Phind
Last reviewed
May 16, 2026
Sources
35 citations
Review status
Source-backed
Revision
v1 ยท 3,524 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
35 citations
Review status
Source-backed
Revision
v1 ยท 3,524 words
Add missing citations, update stale details, or suggest a clearer explanation.
Phind was an AI-powered answer engine built for software developers. The product combined a live web index with fine-tuned large language models to return cited, code-aware answers to programming questions, and it was widely credited as one of the first developer-facing AI search tools to reach mainstream use after the release of ChatGPT in late 2022 [1][2]. The company was founded in 2022 by Michael Royzen and Justin Wei, both undergraduates at the University of Texas at Austin, and joined Y Combinator's Summer 2022 batch under the original name Hello before silently rebranding to Phind in late 2022 [3][4][5].
Over the next three years Phind released a series of in-house models, including the open-weights Phind-CodeLlama-34B family in 2023, the proprietary Phind-70B in early 2024, and the Llama 3.1 derivative Phind-405B in September 2024, each of which posted competitive HumanEval scores against frontier commercial systems at the time of release [6][7][8]. The product expanded from a free web tool into a tiered subscription business with Phind Plus and Phind Pro plans, and at peak in early 2024 the site was serving more than 27,000 monthly searches in third-party tracking samples [9][10].
The company shut down on January 16, 2026, with roughly two weeks of public notice, just over a month after raising a $10.4 million round disclosed via Axios on December 3, 2025 [11][12][13]. Saved searches, chat history, and account data were deleted on January 30, 2026 [12][14]. The closure was widely read as the canonical end of the first wave of "AI wrapper" search startups, since the moat Phind had built around a developer-focused interface and source citations collapsed once OpenAI, Anthropic, and Google added native web search to their own foundation models [12][15][16].
Michael Royzen began building consumer AI products in high school. In 2017, while still a student, he released SmartLens, an iOS app that used a fine-tuned Inception-V3 vision model to identify objects in photos. The app outperformed Google Lens on several public benchmarks and maintained a small but loyal subscriber base for years, which Royzen has cited as his first experience shipping a model-backed product [3][17]. Royzen and Justin Wei met at the University of Texas at Austin, where both were Turing Scholars in the class of 2022. The pair began working on what would become Phind during Royzen's final year, after he discovered a HuggingFace demo on long-form question answering and became, in his words, "obsessed" with retrieval augmented systems [3][17].
The pre-Phind prototype processed roughly 350 million pages from Common Crawl through Elasticsearch and combined the index with then-current open language models to produce direct answers to natural language questions. The early system was framed as a general purpose conversational search engine rather than a developer tool, though the founders noted in interviews that programming questions returned the best results because the underlying web data was dense, well structured, and easy to verify [3][18].
The product launched publicly in January 2022 under the name Hello, hosted at sayhello.so and beta.sayhello.so. The original positioning was a conversational search engine for general knowledge questions. Hello Cognition was accepted into Y Combinator's Summer 2022 batch (S22), making it one of the earliest LLM-native startups in any YC class [4][5][19]. During the batch the founders began emphasising developer queries over general search, on the basis that technical documentation was both easier for early language models to handle and easier for paying customers to value.
The rebrand to Phind happened quietly in late December 2022. Royzen has said in interviews that Paul Graham, in a chance encounter, suggested the unusual "Ph" spelling as a play on the word "find"; Royzen was initially sceptical but acquired the phind.com domain from an elderly Canadian entrepreneur and switched the brand over [3][20]. The Show HN post for phind.com in February 2023 introduced the product to the broader Hacker News audience as a "generative AI search engine for developers" and became one of the most-discussed launches of that quarter [21].
