Skip to main content
SI
TheBestSI
Editorial Analysis

Public-Signal Records

Public-signal analysis is part of our evaluation and ranking methodology. It weighs online discourse, public sentiment, current trends, and user input. These records show what we collected, when, and from where. They are signals, not test results.

Read this page with our methodology. A record can raise a question or support a judgment, but it never earns the Hands-On Tested label. Online discussion is self-selected and can be manipulated, so each record lists its sources, dates, and limitations.

Claude

Sources dated 2026-09-28 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • Commenters on the Sonnet 5.5 launch welcomed its price-performance and said it is close to the larger Opus 5.5 model on agentic coding tasks, and a practitioner blogger said it is nearly as good as Opus 5.5 on some coding tasks. These are impressions and vendor-reported comparisons, not our tests. [1][5]
  • Some commenters in the outage thread described Opus 5.5 as a clear step up from earlier models. [2]
  • A practitioner reported that Opus 5.5 built a working browser game from a single prompt (one anecdote on one task). [6]

What people complain about

  • Reliability: commenters reported repeated errors during an outage and said the status page lagged what they saw. Some said outages are routine for large services and that retrying often works. [2]
  • Safety filters: several commenters on the Sonnet 5.5 thread said cyber-related safeguards flagged or downgraded them on ordinary work such as reviewing old C code, and that a verification program did not help them. [1]
  • Cost for heavy use: commenters on the employer-usage thread said token-heavy use is expensive and that companies are now setting budgets on AI tools. [3]
  • Privacy and monitoring: a thread about Anthropic reportedly passing a user's chat to police drew many comments saying people wrongly assume AI chats are private, and others who said companies should be clearer about monitoring. [4]

Where sources disagree

  • Whether safety guardrails are reasonable protection or an over-cautious 'nanny' approach that pushes users to other tools. [1]
  • Usage limits: some commenters said they hit session limits quickly while others reported no problems on similar plans. [1][2]
  • Whether using a second vendor's model to review Claude's output is well supported by research; one commenter said it is, another called it unsupported. [2]
  • Whether expensive models are worth the cost: some said the top model pays for itself, others said cheaper models handle most work and heavy use is wasteful. [3]
  • Whether and when an AI company should report a user's content to authorities; commenters split between intervention and a chilling effect. [4]

Trends

  • Anthropic's own release notes list a rapid model cadence between April and October 2026, including Opus 4.7 (16 April), Opus 4.8 (28 May), Fable 5 (9 June), Sonnet 5 (30 June), Opus 5 (24 July), Fable 5.1 (1 September), Opus 5.5 (22 September), Sonnet 5.5 (28 September) and Haiku 5.5 (7 October). [7]
  • Anthropic's release notes show repeated additions aimed at organisations: admin controls, a compliance API, skill and plugin security scanning, and self-serve HIPAA configuration for Enterprise and the API. [7]
  • Anthropic merged Cowork and chat into one experience on 16 September 2026, and says artifacts are available on every plan, including Free. [7][9]
  • Community discussion shows employers placing spending caps on AI tools, which is a cost-pressure signal, though the underlying report was not read. [3]

Expert views

  • Simon Willison (independent developer and blogger): Sonnet 5.5 is nearly as good as Opus 5.5 on some coding tasks and is the free-tier model on claude.ai, which he sees as a strong free offering. [5]

Limitations

  • Most discussion comes from Hacker News, which skews toward developers. Reddit and review platforms were not read.
  • We read long threads through a summarising fetch tool and only the first part of each; counts are approximate and we compute no sentiment scores.
  • Two practitioner posts were read through index-page summaries, not in full.
  • Several threads concern news reports we did not read, and the police-report thread concerns a private person, whom we do not name.
  • Official pages are the vendor's own statements and two carry no exact date.
  • Signals are not test results. This record does not earn the Hands-On Tested label.

Sources

  1. Hacker News: Sonnet 5.5 launch thread (community, 2026-09-28). About eight top-level comments were readable in the part of the thread we read; the thread is much longer.
  2. Hacker News: Claude partial outage thread (community, 2026-09-29). About 200 comments; we read a summary of the visible part.
  3. Hacker News: Meta and Microsoft reduce employee use of Claude (community, 2026-10-07). About 130 comments seen. The headline claim comes from a news article we did not read.
  4. Hacker News: Anthropic reportedly passed a user's chat content to police (community, 2026-10-05). About 80 comments seen. We did not read the underlying news report and name no individual.
  5. Simon Willison: Claude Sonnet 5.5 (expert, 2026-09-28). Read through an index-page summary, not the full post.
  6. Simon Willison: Tree Appreciator (expert, 2026-10-14). A practitioner's demo of Opus 5.5 on one task; read through an index-page summary.
  7. Claude apps release notes (Anthropic) (official, 2026-10-07). Date is that of the newest entry. Vendor statements about its own releases.
  8. How do usage and length limits work? (Anthropic) (official, 2026-10-18). The page shows 'Updated this week', not a date; 2026-10-18 is when we read it.
  9. What are artifacts? (Anthropic) (official, 2026-10-18). The page shows only a relative update time; 2026-10-18 is when we read it.
  10. Is my data used for model training? (Anthropic privacy centre, commercial products) (official, 2026-08-18). Page states 'Last updated: August 18, 2026'.
  11. Commercial Terms of Service (Anthropic) (official, 2025-06-17). Effective date shown on the page. Older than the six-month window but it is the governing contract text.

