0

Changelog

Notable updates to Sophon - new features, fixes, and improvements, newest first.

October 29, 2026

  • ImprovedGPQA Diamond, Humanity's Last Exam, SciCode and AIME 2025 scores now come from Epoch AI and MathArena, which publish them directly, instead of through Artificial Analysis. Artificial Analysis stops sharing per-benchmark numbers for free on November 4, so without this change these scores would have stopped updating that day. IFBench, Terminal-Bench Hard, LiveCodeBench, MMLU-Pro, MATH-500 and AIME 2024 have no other public source and will keep the last numbers they had.

September 22, 2026

  • AddedGPT-6 Sol and GPT-6 Luna are in the catalog on launch day, with Artificial Analysis scores and their prices ($2 / $10 and $0.10 / $0.50 per million tokens, half what GPT-5.6 Sol and Luna cost). OpenAI's own launch numbers are here too, marked as vendor-reported, so Sol sits next to Astra on DeepSWE, OSWorld 2.0, Agents' Last Exam and AutomationBench.
  • AddedClaude Opus 5.5 is in the catalog on launch day, with its price ($4 / $20 per million tokens, cheaper than Opus 5), its 1M context window, and the first seven vals.ai scores. The independent leaderboards that usually bring new models in, Artificial Analysis and LMArena, had not added it yet. The rest of its results will show up as those sources publish them.

September 18, 2026

  • AddedJev is in the catalog, three days after TypeSafe AI came out of stealth with it and $40M. Jev is not a chatbot: it returns a typed decision with a probability attached instead of text, so none of the nightly sources would ever have brought it in, and it is not on OpenRouter either. The four workflows TypeSafe published are here as four benchmarks, one page each, every one of them also run on GPT-5.6 Terra and Luna and on Claude Opus 5, Sonnet 5 and Haiku 4.5, which makes the actual trade visible: Jev wins one of the four on accuracy and loses the other three, while costing $0.0001 a decision against Opus 5's $0.0574 and answering in 0.3 seconds against 15. The benchmark is TypeSafe's own and has no human answer key. Its correct answers are whatever GPT-6 Astra and Claude Fable 5.1 agreed on, so every model is being scored on how closely it imitates those two, and all four pages say exactly that. The two independent tests published so far, one on extraction and one on tool-call risk, are noted on Jev's page for what they do and do not cover.

September 7, 2026

  • AddedPeople and Papers now have a full index behind them, linked from the bottom of each page. Both pages hold their whole list in the browser so you can filter it instantly, and that caps them at 2,000 entries: the 2,000 most prolific contributors, the 2,000 most recent papers. The other 18,575 researchers and 20,005 papers were only ever a search or a co-author link away, never something you could simply page through. Now you can, 200 to a page.
  • AddedModel pages now list the head-to-head comparisons that model takes part in, and each comparison page links the other comparisons its two models appear in. All 436 of those pages already existed, but only the twelve on the Compare page were linked from anywhere, so following links from the home page reached 13 of them. "GPT-5 vs Claude Opus 4.5" is the shape of question people actually type, and the page that answers it now sits where the question comes up, on the model's own page.
  • AddedEvery browse page now has a pager at the bottom, so a page of results has an address of its own that you can link to or come back to. Scrolling still loads more as you go. The pager is for the reader who wants page 40 of the environments list without scrolling to it, and for search engines, which do not scroll: they were being shown the first 24 rows of each list with no way to reach the rest, so 2,244 of the 2,268 RL environments were known to them only from the sitemap.
  • FixedSearch engines could not see the catalog. The Models, Evals, RL environments, Leaderboards, Organizations, People and Capabilities pages built their lists in the browser, so anything that does not run JavaScript, Google and the AI crawlers included, was served the page heading and an empty list: no rows, and no links through to the 4,802 model, benchmark, environment and organization pages underneath them. Following links from the home page, only 8% of those pages could be reached at all, and RL environments was the worst of it at 37 of 2,268. The lists are now built on the server, so the links are in the page itself.
  • FixedThe sitemap had not changed since the last deploy three days earlier, and robots.txt was serving a copy 55 hours old under a 24-hour refresh. Both files were being frozen at build time, so everything the nightly sync adds, and it adds models, benchmarks and environments daily, went into the database and was never announced to search engines. Both now refresh on their own. The "last changed" dates in the sitemap also come from when each entry actually changed, instead of from the moment the file happened to be written, which is the difference between a date Google can use and one it learns to ignore.

