TL;DR
Anton Morgunov argues that 'vibe coding'—heavy agentic, prompt-driven code generation—made Cursor economically unsustainable and prompted token-saving engineering that undermined complex coding workflows. He says long-context chat in Google AI Studio with Gemini 2.5 Pro solved the problem for his work.
What happened
After a period when large language models were dominated by a single clear leader, the landscape became fragmented and workflows diverged. Morgunov describes 'vibe coding'—repeated conversational edits that cause message chains and long-token inferences—as highly token-inefficient and expensive. He says Cursor responded by engineering context-sparing techniques (line-windowing, pattern search) to cut token usage, which worked well for small, surgical edits but often omitted semantically relevant code for larger changes. That trade-off, he argues, rendered Cursor less useful for full-time software engineers. To continue larger refactors, he switched to Google AI Studio with Gemini 2.5 Pro, which he reports handled very large code dumps and multi-file refactors reliably, and which he continues to use because it preserves wide context. He also references OpenCode/Claude Code-style tooling and pricing tiers for Claude Pro that influence which models are practical for developers.
Why it matters
- Tool design choices can force a trade-off between cost control and semantic understanding of large codebases.
- Developers relying on LLMs for non-trivial refactors need wide-context support; context-limiting optimizations can break those workflows.
- Pricing and model access shape which developer-facing AI tools remain practical for sustained engineering work.
- How vendors prioritize common short edits versus complex cross-file reasoning will affect adoption of LLM-based IDE assistants.
Key facts
- The author identifies two LLM-in-loop coding use cases: simple surgical edits and complex semantic changes across files.
- Vibe coding (iteratively prompting to change small things) multiplies token usage because each prompt includes prior messages and outputs.
- Cursor reportedly adopted context-saving strategies such as limiting lines read or using pattern search approaches similar to ripgrep to reduce tokens.
- Those context-limiting strategies help cheap, isolated edits but can omit semantically relevant code needed for complex changes.
- The author switched to Google AI Studio with Gemini 2.5 Pro and reports successful handling of very large code contexts (examples cited around 80k–120k tokens).
- Gemini 2.5 Pro is noted in the piece as free in AI Studio with the caveat that Google may use conversations for training.
- The author cites models claiming 1M–2M token windows (Sonnet 4/4.5, Grok 4 Fast) and references an external bench reporting different effective context performances.
- According to the cited bench in the piece, pass@1 scores were reported as: Sonnet 4.5 ~30%, Grok 4 Fast ~45%, Sonnet 4 ~60%, Gemini 2.5 Pro ~90%, GPT-5 High ~80% (reported by the author).
- OpenCode is compared to Claude Code: the latter shows code changes in git-diff format; the author recommends OpenCode combined with Sonnet/Opus models for many coding tasks.
- The author mentions Claude Pro at $20/month includes Claude Code and can be used to authorize OpenCode; higher-paid tiers ($100/$200) permit more frequent Opus usage (as reported).
What to watch next
- Whether Cursor revises its context strategy or product positioning in response to complaints about tunnel-vision optimizations: not confirmed in the source.
- Further empirical comparisons of long-context model performance in real-world multi-file refactors beyond the cited bench: not confirmed in the source.
- Commercial uptake of long-window models (1M+ token claims) inside IDE-style tooling and their real effective context limits: not confirmed in the source.
Quick glossary
- Vibe coding: An informal term for iterative, conversational prompting that asks an LLM to generate or modify code repeatedly, often for UI and small refinements.
- Agentic workflow: A setup in which an LLM or set of tools performs multi-step actions autonomously or semi-autonomously, often reading and writing project files.
- Context window (or context length): The amount of input (tokens) a language model can consider at once; larger windows let models reason over more code or text in a single prompt.
- Ripgrep (rg): A fast command-line search tool used to find patterns in files; referenced as a technique to locate relevant code fragments to include in LLM context.
- Git diff: A representation of changes between file versions; some coding assistants show suggested edits in diff format to make modifications explicit.
Reader FAQ
Why did the author stop using Cursor?
He says Cursor's token-optimization choices reduced useful context for complex, cross-file engineering tasks, undermining his workflow.
Is Gemini 2.5 Pro the best model for large refactors?
The author reports strong long-context performance from Gemini 2.5 Pro in AI Studio for his work, but broader conclusions are not confirmed in the source.
Do models actually support 1M+ or 2M token windows?
Some models claim those windows; the piece notes effective context lengths can be much smaller and cites a bench with differing pass@1 results.
How much does Claude Pro cost and what does it include?
The author reports Claude Pro at $20/month includes Claude Code and can authorize use of OpenCode; higher-priced plans ($100/$200) are said to allow more Opus usage.
Back to all posts Vibe coding killed Cursor January 1, 2026 · Anton Morgunov For almost 1.5 years after the release of ChatGPT, the question of which LLM should you…
Sources
- Vibe Coding Killed Cursor
- Together we can fix this pressing issue! #aicoding #ai #cursor
- Cursor lacks spending caps, researchers warn
- I Literally Just Said Vibe Coding Was Gone… Then Cursor …
Related posts
- OpenAI plans audio-first hardware and upgraded ChatGPT voice models
- Tesla Annual Sales Drop 9% as BYD Overtakes It as Global EV Leader
- Satya Nadella starts a public blog to push a new AI ‘theory of the mind’