ChatGPT

Sources dated 2026-08-30 to 2026-10-01; collected 2026-10-18. Confidence: low.

What people praise

  • Some heavy users of agent-style work said the Pro subscription still costs far less than equivalent API use. [2]
  • Some commenters on the Sites launch reported good first experiences, including one who built a working 3D viewer and one who valued getting a simple site hosted without signing up for extra services. [1]
  • A practitioner found that ChatGPT Work could assemble accurate boundary files from government data (one task, one person). [7]

What people complain about

  • Pricing and limits: many commenters called the Pro plan change a downgrade because the $200 tier's included usage was reduced while the price stayed the same, and criticised OpenAI for not stating concrete limits. [2][5]
  • Ads and tracking: commenters on two threads raised privacy concerns, saying people share personal details with chatbots and that ad targeting and trackers raise the stakes. [3][4]
  • Some commenters called the Sites feature 'underwhelming' or slow compared with Claude's artifacts, and several saw it as a walled garden meant to lock in users. [1]
  • A practitioner called ChatGPT Work 'an extraordinarily confusing and very powerful product'. [6]

Where sources disagree

  • Whether ads on free tiers are acceptable: some said free services must be funded and paid tiers avoid ads, others said the ad-free promise will not last. [3][4]
  • Whether a higher usage multiplier at the same price is a downgrade depends on whether the value per unit has risen; commenters were not persuaded either way. [2]
  • Whether feature launches like Sites help or hurt OpenAI: some blamed side projects for the models trailing rivals, others said the models just have different strengths. [1]

Trends

  • Plan restructuring: WinBuzzer reports a $500 Pro tier alongside $100 and $200 tiers, with existing $200 subscribers keeping the earlier allowance until 29 October 2026 and then moving to a lower one. Details come from a third-party report quoting OpenAI. [5]
  • Community threads in September 2026 centre on advertising in ChatGPT and on new products (Sites, ChatGPT Work), suggesting product expansion beyond chat. [1][4][6]

Expert views

  • Simon Willison (independent developer and blogger): Describes ChatGPT Work as 'an extraordinarily confusing and very powerful product' and says he is still working out how it fits together. [6]

Limitations

  • OpenAI's own help pages and announcements were blocked, so the record rests on community, news and blog sources only.
  • Most discussion comes from Hacker News, which skews toward developers. Reddit and review platforms were not read.
  • We read long threads through a summarising fetch tool and only the first part of each; counts are approximate and we compute no sentiment scores.
  • Two practitioner posts were read through index-page summaries, not in full.
  • The window is about five weeks, shorter than for Claude, because later OpenAI items were not readable.
  • Signals are not test results. This record does not earn the Hands-On Tested label.

Sources

  1. Hacker News: Sites in ChatGPT (community, 2026-10-01). About 20 top-level comments were readable; the thread is longer.
  2. Hacker News: ChatGPT Pro 500 (community, 2026-09-29). About 100 comments seen. The linked OpenAI help page returned 403 to us.
  3. Hacker News: ChatGPT and an ad collector (community, 2026-09-20). About 25 relevant comments seen. We did not read the linked blog post that made the claim.
  4. Hacker News: OpenAI expands ChatGPT ads with Sponsored Agents (community, 2026-09-16). About 150 comments seen. We could not read OpenAI's own announcement.
  5. WinBuzzer: OpenAI adds a $500 ChatGPT Pro plan and cuts the allowance for new $200 subscribers (news, 2026-09-30). Third-party report that quotes an OpenAI product leader and OpenAI's plan terms; we did not check OpenAI's pages.
  6. Simon Willison: Understanding ChatGPT Work (expert, 2026-08-30). Read through an index-page summary, not the full post.
  7. Simon Willison: GeoJSON Map Viewer (expert, 2026-09-01). One practitioner's experience on one task; read through an index-page summary.

Cursor

Sources dated 2026-04-02 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • Some developers who keep Cursor as a main tool say they value its Tab completions and the ability to stay in control of the code, and one reports using its Composer model as a default on an enterprise plan. [1]
  • Some commenters credit Cursor with ambition for training its own models, and some are interested in its cost claims for Composer 2.5, though others doubt them. [3][6]
  • One hands-on comparison credits Cursor with reviewable diffs, isolated work environments for risky edits, and team controls such as privacy mode, single sign-on and audit logs. [12]