September 4, 2026

  • FixedGPT-6 Astra's page had one benchmark wrong and four others described loosely. It was built from launch-day reporting because OpenAI had published nothing itself yet: no launch post, no system card, no entry in its own API listing. That page is up now, and OpenAI's table says ARC-AGI-3 is 99.9%, not the 98.6% that was reported at the time. The cybersecurity results now carry OpenAI's own wording, "without production safeguards", instead of the name of an access program, which matters because Astra as it ships refuses that work outright. Every number on the page now comes from OpenAI's table and links back to it.
  • ImprovedGPT-6 Astra now shows its whole launch table, 27 results instead of 9, adding Terminal-Bench 4.0, Agents' Last Exam, FrontierCode, Humanity's Last Exam, ARC-AGI-1 and ARC-AGI-2, the long-context MRCR pair and the science and health benchmarks. Its context window and input types are filled in from OpenAI's model docs. Artificial Analysis published Astra overnight, so the page also carries an independent index score and its six effort tiers next to what OpenAI reported.

September 3, 2026

  • AddedGPT-6 Astra is in the catalog the day it launched, with the nine benchmark results OpenAI published for it and its announced $10/$50 price. The nightly sync could not have brought it in: none of the three sources allowed to add a model carried it yet, and it is not in OpenAI's own API listing either, so the model would have been missing here for days while every outlet wrote about it. Five of the benchmarks in its launch table were new to the catalog as well, among them DeepSWE, ARC-AGI-3 and BenchCAD. The conditions attached to each number are recorded alongside it, so the 100% on ExploitBench reads as what it is, a result from the restricted access program rather than from the model as it ships.

September 2, 2026

  • ImprovedThe RL environments and datasets page opens on what is new and what people are using, instead of the same unchanging list. It ordered by how many benchmarks an environment improves, a number only 251 of 2,267 entries carry, so everything else tied at zero and the top of the page stayed put while the hubs published around 186 new environments a month underneath it. The order is now recency with a head start for environments people have starred: one star is worth roughly two weeks of freshness, so a fortnight-old environment with ten stars still opens above today's uploads without burying them. Every row and card now also shows the two numbers its position comes from, stars and age, so the order can be read instead of guessed.
  • FixedSorting environments by "Last updated" now orders by when they were last changed. The date was never stored, only printed in the page text, so the control fell back to publication date and quietly answered a different question. 1,979 entries now carry a real date from Prime Intellect or OpenReward, and 916 of them were changed after they were first published, which is what separates a maintained environment from a dormant one.
  • FixedFiltering models by licence no longer splits one licence in two. The page links to /models?license=<value>, and the stored values disagreed on capitalisation, so "MIT" and "mit" were separate filters - across 447 models, including every Apache-2.0 and proprietary one. Six such pairs were folded together. 34 models that showed no licence at all now show one, read from the model's Hugging Face card; Qwen3 Coder and DeepSeek R1 had been reading "unknown" where their cards say Apache-2.0 and MIT.
  • FixedLlama 3.1, Gemma 2 and 3, Kimi K2, MiniMax and Qwen2.5-VL no longer read "Closed weights" - all of those publish downloadable weights, and 15 models were labelled the opposite of the truth. A further 46 had no weights answer at all, so the tag simply did not appear on them: Artificial Analysis stopped sending that field, and every nightly sync since recorded the absence as fact rather than noticing. 384 of 999 models were affected, almost all of them recent - 69% of everything released this year. Weights are now read from whether the model publishes to Hugging Face, and the sync warns instead of staying quiet when a field it depends on disappears upstream.
  • ImprovedThe leaderboards page now opens as cards, and each card shows who is currently on top of that board: the three leading models with their scores. Boards that publish no score of their own, like the Open LLM and SWE-bench leaderboards, show the leaders on the benchmark they track most widely, named above the list. The table view is still one click away.
  • ImprovedThe home page now shows what just shipped. It led with the standings and stopped there, so anything new sat behind a Feed page you had to know about: the ten newest evals, tools, models and papers now sit right under the frontier chart, with a link through to the full feed. The sidebar's separate Feed row is gone, and Collections now shows how many items you have saved.
  • ImprovedPages you have opened and comparisons you have run moved off the home page onto their own History page, in the sidebar next to Collections. Home keeps the things worth acting on: what you saved, and what is new in your saved views.
  • FixedEleven organization addresses have been returning "not found" since the organizations behind them were folded into a twin, among them /organizations/arcee, /organizations/prime-intellect and /organizations/liquidai. They now open the surviving page, and a future merge cannot ship without its redirect.
  • FixedOrganization pages now list the benchmarks and environments they published. 211 of them had been sitting unattributed - 111 Atropos environments belonging to Nous Research, 72 Prime Intellect ones, 28 OpenEnv ones belonging to Meta - because nothing ever filled in the publisher after ingest. A further ten OpenEnv environments had been credited to Hugging Face, which hosted the announcement but not the code. That now runs with the nightly sync, so newly ingested work arrives attributed instead of adding to the pile.
  • FixedTwelve organizations that existed twice have been folded into one, so their pages are no longer half-empty. ARC Prize is the clearest case: one page listed Greg Kamradt and the foundation's tweets, a second page under a near-identical name listed François Chollet, and neither linked to ARC-AGI at all. Same story for LMArena, whose five leaderboards were split across two rows, the Center for AI Safety, Stanford, Oxford, Washington, Waterloo, Toronto, Numina, ServiceNow Research, XLANG Lab and the Mathematical Association of America. The retired addresses redirect to the surviving page. LMArena had been merged once before and the nightly sync kept recreating the twin; it no longer does.