From mid-2023 onward Phind invested heavily in training in-house models on top of open base weights. The company described this as a strategy to control quality on coding-specific queries while keeping inference cost lower than calling frontier commercial APIs for every search.
| Model | Release | Base | Notable result | Notes |
|---|---|---|---|---|
| Phind-CodeLlama-34B-v1 | August 2023 | CodeLlama-34B | 67.6% pass@1 on HumanEval | Released with open weights on Hugging Face; matched GPT-4 on HumanEval at release [6][22] |
| Phind-CodeLlama-34B-v2 | October 2023 | CodeLlama-34B | 73.8% pass@1 on HumanEval | Fine-tuned on an additional 1.5B tokens of high-quality programming data; ranked #1 on the BigCode Leaderboard for open-source models [6][7][23] |
| Phind-70B | February 2024 | CodeLlama-70B | 82.3% on HumanEval | Trained on 50B additional tokens; served at roughly 80 tokens per second on H100 GPUs using NVIDIA's TensorRT-LLM library; 4x faster than GPT-4 Turbo in Phind's reported benchmarks [8][24][25] |
| Phind Instant | September 2024 | Phind in-house | Not separately published | Speed-tuned smaller model designed to return first tokens within roughly a second; used as the default for short queries on the free tier [26][27] |
| Phind-405B | September 2024 | Meta Llama 3.1 405B | 92% on HumanEval | 128K-token context window with a 32K window at launch; trained on 256 H100 GPUs using FP8 mixed precision; reported to match Claude 3.5 Sonnet on HumanEval at release [26][27][28] |
The Phind-CodeLlama-34B series was the company's first widely cited in-house release. Built on Meta's CodeLlama-34B base, the model was fine-tuned on roughly 70 billion high-quality code and reasoning tokens drawn from open-source repositories and curated technical documentation. The version 1 release in August 2023 scored 67.6% pass@1 on HumanEval, which matched OpenAI's published GPT-4 score on the same benchmark at the time, and was released with open weights under a permissive license [6][22]. The version 2 update in October 2023 raised the score to 73.8% and briefly held the top spot on Hugging Face's BigCode Leaderboard for openly licensed coding models [7][23]. Royzen described the v2 release as the moment when in-house models became commercially viable for Phind, because it could now answer most coding questions with a model the company controlled, rather than paying per-token rates to OpenAI [3][7].
Phind-70B, released in February 2024, was the company's first model built specifically for speed at production scale. The model was trained on top of CodeLlama-70B with 50 billion additional high-quality tokens and was served on H100 GPUs through NVIDIA's TensorRT-LLM library. Phind reported a throughput of roughly 80 tokens per second per stream, compared with the roughly 20 tokens per second that GPT-4 Turbo was producing in similar tests at the same time, and an 82.3% HumanEval score [8][24][25]. The launch was the first time a third-party model from a smaller company posted a competitive HumanEval result while also being roughly four times faster than a frontier OpenAI model, and it received heavy coverage in developer media [24][25].
In September 2024 Phind released two models in tandem. Phind-405B was a fine-tune of Meta's Llama 3.1 405B with a 32,000 token context window expandable to 128,000 tokens, trained on 256 H100 GPUs using FP8 mixed precision; the company reported 92% on HumanEval, matching Anthropic's Claude 3.5 Sonnet at release [26][27][28]. Phind Instant was a much smaller, latency-optimised model designed to return the first tokens of an answer in roughly one second; the company used Instant as the default for short queries on the free plan and to power the assistant's autocomplete-like interactions on Pair Programmer [26][27]. Phind-405B was made available to Phind Pro users on launch day [27].
The core Phind product was a chat-style search box that accepted natural language questions, ran a live web search, retrieved a set of source pages (heavily weighted toward technical documentation, Stack Overflow, GitHub, and engineering blogs), and used a language model to produce a direct cited answer. Each fact in the answer carried an inline citation linking back to the source page, similar in style to Perplexity AI's answer engine but tuned for developer queries [1][2]. Users could expand any answer to see the full source list and could ask follow-up questions that maintained context across the thread.
The product supported what Phind called "multi-step reasoning," which in practice meant that for complex questions the assistant would issue several internal search queries against the web index, read multiple pages, and synthesise an answer that wove the sources together. This was the feature most often mentioned in early reviews and was credited with making Phind useful for questions where the answer lived across several pages of API documentation rather than in a single Stack Overflow post [1][2][29].