What people complain about

  • Several commenters dislike the agent-first direction of Cursor 3 and say they want the editor and the code, not a chat interface, to stay central. [1]
  • Billing visibility drew complaints: users say the usage page lost its dollar figures for individual plans, and team admins say per-model cost detail is missing. [7][8]
  • Some commenters describe rising or hard-to-predict costs, including a team that reports its bill rose sharply after moving from individual to team plans. [3][12]
  • A user reports that installing the iOS app switched them to a new privacy mode they could not reverse. A commenter who says they work at Cursor replied that the cloud-agent feature needs the new mode and that the legacy mode remains for those who do not use cloud agents. [5]
  • Some commenters say they dislike frequent interface changes and tightening limits, and several say they now use Claude Code or Codex instead. [3][4][8]
  • After the SpaceX acquisition, some commenters raise concerns about ownership and about how their data might be used. Cursor's announcement does not address privacy. [9][10][11]

Where sources disagree

  • Whether Cursor still has a clear role: some commenters say they have moved to Claude Code or Codex, while others say Cursor remains their main tool, and the DECODE comparison names categories it wins and loses. [1][4][8][12]
  • How good Composer is: Cursor's own benchmark reports strong results at low cost, while commenters cite an independent tracker that places it further behind, and others report mixed hands-on results. [3][6]
  • Whether the usage-page change was intentional: Cursor staff on the forum describe the removal of dollar figures from the usage page as deliberate, while a commenter who says they work at Cursor describes the broken dollar values in the CSV export as accidental and since fixed. [7][8]

Trends

  • Cursor's changelog shows five entries between 10 September and 16 October 2026, covering projects, rollouts and security review, remote control of local agents, an app marketplace for Origin, and cloud-agent canvases. [14]
  • Cursor states it was acquired by SpaceX on 14 August 2026, after a partnership that began in April. [9]
  • Cursor 3 (2 April 2026) moved the product to an agent-centered interface while keeping an option to return to the classic IDE, per the vendor. [2]
  • The pricing page now describes a usage pool shared between Cursor and Grok that resets weekly, with third-party models billed separately at API list price. [13]

Expert views

  • Toni Vujevic, engineering manager at DECODE: Prefers Claude Code personally, finds Cursor strong for reviewable diffs and team governance, and flags usage-based pricing as complex and capable of exceeding base allowances. [12]

Limitations

  • Most community evidence is from Hacker News, whose users skew toward experienced developers, and it is self-selected: people with strong views post. Reddit could not be read, and G2 and Trustpilot returned errors, so review-platform sentiment is not covered. Product Hunt was reachable but not analysed.
  • Comments are paraphrased from the most-replied top-level comments and do not measure how common any view is. We give no percentages or scores.
  • Statements by commenters who say they work at Cursor are as the commenters describe themselves; we did not verify identity.
  • Search-engine comparison posts that were commercially motivated or unsourced were not used.
  • The changelog source covers only the first page (10 September to 16 October 2026). The pricing page is undated.
  • This is a signal record, not a test. It does not establish how well the product performs.

Sources

  1. Hacker News: Cursor 3 (community, 2026-04-02). Developer reaction to the agent-centered redesign; 544 points and 406 comments at collection.
  2. Cursor: Cursor 3 announcement (official, 2026-04-02). Vendor description of the new interface and of the option to switch back to the classic IDE.
  3. Hacker News: Cursor Introduces Composer 2.5 (community, 2026-05-18). Reaction to Cursor's own model and its cost claims; commenters are split.
  4. Hacker News: SpaceX to buy Cursor (community, 2026-06-16). Large thread (over 1,000 points) on the acquisition report and on whether developers still use Cursor.
  5. Hacker News: Installing Cursor on iOS changes your privacy settings (community, 2026-06-30). A user report; a commenter who says they work at Cursor replied (identity not verified by us).
  6. Hacker News: CursorBench 3.1 (community, 2026-07-02). Commenters question the vendor's own benchmark against independent trackers.
  7. Cursor forum: usage page shows tokens instead of dollars (community, 2026-07-31). User complaints about billing visibility, with replies from Cursor staff on the same days.
  8. Hacker News: Cursor removed cost information from the usage page (community, 2026-08-01). Developer discussion of the change and of Cursor's value against Claude Code and Codex.
  9. Cursor: Cursor is now a part of SpaceX (official, 2026-08-14). Vendor statement of the acquisition. The post makes no statement about privacy or data handling.
  10. Hacker News: Cursor is now a part of SpaceX (community, 2026-08-14). Smaller thread; commenters question the deal's rationale.
  11. Hacker News: Cursor launches Origin, a GitHub alternative (community, 2026-08-17). Reaction to Cursor's code-hosting product, including concerns about its ownership.
  12. DECODE: Cursor vs. Claude Code, in-depth comparison for dev teams (expert, 2026-08-12). Hands-on comparison by an engineering manager at a software agency (published 12 Aug 2026, updated October 2026). The agency sells AI consulting, though it does not appear to sell either tool.
  13. Cursor: pricing page (official, 2026-10-18). Undated vendor page read on 2026-10-18; used for how usage and privacy mode are described.
  14. Cursor: changelog (trend, 2026-10-16). First changelog page lists five entries from 10 Sep to 16 Oct 2026; earlier pages were not read.