September 1, 2026

  • ImprovedSearch is roughly 20 to 60 times faster and now leads with what you named. Typing into ⌘K fired a query for every keystroke, and each one re-read the full text of every matching entry to score it: "re" took 19.7 seconds, "cl" 9.3, "arc" 3.5, so the results for what you finally typed queued behind the ones for what you had half-typed. Scoring now reads a prepared index instead, and the same searches land in 70 to 340 milliseconds. Titles and slugs also count for far more than a passing mention in a body, so "arc prize" opens with ARC Prize and the ARC Prize Foundation rather than with people who happen to be mentioned on their pages, and "arc agi" opens with the ARC-AGI benchmark rather than with community forks of it.
  • ImprovedSearching a model by its full name now finds that model. Version numbers were being thrown away before the search ran, so "claude-opus-4-8" searched for "claude" and "opus" and returned 25 assorted Claude models; it now returns Claude Opus 4.8. Same for "gpt-5", which led with GPT-5.1 and now leads with GPT-5.
  • ImprovedA model can now be found by the id you have in your code. The dated and vendor-prefixed ids - "claude-opus-4-5-20251101", "openai/gpt-5-2025-08-07", "gemini-3-pro-preview" - were not searchable at all, so pasting one returned whatever the fuzzy fallback could scrape together: "claude-opus-4-5-20251101" answered with an unrelated FastAPI environment. 835 such ids across the catalog now resolve to their model.
  • ImprovedSearching a topic now offers the page that gathers everything on it. Capabilities were the one section of the site search could not reach, so "tool use" returned three RL environments that happen to mention it rather than the page listing all 203, and "memory", "rag" and "long context" behaved the same way. They now lead for a bare topic word, while anything actually named after the topic still comes first.
  • ImprovedSearch results no longer repeat themselves. Every RL environment ingested from Prime Intellect and OpenReward is stored as both a tool and an eval, and popular ones are re-published by several owners, so "arc agi" returned 40 rows in which the same handful of things appeared up to six times each. Identical entries now collapse to the best-ranked one, and the same search returns 25.
  • FixedBenchmark numbers on 95 model pages were reading well below what the benchmark actually reports, and now match it. When a source starts publishing a model at several reasoning settings, the older single number stays in the table with no setting attached, and it was outranking the settings that replaced it while quietly serving whichever one it had been copied from: DeepSeek V4 Pro showed 71.7 on GPQA Diamond against a published 92.8, 8.2 on HLE against 41.0. 710 scores changed, 658 of them upward by 15 points on average. The 52 that moved down are models whose reasoning mode scores below their non-reasoning one, and a model is represented by its strongest setting.
  • FixedDeepSeek V4 Flash now means the version DeepSeek ships today. The page under that name held the April build and scored 42.1 where Artificial Analysis shows 51.8, because the current July weights sat on a second page called "DeepSeek V4 Flash 0731". The name now belongs to the July build, and April keeps its own page as DeepSeek V4 Flash 0423 with the Elo and benchmark history it earned. The two are different weights, not two names for one thing: GPQA Diamond goes from 71.6 to 90.8 between them.
  • FixedArtificial Analysis scores read the model's best published setting again. AA lists some models more than once, a dated snapshot beside the rolling name, a ChatGPT-tuned build, a non-reasoning mode, and the weakest of those was writing the headline number: GPT-5 showed 15.4, the ChatGPT build, where AA shows 35.3, and DeepSeek V4 Pro showed a stale April snapshot at 45.3 instead of 53.2. Six models were affected, and their price and speed came from the same wrong entry.
  • Fixed1,405 RL environments carry their real names again. A cleanup script had rebuilt every title from the last two pieces of its address, so "prime-ryanznie-arc-agi-3-env" read "3 ENV RL Env (Community)", and short words came out shouted, as in "Fruit BOX" and "Goofy MATH". Names now come from the source listing, with the owner added only where two environments would otherwise share one.
  • FixedSearch shows a thing's current name again after it is renamed. The nightly indexer decided what to re-read from a date field that only ever moves when the upstream source publishes a revision, so anything corrected on our side kept its old name in search indefinitely: LMArena was listed as "Arena", and fifteen papers carried a title arXiv had since revised. The indexer now notices the name itself changing; 18 entries were stale and are fixed.
  • FixedPasting a paper link into search now finds the paper, or says nothing was found. An arXiv link is read for its paper id; other links no longer have their scheme and host searched as if they were words, which had been returning a page of unrelated benchmarks for a paper the site does not carry.
  • FixedA model that keeps its weights private no longer looks retired. Every closed-weights model carried a bare gray "Closed" tag beside its name, sharing that row with tags like "Withdrawn" and "Not available", so it read as a model that had been pulled rather than one you can only reach through an API. Weights availability now sits with the other specs, next to context length and price, and reads "Weights: Closed, API only".