In 2024 Phind launched Pair Programmer, a separate mode that emphasised long, code-focused interactions over short web search answers. Pair Programmer accepted full code snippets and stack traces, kept richer context across follow-up questions, and could pull in repository structure when given a project description. The mode was marketed as a complement to dedicated IDE tools like Cursor and GitHub Copilot for developers who preferred a web-based chat interface to a code editor integration [26][30].
Phind shipped a Visual Studio Code extension that put the Phind chat panel inside the editor and gave the assistant access to the open file as context. The extension was not the primary route for most users; site traffic data suggested the web app remained the dominant surface throughout the company's life [29][30]. A browser extension shipped later let users invoke Phind from any web page.
Phind operated a freemium model with two paid tiers. Pricing shifted several times between 2023 and 2025; the structure below reflects the configuration that was in place when the company announced its shutdown.
| Plan | Price | Included | Best-model uses |
|---|---|---|---|
| Free | $0 | Phind Instant, limited best-model queries per day, basic citations | Limited daily quota |
| Phind Plus | $20 per month or $200 per year | Higher daily quota of best-model queries, longer context length, faster servers | Several hundred per day |
| Phind Pro | $30 per month or $300 per year | 500+ best-model uses per day (GPT-4 class), input length up to 12,000 characters, priority access to Phind-70B and Phind-405B | 500+ per day [9][31] |
In 2024 the paid plans were renamed and re-tiered after Phind-405B's launch, with Pair Programmer access bundled into the paid tiers. The free plan remained a real product through to the shutdown, supported by the cheaper inference cost of Phind Instant and the in-house Phind-70B [27][31].
On January 2, 2026, Phind announced that the product would shut down on January 16, 2026. The notice was posted on the company's website and on its X account, giving users two weeks to export saved searches and chat history. User accounts and stored content were scheduled to be permanently deleted on January 30, 2026 [11][12][14].
The announcement came roughly six weeks after Axios reported on December 3, 2025 that the company had closed a $10.4 million round, which Royzen described in the same interview as Series A funding intended to expand the product's visualisation features and improve the answer engine [13]. The proximity of a fresh fundraise to a complete shutdown drew sharp commentary from observers including Ed Zitron, who pointed out on X that Phind had "shut down just over a month after raising over $10m" [15][32]. Hacker News discussion threads on the shutdown were among the most active in the first half of January 2026 [11][12].
Neither Royzen nor the company published a single definitive explanation. Several factors were consistently cited in coverage of the shutdown and in community discussion.
Frontier models absorbed the feature. Through 2024 and 2025, OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini all added native web search with citations directly inside their consumer products. Once a $20 per month ChatGPT subscription returned cited web results inside the same chat where users were also writing code, the case for a dedicated developer search engine became weaker. As one widely shared analysis put it, "When OpenAI, Google, and Anthropic add web search to their foundation models, a dedicated developer search engine becomes redundant" [12][16]. This was the dominant explanation in reviews and post-mortems.
Unit economics on premium inference. Phind's competitive edge depended on serving best-in-class answers cheaply, but it was simultaneously paying to train its own large models (Phind-70B and Phind-405B both required significant H100 GPU time) and running an expensive live retrieval pipeline. Several Hacker News commenters who described themselves as former or current employees noted that the cost of frontier inference kept climbing while the price ceiling on developer subscriptions did not budge from roughly $20 to $30 per month [11][12].
Traffic decline from a 2024 peak. Third-party tracking data referenced in shutdown coverage suggested Phind's monthly search volume fell roughly 91% from its peak in early 2024 to early 2026. The peak appears to have come during the Phind-CodeLlama and Phind-70B release cycle, when the in-house models gave the company a genuine quality lead on coding questions. By mid-2025 Phind's quality lead over frontier chat products on programming queries had been largely closed [12][16].