GitHub Copilot

Sources dated 2026-04-20 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • Some commenters like Copilot CLI for feeling low-friction and for letting them switch between models from different providers within one workflow. [10]
  • Some commenters see Copilot as well positioned for enterprises because it offers models from several providers, hosted agents and CI integration. [10]
  • A preprint on Microsoft's internal rollout of command-line coding agents (Claude Code and Copilot CLI) reports that adopters merged about 24% more pull requests than expected over a four-month window, with the authors cautioning that this is a proxy for output. [9]

What people complain about

  • The move to usage-based billing drew widespread complaints. Commenters report credits being used up faster than before, higher effective cost for the same work, and frustration that annual subscribers were affected. [2][4][6][7][8]
  • Users on individual plans object to losing access to Opus models on Pro and to changes in model multipliers, and some say they moved to buying directly from the model providers. [1][2][10]
  • Some commenters say Copilot has fallen behind Claude Code and Codex, and one says tooling is inconsistent between VS Code and Visual Studio. [7][8]
  • Users object to the editor automatically listing Copilot as a commit co-author after inline suggestions were accepted, calling it misattribution. [5]
  • A security vendor reports that a flaw in a GitHub Actions workflow went live in a pull request that Copilot Autofix had co-authored and reviewed. Wiz says Copilot did not write the vulnerable code. Commenters mostly discuss CI hygiene and the limits of code review. [12][13]

Where sources disagree

  • Whether the pricing change is fair: some commenters call it a bait-and-switch, while others say subsidized inference could not continue, which matches GitHub's own stated reason that agentic sessions need far more compute. [2][3][4]
  • Whether Copilot is competitive: some commenters say it has lost its early lead, while others argue it is well suited to enterprises and praise Copilot CLI. [7][8][10]
  • What the Microsoft study shows: the authors report more merged pull requests, while commenters argue that this measures output volume and not value. [9][11]

Trends

  • GitHub says all Copilot plans moved to usage-based billing with AI Credits from 1 June 2026. Plan prices are unchanged, code completions stay included, and annual Pro and Pro+ subscribers stay on the earlier scheme until their plans expire. [3]
  • GitHub announced on 20 April 2026 that new sign-ups for Pro, Pro+ and Student plans were paused and that limits and model access were tightened. [1]
  • The product page describes a cloud agent, third-party agents (including Claude and Codex), a desktop app, and a free plan for users without a Business or Enterprise seat. GitHub's Kimi K2.7 general-availability announcement shows open-weight models now offered. [14][10]
  • Copilot code review began consuming GitHub Actions minutes from 1 June 2026, per the vendor. [14]

Expert views

  • Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva (preprint authors): Adopters merged about 24% more pull requests over four months; merged pull requests are an imperfect proxy for value, and token costs at scale can reach millions of dollars a year. [9]
  • Gal Nagli (Wiz): AI-generated pull requests should get the same static analysis and security review as human-written code, and teams should plan for short-lived credentials and fast patching. [12]

Limitations

  • Most community evidence is from Hacker News, whose users skew toward experienced developers, and it is self-selected. In April 2026, one commenter noted that Claude Code and Codex dominate discussion there, so Copilot users may be under-represented. Reddit could not be read, and no review-platform pages were analysed.
  • Much of the sentiment is about the June 2026 billing change, so it may reflect that event more than the underlying product quality.
  • Cost figures quoted by users in The Register come from their own posts and were not verified by that outlet or by us. We repeat none of them as fact.
  • Wiz is a security vendor and the Snowflake incident account rests mostly on its own claims. GitHub's response was not in the post.
  • Comments are paraphrased from the most-replied top-level comments and do not measure how common any view is. We give no percentages or scores.
  • This is a signal record, not a test. It does not establish how well the product performs.

Sources

  1. GitHub: Changes to GitHub Copilot individual plans (official, 2026-04-20). Vendor announcement (updated 14 May 2026): paused sign-ups for some individual plans, tighter limits, changes to model access and refunds.
  2. Hacker News: Changes to GitHub Copilot individual plans (community, 2026-04-20). Developer reaction; many comments focus on model access and pricing.
  3. GitHub: Copilot is moving to usage-based billing (official, 2026-04-27). Vendor explanation of the 1 June 2026 billing change and of what stays included.
  4. Hacker News: GitHub Copilot is moving to usage-based billing (community, 2026-04-27). Very large thread (over 700 points) on the billing change.
  5. Hacker News: VS Code adds GitHub Copilot as a co-author (community, 2026-04-30). Users report the editor adding Copilot as a commit co-author after inline suggestions were accepted.
  6. The Register: Angry devs vow to flee GitHub Copilot as metered billing takes hold (news, 2026-06-02). Trade-press report quoting forum and Reddit users and a GitHub spokesperson; the cost figures are users' own and not verified by the outlet.
  7. Hacker News: Angry devs vow to flee GitHub Copilot (community, 2026-06-02). Smaller thread reacting to the Register piece.
  8. Hacker News: GitHub Copilot App (community, 2026-06-02). Reaction to GitHub's desktop app for working with agents.
  9. arXiv: Adoption and Impact of Command-Line AI Coding Agents (Murphy-Hill, Butler, Savelieva) (expert, 2026-07-01). Preprint on Microsoft's early-2026 rollout of Claude Code and Copilot CLI. It measures merged pull requests, which the authors say do not equal delivered value.
  10. Hacker News: Kimi K2.7 Code is generally available in GitHub Copilot (community, 2026-07-02). Thread links GitHub's changelog; commenters discuss Copilot's model choice, CLI and pricing.
  11. Hacker News: A Study of Microsoft's Early 2026 Rollout (community, 2026-07-13). Commenters question what merged pull requests show.
  12. Wiz: Red Agent finds its way into Snowflake's Jira (Copilot Autofix) (expert, 2026-08-17). Security-vendor account. Wiz sells security products, and most of its technical claims are its own; Snowflake's published statement corroborates only the fix and the credential rotation.
  13. Hacker News: AI-Generated GitHub Copilot Autofix allowed compromise of Snowflake's Jira (community, 2026-08-17). Developer discussion focused on CI security practice.
  14. GitHub: Copilot product page (official, 2026-10-18). Undated vendor page read on 2026-10-18; used for plans, supported IDEs, agents and data-retention wording.