August 25, 2026

  • AddedHead-to-head pages for the models people actually compare: /compare/claude-opus-4-8-vs-gpt-5.5, /compare/gpt-5-vs-gpt-5-mini and 433 more. Comparing two models used to live behind a temporary address that could not be linked or found, so the single most common question in the field - "is A better than B?" - had no page to land on. Each pair now has its own, with the benchmarks both models report, price per million tokens, context window, release dates, and a short plain-language answer on who leads and where. The most-compared pairs are listed on /compare.
  • FixedA model with no published price no longer reads "$0.00" in a comparison. 195 of 454 price records store a zero to mean "not reported", which made a paid model look free; those cells now show a dash.
  • ImprovedSearch engines are now pointed at the pages that carry Sophon's own work rather than at everything. Roughly 130,000 of the 205,000 pages are a name attached to a single paper, or an abstract copied from arXiv with nothing added; those stay on the site and stay linked, but are no longer offered up for indexing. What is offered is the catalog itself plus the 42,000 people and papers with a real profile, an employer, a contribution, citations, code, or a link into the benchmark graph.

August 19, 2026

  • FixedModel names no longer carry a stray "(batch)" tag. OpenRouter recently began listing a discounted batch endpoint alongside each model, and that listing's name was overwriting the real one, so the flagship read "Claude Opus 5 (batch)" on its own page, in every leaderboard row and in search results. 49 models were affected, among them GPT-5.5, Gemini 2.5 Pro and Claude Sonnet 5. The same listing was also setting their cached-input price to the batch discount while input and output stayed at standard rates; all three now read as the standard rate.
  • FixedBenchmark scores again come from a model's strongest setting when the benchmark is not Artificial Analysis. The rule that picks between a model's reasoning settings only ever applied to AA's numbers, so on a source that publishes several settings the most recently written one won at random: Claude Opus 4.6 showed 68.8 on Thematic Generalization where it actually scores 80.6, which is first place. 76 scores were affected.
  • FixedLMArena ratings are current again. They had not refreshed since 28 May, so the site showed three-month-old human-preference Elo and was missing every model released since. The boards now refresh weekly on their own schedule, and a missed run shows up in the ingest health view instead of ageing unnoticed.
  • FixedA model no longer risks appearing several times on an LMArena board, once per reasoning setting. LMArena lists "claude-opus-5", "claude-opus-5-high" and "claude-opus-5-max" as separate entries; each now folds into the one model, represented by its strongest setting that has the votes to back it.
  • AddedTwo benchmarks from Lech Mazur's independent suite: Creative Story-Writing (each model writes a story that must use 10 assigned elements, graded by a panel of LLM judges) and Thematic Generalization. 71 model scores, with each reasoning setting kept separate. Claude Opus 5 at extra-high effort leads the writing board with a 96% expected win rate against the field.
