1. Long‑context windows are a double‑edged sword
“The 1 M window has been holding up pretty well until about 700k+. Sometimes it would continue to do okay past that, sometimes it started getting a bit dumb around there.” – a_e_k
“I’ve been using the 1 M context at work through our enterprise plan … it seems to have been holding up pretty well until about 700k+.” – a_e_k
“The 1 M context has been a welcome addition, but the per‑task cost goes up while the time‑to‑correct‑output drops significantly.” – gregharned
Users praise the ability to avoid mid‑task compaction, but many note that coherence degrades after ~600‑700 k tokens and that compaction can erase useful information.
2. Workflow engineering is key to making long context useful
“I start with a PRD, ask for a step‑by‑step plan, and just execute on each step at a time.” – frannky
“I keep a CLAUDE.md file in my project root with key decisions and context.” – nvardakas
“Plan mode is great, but the plan file names are randomly generated, so it can delete the plan without asking.” – hedora
Successful users rely on explicit planning, markdown logs, sub‑agents, and frequent context resets to keep the model focused and avoid “dumb” loops.
3. Pricing and token‑budget anxiety dominate conversations
“I’ve been poking at it today, and it definitely changes my workflow – I feel like a full three or four hour parallel coding session with subagents is now generally fitting into a single master session.” – vessenes
“The 1 M context will be great for this, but it’s expensive – 500‑1000 $ for a session.” – tudelo
“I’m paying $200/mo for the most expensive plan, but I still get charged extra for 1 M context usage.” – LoganDark
Users are torn between the promise of larger windows and the reality of higher per‑token costs, especially when compaction or tool calls inflate the input size.
4. Model differences matter – Claude vs Codex vs Gemini
“Codex continues working great post‑compaction since 5.2.” – furyofantares
“Gemini gets real real bad when you get far into the context – it gets into loops, forgets how to call tools, etc.” – girvo
“Opus 4.6 is nuts. Everything I throw at it works.” – frannky
Users compare the same prompt across models, noting that Claude’s long‑context performance is still uneven, while Codex and Gemini often suffer from hallucinations or tool‑forgetting.
5. Real‑world use cases reveal both promise and limits
“I’ve had Opus fix a Rust B‑rep CSG classification pipeline successfully over the course of a week, unsupervised.” – virtualritz
“I’m building a game with Opus – it can write a test harness, but it also creates a test suite for an existing tool.” – sarchertech
“I’m using it for refactoring, code reviews, and generating feature ideas, but I still have to step in for architectural decisions.” – avereveard
While many developers can generate large amounts of code quickly, they still need to review, correct, and sometimes rewrite the output, especially for complex or safety‑critical tasks.
6. Community sentiment oscillates between hype and skepticism
“I’m not sure if AI will replace human developers, but it’s already changing how we work.” – popcorncowboy
“I think we’re just getting better at using LLMs across the board, not that the models themselves are getting better.” – fbrncci
“The shift is that 1 M context makes ‘load the whole codebase once, run many agents’ viable, whereas before you were constantly re‑chunking.” – gregharned
Debate centers on whether the technology is truly transformative or simply a productivity boost that still requires human oversight.