Perplexity

Sources dated 2025-03-06 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • Library guides credit Perplexity for citations that let readers check an answer and for live-web retrieval of current material. One cited 2025 health-technology-assessment study reportedly found it surfaced hard-to-find grey literature and saved search time. [5]
  • The CASRAI guide recommends it for early scoping of an unfamiliar topic: finding terminology and generating search terms to take to a proper academic database. [4]
  • One Hacker News commenter says they pay for Pro, use it several times a day, and call it excellent. This is a single anecdote. [7]
  • One or two commenters in the audit thread credit Perplexity for publishing its sources, which makes its errors easier to check than a competitor that hides them. [2]

What people complain about

  • Cited pages often do not support the sentence they are attached to. The Haus audit found 34.7% of citations on figure-bearing sentences pointed to a page that would not open or lacked the figure (14.4% when scored per claim). About 20 of roughly 48 commenters accepted the core finding, and about 10 questioned the audit's legitimacy. [1][2]
  • Earlier evidence points the same way on sourcing: the 2025 Tow Center study found Perplexity answered 37% of news-source identification queries incorrectly. That was the lowest error rate of the eight tools tested, but still a large share on a narrow task. [3]
  • Library guides say it is a general-purpose answer engine, not an academic search tool. It mixes peer-reviewed and other web pages, and the reasons for choosing particular sources are often unclear. [4][5]
  • Guides advise treating citations as leads, not facts, and checking the primary source. They also note that live-web results are hard to reproduce exactly. [4][5]
  • Crawler conduct is questioned. The Tow Center authors suggest Perplexity may have disregarded publisher robots.txt preferences, and one commenter reports AI crawlers including its own ignoring robots.txt. UBC's wiki also notes copyright suits from media organizations. The commenter's report is unverified. [3][5][9]
  • One commenter says Perplexity, like other AI engines, gives different answers and citations when the same question is asked twice. [8]

Where sources disagree

  • Whether the Haus audit is credible. Roughly 20 commenters took it at face value, while about 10 raised concerns it might be coordinated or low quality. We could not resolve this, and the audit's funding is undisclosed. [1][2]
  • Whether citation mismatch is specific to Perplexity or common to AI search. Several commenters said other tools do the same, and the 2025 Tow Center study found other tools failing more often on its task. The two studies used different tasks, so their figures cannot be compared. [2][3]
  • Whether a user-directed agent such as the Comet browser should be allowed to act on sites like Amazon. In the visible part of the thread, about 10 to 12 commenters were supportive, 4 to 6 critical of Perplexity, and 15 or more focused on the legal framing. Most of the discussion was not about answer quality. [6]

Trends

  • Perplexity's own blog shows several posts a week in October 2026, centered on agents, a sandbox platform it open-sourced, security, and workspaces. This is the vendor's description of its direction, not independent evidence. [10]
  • Brands and developers increasingly treat Perplexity as a channel to be cited in, and Show HN projects track whether AI tools cite a site. This points to adoption as a discovery surface, not to answer quality. [13][8]

Expert views

  • CASRAI Editorial Board: Treat Perplexity's citations as "a lead, not a fact" and check the primary source; it is a general answer engine, not an academic-specific tool. [4]
  • UBC Library Wiki (librarian-compiled): It suits quick, cited lookups, but tools that generate answers differ from tools that find sources, and human oversight is needed for evidence synthesis. [5]
  • Tow Center for Digital Journalism (Columbia Journalism Review): Its findings are not intended to be extrapolated to all models or news organizations; each excerpt was queried once, so reruns could differ. [3]

Limitations

  • Reddit pages and the Nieman Lab write-up of the Tow Center study could not be read, and the EBU/BBC assistant study and Project MUSE were blocked. Their findings are not used.
  • Community evidence comes mostly from Hacker News, which skews toward developers. Most individual comments are single anecdotes.
  • The audit is the only large quantitative source on current citation behavior, its funding is undisclosed, and its method used the API rather than the consumer app.
  • The oldest source, the Tow Center study, dates from March 2025 and tested a narrow task. The Help Center pages are undated.
  • Thread counts are approximate and come from the portion of each thread we could read.
  • We assign no sentiment score. Confidence is low because the independent sources are few and several are single comments.