Competitive convergence. Perplexity AI expanded into developer use cases through 2024 and 2025, while Cursor and other IDE-integrated tools captured the deeper coding workflows where users wanted the AI inside their editor rather than in a browser tab. Phind sat between two adjacent product categories that were both moving faster [33][34].
Founder transition. By the time of the shutdown, Royzen had been listed publicly as the CEO of Standard Signal, a new Y Combinator P26 batch company, which suggested at least part of the founding team had moved on to a fresh project before the closure was announced [5][35].
The two-week notice window was widely criticised, particularly because many developers had stored long chat histories and saved searches inside Phind as a kind of working memory. Coverage in industry newsletters and on developer forums emphasised the speed of the shutdown more than the fact of it [11][12][14]. The Phind website was kept up through January 16, 2026 to allow data export, then taken offline. The Phind-CodeLlama-34B-v1 and v2 weights remained available on Hugging Face under their original permissive license after the company closed [22][23].
Phind's main competitors evolved over its life. At launch the closest product was Perplexity AI's answer engine. By 2024 the practical comparison had widened to include frontier chat products with web search and IDE-integrated coding assistants like Cursor and GitHub Copilot.
| Product | Primary surface | Strength on coding | Citations | Status as of May 2026 |
|---|---|---|---|---|
| Phind | Web app, VS Code extension, browser extension | High on focused programming queries, particularly with Phind-70B and Phind-405B [8][27] | Inline source citations by default [1][2] | Shut down January 16, 2026 [11][12] |
| Perplexity AI | Web app, mobile apps, Comet browser | Good on general technical questions, less code-specific tuning [33] | Inline source citations by default | Active |
| ChatGPT | Web app, mobile apps, desktop apps, ChatGPT Atlas browser | Strong on code generation; gained native web search in 2024 [16] | Citations available with web search mode | Active |
| Claude | Web app, API, IDE integrations | Strong on long-context refactoring and reasoning; gained web search in 2025 [16] | Citations available with web search | Active |
| Cursor | VS Code-style desktop editor | Built around the codebase rather than the web; integrates with multiple frontier models [34] | Codebase-aware grounding rather than web citations | Active |
| GitHub Copilot | IDE plug-in | Built around the codebase and developer workflow; supports model selection | Inline code suggestions rather than cited answers | Active |
A recurring observation in side-by-side reviews from 2024 and 2025 was that Phind was best when the developer was searching the web for an answer, Cursor was best when the developer was already inside the codebase, and Perplexity was best for the research and architecture decision phase. The split worked while each tool had a clear niche; the trouble for Phind was that the frontier chat products gradually became acceptable substitutes for the web-search-for-developers slot [29][33][34].
Phind's run between 2022 and 2026 mattered for reasons beyond its commercial outcome. The Phind-CodeLlama-34B-v2 release in October 2023 was, for a brief window, the top-scoring openly licensed coding model on the BigCode Leaderboard, and it was used widely as a base for further fine-tuning by other groups [7][23]. The Phind-70B release in February 2024 helped establish that an LLM-native startup could train and serve a competitive in-house model at meaningful scale without becoming a foundation model lab, a template that several subsequent companies followed [8][24].
The shutdown itself became a reference case in commentary about AI startup defensibility. Posts that framed Phind as a "wrapper company" undone by frontier model expansion were widely circulated through January and February 2026, and the company was repeatedly cited as a cautionary example in discussions about whether application-layer AI startups could build durable moats against the frontier labs whose APIs they sometimes resold [12][15][16][32]. Whether that framing is fair is contested; Phind shipped multiple in-house models and was not purely an API reseller, and Royzen has said publicly that the company was profitable on some product lines before the shutdown [3][12]. The simpler reading is that the developer search niche turned out to be too narrow to support a standalone company once frontier chat tools could do most of the same work [12][16].
The Phind-CodeLlama weights are still hosted on Hugging Face under their original license and continue to be downloaded for research and self-hosting [22][23]. As of May 2026, Royzen is reported to be running Standard Signal, a separate Y Combinator company, and Justin Wei's current activities are not publicly listed [5][35].