMind showed AlphaProof Code — an agentic framework for formal verification. Humans and agents together describe what exactly needs to be proven, and agents on Gemini Carbon write mathematical proofs in Lean.

They showed the libsignal case — the code of the Signal protocol: the proofs reached the compiled machine code rather than stopping at the design level. Over weeks, the agents generated a huge volume of Lean proofs, leaving only a small portion for human re-review.

How Hacker News reacts
https://news.ycombinator.com/item?id=50442303
https://news.ycombinator.com/item?id=50471431
The biggest complaint is not the model itself but access. For ordinary users, Carbon was initially opened only on Ultra; different “Pro” plans have different models; in the app some see Carbon, some do not. Some people simply do not understand how to pay or where to select the model they need.

In the Hydrogen thread, the main topic is the awkward names. Hydrogen, Argon and Carbon do not intuitively line up as “bigger — smaller”; people — myself included — confuse which is stronger. The models are competitive, but Google still loses not on LLM quality but on confusion in names, tiers, quotas and access, plus the unfinished Antigravity.

https://www.youtube.com/watch?v=Sc6TriRAGnI

Carbon independently edited a long video: it removed pauses and failed takes, added graphics, charts, chapters and even a subscribe button, and it navigated the visual content rather than just the transcript.

The /swarm command in the Antigravity App and CLI (Pro and Ultra, still preview) turns the current chat into a coordinator: it splits the task into non-conflicting parts (for example, a separate file or package) and creates named workers. For a large task it really works, but it is easy to burn through quota.

Carbon + /swarm https://antigravity.google/docs/swarm/ reverse-engineered Factorio: the agent understood the game’s logic, built its own harness and played it; the swarm coordinated resource gathering, production, research and railways, while a separate long run spent hours progressing toward launching a rocket.

The author launched many cheap agents in the background — and anti-abuse temporarily blocked the account; support later restored it.

#openai #cursor #anthropic #claude #google #gemini #alphaproof #antigravity