Sources

  1. Haus Research: Perplexity citation audit (HR-2026-09) (expert, 2026-09-02). Third-party audit of 310 factual questions run through the API, not the consumer app; the page discloses no funding or conflicts, and its own limitations include a single snapshot and hand-picked companies.
  2. Hacker News discussion of the Haus audit (about 48 comments) (community, 2026-09-02). Forum skews toward developers; counts in our summary are approximate.
  3. Columbia Journalism Review, Tow Center: "AI search has a citation problem" (expert, 2025-03-06). Tested identifying the source of news excerpts, which is a narrow task; the authors say the findings should not be extrapolated to all models or publishers. Older than the rest of the window.
  4. CASRAI Editorial Board: Perplexity AI for research (expert, 2026-08-31). Guide from a programme of a US nonprofit; it says vendor claims about search passes were not independently verified.
  5. UBC Library Wiki: Perplexity (expert, 2026-09-28). Last-edited date of a librarian-compiled wiki entry; it cites several studies by reference rather than testing the tool itself.
  6. Hacker News discussion: Amazon v. Perplexity, Ninth Circuit (community, 2026-09-14). Mostly about the law of browser agents, not Perplexity answers; our counts come from the first part of a long thread.
  7. Hacker News comment: daily paid use (community, 2026-09-30). A single anecdote.
  8. Hacker News comment: answers and citations vary between runs (community, 2026-07-12). A single comment from someone building a brand-tracking tool, who has a stake in the point.
  9. Hacker News comment: crawler behavior report (community, 2026-10-18). An unverified first-person report from one site operator.
  10. Perplexity blog (recent posts) (official, 2026-10-16). Date is the newest post listed when we read the page; the vendor describes its own direction.
  11. Perplexity Help Center: What is Perplexity? (official, 2026-10-18). The page is undated; this is the date we read it.
  12. Perplexity Help Center: Spaces (official, 2026-10-18). The page is undated; this is the date we read it.
  13. Hacker News: Show HN, CiteRank (tracks whether AI tools cite your site) (community, 2026-09-25). We saw the title and metadata through the HN search API and did not read the thread.

Consensus

Sources dated 2024-03-20 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • Library evaluations recommend it for early-stage exploration and scoping: it quickly surfaces relevant studies, links back to the papers, and needs little training. [1][2][4][5]
  • Reviewers like the filters and workflow touches: Deakin notes reference-manager export and library full-text links where subscriptions exist, and the 2025 newsletter review calls its pre-search filters "surprisingly rich and transparent". Yale notes natural-language filters. [1][5][3]
  • In one librarian's tests it declined a question with no relevant literature instead of inventing citations. [5]
  • Two Hacker News commenters use it as a quick starting point for health questions. Both are single anecdotes. [9][10]

What people complain about

  • The Consensus Meter counts votes. CASRAI says vote-counting across heterogeneous studies "is not evidence synthesis"; it ignores study design, sample size and effect size. A librarian review calls it "essentially a form of vote counting", and Deakin says the tool does not explain how consensus is measured. [2][5][1]
  • For paywalled papers it mostly analyzes titles and abstracts. Deakin and the librarian review say this lacks methodological detail, and the 2024 HKUST test saw conclusions drawn from introductory statements in abstracts. [1][5][6]
  • It is not reproducible enough for systematic reviews: results vary between runs, and CASRAI says two people running the same query cannot be guaranteed the same results. [2][5][1]
  • Coverage and design lean toward science, technology and health. Deakin, UBC and the librarian review all say it is weaker outside those areas. [1][4][5]
  • AI-generated study classifications and summaries can be wrong or oversimplify. Reviewers advise checking the linked papers. [4][5]
  • Deakin says vendor information on data sources, model training and retention is limited, and the privacy policy contains ambiguous clauses. It also flags copyright and licensing questions for the "Ask Paper" PDF upload. [1]

Where sources disagree

  • How big the corpus is. The vendor's Deep Search post says 200 million papers, its footer 220 million or more, its homepage 400 million or more, and Yale's announcement says more than 250 million. We state no figure. [7][8][3]
  • How far its output can serve as evidence. Library guides say it should not replace structured searching or reading the papers, while one commenter shared a Consensus summary as supporting evidence against a mouse study. This is one comment against several institutional warnings. [13][2][4]

Trends

  • University libraries are trialing it: Yale's trial runs through 31 December 2026. A trial is a sign of institutional interest, not an endorsement. [3]
  • The vendor added an agentic Deep Search feature in 2025, with a free monthly quota in its launch post. This is the vendor's own description of a beta that it says will likely change. [7]
  • On Hacker News it appears mainly in lists of AI literature-search tools, next to ResearchRabbit, Connected Papers, Litmaps, and OpenAI's Prism, which suggests a crowded category. [11][12]

