Perplexity on X

What draws praise is range of use. One person credits a couple of days of prompting with a full Halo-style game — a project file, screenshots from real Halo games, sponsor logos turned into gear — while others describe trip planning, school worksheets and presentations, or an assistant they ask everything. A paying user says they would drop their other subscriptions before Perplexity.

The disappointed are fewer but blunt: one calls it "utter crap" and cancelled after two weeks, others ask what happened to it or say nobody uses it, and one user posted a plain goodbye.

Recent talk is about two changes. Ultrafast, shipped as a speed setting rather than a separate model pick, draws praise for long Computer tasks: less waiting between asking and getting finished work back. Running the model and agent on your own Mac is called "the version of an agent we'd trust with real files". Skeptics want the claims stress-tested past the math benchmark, worry about agents committing to main, and ask whether local holds up when the Wi-Fi drops.

Price complaints centre on gating: Ultrafast Astra and Sol are only on the Max plan, one person doubts the extra speed is worth double the price, and a Pro subscriber who got a week of use was told to wait until next year for more credits. Others describe a free Pro offer followed by a card charge, a cancellation answered with a warning of further charge attempts within 24-48 hours, credit-card details that can't be deleted, and no automatic invoice emails on enterprise plans.

On accuracy, people describe names introduced without checking their relevance, a chart drawn wrong the first time and right the second, and an answer that is usually "according to a website we didn't pay". Two say it is wrong constantly, or reframes their complaint instead of correcting the answer.

Smaller counts: fully local mode wants 36GB of memory and the app takes 20GB, one user keeps getting GPT-6.1 after choosing another model and threatens to cancel, and Ask and Discover Daily are gone.

The comparisons don't converge. One user says Claude and DeepSeek have a more reasonable chat UI, and that Google AI Studio gives more control over context by editing or removing messages. One asks how many people actually move off Bloomberg, doubting the habit breaks. One argues Perplexity is fading because OpenAI, Google and Anthropic put web search and source citations inside their own chatbots. One expects Computer's background tasks and browser automation to make paid job-search tools redundant for job hunting and networking. One user who hit usage limits says they built and run their own models without caps. And one lists Perplexity among the alternatives once Gemini limited free accounts to its Lite model.

None of these posts travelled far.

Six people ask some version of what happened to @perplexity_ai, from rarely seeing it around to "Is it dead?" and not hearing updates for months, through to one who hasn't thought about it in a while and asks whether we're still in the game; another wonders whether Perplexity has turned into an agent.

Local mode raises the design questions. Does a local task that needs a web lookup escalate to the cloud automatically, or stop and say so? Which tasks still need the cloud once the on-device model hits its limits, and how does a hybrid setup run? Two people want numbers first: how much RAM before it feels usable, and how many tokens per second it produces.

The fast models get the same test: same quality, or a lighter variant that trades accuracy for speed? Others ask how two agents committing to the same repo resolve conflicts, what actually breaks inside a 0.0002% error rate, what verification adds in latency, and what one real Computer workflow costs a month.

The routing group wants Perplexity to decide for itself whether a request runs on the local model or in the cloud — "a built in auto-routing mechanism", one person calls it, and sticky too. Another would allow local models in chat while flagging that cloud gives better answers, and suggests hosting the routing classifier on the device.

Level with them: people asking for open source and a fully local build, and people who want Perplexity to become a business platform, either as a verified-outcome platform or as "the orchestrator of the business operating system".

The single asks are concrete: local mode working on the mainstream 16GB laptops people buy, a Perplexity Code product shipped soon, faster work on Windows, PPLX Radio brought back, an append-only commit log recording which agent and session made each change with clear merge rules, and controls to check a long task before it finishes.