Cognition (Devin) on X

Firsthand users carry the section. They run Devin CLI and cloud sessions daily, describe orchestrating many sessions with a Python script through Dynamic Workflows, and argue cloud agents are where the work goes: "the future is cloud agents". One reports Devin built and submitted an iOS game in six hours; another hands heavy work from a chat client to a cloud machine and gets results back in the chat. A user likes it enough to build in public with it.

The cheaper setup is a second thread of praise: 10 people read the lead-plus-sidekick benchmark as a win, several citing 31% lower cost per task at a higher score, and expect a frontier lead with a cheap sidekick to become the standard agent topology.

Objections are smaller. 5 see no advantage over other harnesses or doubt the benchmark carries: "benchmark wins evaporate when customers bring real codebases". 4 are sour — a Max subscriber still needs a team plan to collaborate, and others lost sessions and tokens to an update. 2 like parts of it (SWE-2, cloud sessions) but downgraded a plan or prefer one part over another.

The reliability complaints are specific: the CLI crashes, updates arrive and sessions crash or lose all their information — a user wakes up to "another update and lost all my session information", watching work and tokens go down the drain — routing is "definitely broken", and one developer traces a failed local session to Devin negotiating ACP v2 while its request omits a required field. One user says they were muted in the Discord server after pointing out the CLI crashes and that SWE-2 repeatedly drains external services.

On price, people say "cost is high", that Claude models cost about five times what a Claude plan charges, that SWE-2 is "less capable and more expensive than Opus 5.5", and that a Max plan still needs a team plan before anyone can collaborate.

Limits hit paying users: a Pro subscriber kept meeting limits where unlimited usage had been claimed, and a $20 subscriber hits short-term SWE-2 rate limiting while working on two things. 2 describe losing track of long runs — session titles never update, so managing several is guesswork.

The Claude Code comparisons are mostly people choosing to stay: they prefer Opus 5.5 on a Claude Max subscription, use Sonnet 5.5 daily, and note that Sonnet at max in plain Claude Code reaches the same 68 on the chart, so the win is price, not score. Others keep Claude Code locally because cloud agents lack local files and keys, or object that Claude models cost several times a Claude plan inside Devin. One calls SWE-2 "less capable and more expensive than Opus 5.5".

The two-model idea is familiar to some: 2 people describe running their own split — Claude Code building with Codex reviewing, or Opus for the heavy work and a local model behind another client for the rest — and ask how much of the 31% comes from the sidekick rather than fewer retries.

Single voices compare elsewhere: Windsurf's analytics page is "cleaner than Devin's", FactoryAI is rated well above Cognition by one, a Droid user reports Opus with Sol reviewing is "virtually no bugs and cheaper", one waits for Kimi K3.1 Pro to be offered alongside Opus and SWE-2, and one asks where it sits against Cursor.

The limit questions are practical: whether the plan still gives unlimited SWE-2 usage, whether SWE-2 can be free forever and until when, what a 5x allowance gets you, and why desktop shows 8% usage while cloud shows 18% — "which limit is the correct ONE". One asks how far a paid plan's usage goes.

On the sidekick, people want to know what it contributes beyond the bill: "what's luna actually doing in there apart from cutting the bill?", whether the cheap model is doing the grunt work or the win is just two models running at once, whether a Fusion run with a Claude lead bills the GPT sidekick to a ChatGPT account, and who decides when the lead hands off — the lead itself or a separate router — and whether the user can see it.

Scope questions are basic: is it an agent like Claude Code or Codex or does it do more, is mobile for checking in or starting work, is it good for non-coding knowledge work, can it open a tested PR before anyone reviews. 3 ask whether a Claude or ChatGPT subscription can be used inside Devin; 2 whether the cost win holds on long multi-file refactors; 2 whether they can try it and whether it is on OpenRouter; 2 what qualifies someone for the ambassador program and deals.

The requests for evidence are specific: task-by-task results next to the average so a switcher can see where the cheaper setup struggles, run-to-run variance, and the "tool budget and prefix cache hit rate" behind the score, plus a test that seeds a mistake in the lead's brief to see whether the sidekick catches it. One wants a cost comparison against Sonnet, which "scored slightly higher on its own", as the baseline before adding a second model. Another asks what "efficient" means for approvals — tokens per merged PR or per busy-looking session — and argues for measuring accepted changes before capping.

A mobile app is asked for impatiently, with Android named: "can you pls just make an mobile android app and I'm sold". On usage, people want free or generous allowances and caps that reward people "getting real work out of it" instead of blanket limits.

Single requests: Chinese language support in the desktop app, a cloud run that flags missing files and service access before it starts rather than failing midway, and a worked end-to-end example showing how deployment context changes the priority of a security finding.