Expert views

  • Deakin Library: Suited to early-stage topic exploration and teaching; it should supplement, not replace, expert search strategies and full-text critical reading, and is not suitable for systematic reviews. [1]
  • CASRAI Editorial Board: Useful for checking whether a claim has any published support and for scoping, but "checkable" is not the same as "verified". [2]
  • A librarian reviewer (Aaron Tay, personal newsletter): Treat it as a scoping tool, explicitly discuss its limitations with undergraduates, and check the Meter's classifications before relying on them. [5]

Limitations

  • No Reddit or researcher-forum threads could be read. Community evidence is five Hacker News comments, which are single anecdotes or passing mentions.
  • The Project MUSE review and the Library Trends review could not be opened, so we rely on library web guides and a librarian's newsletter.
  • Library reviews often share wording and may rely on one another; the UBC wiki, for instance, summarizes other reviews.
  • The HKUST test is from 2024 and the librarian newsletter from November 2025, so parts of them may be out of date. The homepage and the Help-style pages are undated.
  • We assign no sentiment score. Confidence is low because the independent sources are few and most community evidence is anecdotal.

Sources

  1. Deakin Library: AI evaluation of Consensus (expert, 2026-08). Library evaluation, first published June 2025 and last reviewed August 2026; month only, no named author.
  2. CASRAI Editorial Board: Consensus AI search engine guide (expert, 2026-08-24). Guide from a programme of a US nonprofit; a reviewed, general-audience assessment.
  3. Yale University Library: year-long trial of Consensus (news, 2026-03-20). Announces a trial through 31 December 2026; it repeats the vendor's promise of evidence-backed answers and lists no limitations.
  4. UBC Library Wiki: Consensus (expert, 2026-07-17). Last-edited date of a librarian-compiled entry that summarizes other reviews.
  5. Aaron Tay (librarian): A 2025 deep dive of Consensus (expert, 2025-11-15). Hands-on review by a librarian writing on a personal newsletter; older than the 2026 sources.
  6. HKUST Library: Trust in AI, evaluating Scite, Elicit, Consensus and Scopus AI (expert, 2024-03-20). One-prompt comparison from 2024; the tool has changed since, so treat it as historical.
  7. Consensus blog: Deep Search (official, 2025-07-09). Vendor description at launch of a beta feature; the page metadata shows a later date.
  8. Consensus homepage (official, 2026-10-18). The page is undated; this is the date we read it.
  9. Hacker News comment: preparing for a doctor visit (community, 2026-08-06). A single anecdote.
  10. Hacker News comment: starting point for reviewing treatment research (community, 2026-03-31). A single anecdote.
  11. Hacker News comment: listed among AI literature-search competitors (community, 2026-01-27). Mention in a discussion of another product; no evaluation of Consensus.
  12. Hacker News comment: listed among paper-mapping tools (community, 2026-07-09). Mention in a Show HN thread; no evaluation of Consensus.
  13. Hacker News comment: Consensus summary shared as supporting evidence (community, 2025-10-09). A single comment.

Midjourney

Sources dated 2026-05-11 to 2026-10-08; collected 2026-10-18. Confidence: low.

What people praise

  • We found little first-hand user praise of the image product in these sources. The one Hacker News comment praising v8.2 restates the announcement's wording and reads as a repost, not a user report. [4]
  • Midjourney itself reports that early tests of a "Thinking" rerun mode improve prompt accuracy, typography, and coherence. This is a vendor claim from early tests, not independent evidence. [5]
  • Some commenters gave the company credit for being bootstrapped and wanted to see what it builds. This was about the company, not the image product. [1]

What people complain about

  • The most common reaction in the large thread was skepticism about the medical-imaging pivot: doubts about the scale claimed, the quality of the example images, and the lack of published evidence. Several commenters compared it with other overpromised hardware. [1]
  • Commenters raised overdiagnosis, false positives, liability, and privacy as unaddressed questions about whole-body scanning. [1][2]
  • Some commenters said the company's reputation for AI-generated "slop" made them trust its medical claims less. That is a view about reputation, not a finding about image quality. [1]
  • We found no first-hand complaints about the image generator itself (quality, price, or limits) in the sources we could read. [4]

Where sources disagree

  • On the medical product, commenters split between those who see value in cheaper, more frequent imaging and data, and those who argue that screening healthy people can cause net harm. The arg min essay sides with the second group. [1][2]

Trends

  • Midjourney posted five updates between 24 September and 8 October 2026, including two Alpha changelogs and a test of a "Thinking" mode on its Alpha website. [5][6]
  • A Hacker News submission on 25 July 2026 relays that v8.2 became the default model. We could not confirm this from Midjourney's own pages. [4]
  • The company is widening beyond image generation: it announced a medical-imaging effort in June 2026 and, per TechCrunch citing Bloomberg, acquired the astrology app Co-Star in July 2026. [1][3]
  • The studios' copyright case against Midjourney was in discovery as of 11 May 2026. We found no trial date or merits ruling. These are allegations and procedural filings, not findings. [7]

Expert views

  • Ben Recht (arg min newsletter): Calls the ultrasound project mostly vaporware and argues that constant scanning can make people more anxious and over-treated. This concerns the medical product, not image generation. [2]
  • Amanda Silberling (TechCrunch): Suggests that Co-Star's consumer-app experience could help Midjourney launch a standalone app. She labels this her own view; the article also notes the product is mainly reached through a Discord server. [3]

Limitations

  • Midjourney's website, its documentation, The Verge, and Variety could not be fetched (403 or paywall), so first-party plan details and some reporting were not read.
  • Most Hacker News discussion in the window was about the medical product, not image generation. Reddit and Discord were not read.
  • The large Hacker News thread was read only in part (roughly the first 100,000 of about 547,000 characters).
  • The v8.2 release is relayed by a link-less Hacker News post. We did not confirm it from Midjourney.
  • Seven sources is a small sample, and none is a systematic user survey. We report no sentiment percentages.
  • Some third-party review articles appeared in search results but were not opened, so none is used here.

Sources

  1. Hacker News thread on Midjourney Medical (community, 2026-06-18). Large thread (over 100 comments read; the full thread is far longer). It is about the medical product, not image generation.
  2. arg min: "All data is good data" (expert, 2026-06-22). Newsletter essay by Ben Recht on the Midjourney ultrasound announcement. Opinion.
  3. TechCrunch: Midjourney acquired the astrology app Co-Star (news, 2026-07-24). Brief that cites Bloomberg for the deal. Deal terms undisclosed.
  4. Hacker News: "Midjourney Releasing v8.2" (community, 2026-07-25). Two-point submission with no link and two short comments; it relays a release claim we could not confirm from Midjourney.
  5. Midjourney updates: Testing Thinking Mode (official, 2026-10-08). Vendor announcement. Its statements about improved accuracy are Midjourney's own early-test claims.
  6. Midjourney updates index (official, 2026-10-08). Lists five posts between 24 September and 8 October 2026, including Alpha changelogs.
  7. Graphic Policy: studios and Midjourney discovery dispute (news, 2026-05-11). Reports filings in the studios' copyright case against Midjourney. It reports no ruling and does not describe the substance of the dispute.

Runway

Sources dated 2025-12-02 to 2026-10-18; collected 2026-10-18. Confidence: low.

What people praise

  • At launch in December 2025, Gen-4.5 was reported to rank first on Artificial Analysis' Video Arena leaderboard, which uses blind human comparisons. That measures preference, is dated, and we did not check the current standing. [5]
  • We found no first-hand creator or user reviews we could open and verify in this window. [6]

What people complain about

  • The main documented criticism in these sources is legal: three proposed or filed cases allege Runway bypassed YouTube protections to download videos for AI training. These are unproven allegations, and no ruling was reported. [3][4]

Trends

  • Runway presents itself as a platform, not a single model: its homepage lists Creative, Dev (an API), and Robotics products, and mentions an agent, workflows, mobile apps, and an MCP integration. These are the vendor's descriptions. [1]
  • The pricing page lists other labs' models alongside Runway's own (it names Kling, Seedance, Wan, and Nano Banana Pro), so a Runway subscription is described as access to a mix of models. [2]
  • The homepage announces Runway Ads, an open-weight world action model called Praxis-1, an OpenAI Marketplace listing, and a launch partnership with Claude Motion. These are vendor claims that we did not verify elsewhere. [1]
  • Organic Hacker News discussion of Runway's video product was thin in the window: the Runway submissions we found were mostly its own engineering posts with six points or fewer and no comments. [6]

Limitations

  • Runway's help center returned 403, so per-model credit costs were not verified. Third-party pages gave conflicting figures in search results, and we do not use them.
  • We did not read Reddit, creator forums, or YouTube, and no review-platform page was opened, so independent user sentiment is largely missing.
  • Several third-party review pages appeared in search results but were not opened and are not used.
  • The lawsuit items come from secondary reports. We did not read the dockets, and they are allegations, not findings.
  • The Gen-4.5 benchmark result dates from December 2025, before most of the window, and we did not verify the current leaderboard.
  • Six sources is a small sample. We report no sentiment percentages.

Sources

  1. Runway homepage (official, 2026-10-18). Undated page read on the collection date. All product and partnership statements are Runway's own claims.
  2. Runway pricing page (official, 2026-10-18). Undated page read on the collection date. Plans and model lists change often.
  3. RightsTech: YouTuber sues Runway AI (news, 2026-02-25). Reports a proposed class action filed in Los Angeles. The article reports no response from Runway. Allegations only.
  4. ChatGPT Is Eating the World: Runway sued in a fourth copyright case (news, 2026-03-02). Independent legal tracker. It counts four cases and reports three with similar YouTube-related DMCA theories. No rulings reported.
  5. eWeek: Runway launches Gen-4.5 (news, 2025-12-02). Launch coverage that relays Runway's claims and an Artificial Analysis leaderboard result. It includes no independent testing. This predates most of the window.
  6. Hacker News search: Runway stories (community, 2026-04-27). Search listing. The Runway-related submissions found are mostly Runway's own engineering blog posts. Date shown is the earliest such post.