  • AddedSix new LMArena leaderboards, ranked by human votes rather than an LLM judge: Creative Writing, Writing and Literature, Instruction Following, Hard Prompts, Coding and Math. Claude Opus 4.6 leads creative writing at 1506 Elo, Claude Fable 5 leads writing and literature. The site carried only LMArena's blended overall rating until now, so there was no way to ask which model writes best.
  • AddedCreative-writing and emotional-intelligence scores are now in the catalog, from EQ-Bench: 214 scores across its Longform Creative Writing board (a 0-100 rubric on an 8-chapter story), Creative Writing v3 and EQ-Bench 4 (both Elo). Claude Opus 5 leads all three. Nothing in the catalog covered writing quality or emotional intelligence before, and the boards refresh with the nightly sync.

July 26, 2026

  • ImprovedPaper pages now live at readable addresses named after the paper - /papers/attention-is-all-you-need instead of /papers/pwc-19007 - across all 61,000 papers. Old links keep working via redirect.
  • ImprovedSearch-result titles now say what a page actually offers: benchmark pages read "GPQA Diamond benchmark: model scores & leaderboard" instead of just the benchmark's name, leaderboard pages name the current top models in the preview text, and a person's page leads with who they are ("Bob McGrew - Former Chief Research Officer of OpenAI") rather than a generic "researcher" label.

July 25, 2026

  • FixedA model no longer shows up once per reasoning-effort setting. Artificial Analysis benchmarks each setting separately, so Claude Opus 5's launch filled the feed with five near-identical cards - low, medium, high, xhigh and max - and the same duplication ran through the models list, search and compare. Each family is now a single entry, and new launches stay that way. 26 duplicate entries were folded back into 11 models, including Claude Opus 5, Claude Sonnet 5, and GPT-5.6 Sol, Terra and Luna, which each carried six.
  • AddedModel pages now show how the model scores at each reasoning effort, on 90 models. Claude Opus 5 ranges from 60.7 at max effort down to 50.6 at low, on identical pricing; GPT-5.6 Luna spans 51.2 to 26.6. Previously you could only see that by comparing what looked like separate models.

July 24, 2026

  • AddedFirst-time visitors now get a short three-step intro: what the catalog covers, how to search all of it with ⌘K, and how to save a set of feed filters as a view and have new matches emailed to you. Each step shows the actual screen it describes. It replaces the "Start here" band that used to sit on the home page.

July 21, 2026

  • FixedBenchmark scores on model pages now come from the model's strongest setting. Vendors ship many models at several reasoning efforts (GPT-5 at high, medium, low and minimal; GLM-5 with reasoning on or off), and every setting's scores were overwriting the others, so a page could show the weakest setting's numbers under the model's plain name - and which setting won could change from one night to the next. 81 models were affected. GPT-5's GPQA Diamond reads 85.4% instead of 80.8%, and its HLE 26.5% instead of 18.4%.
  • FixedSearch no longer returns dead results. About 500 entries still pointed at models, tools, leaderboards, organizations, and evals that had been merged into another entry or removed, so they either led nowhere or pushed a real result off the list. Search now drops such entries as soon as the entry they point to goes away.
  • FixedThe nightly refresh finishes again. It had been running over its time limit and getting cancelled partway on 4 of the last 10 nights, so fresh benchmark scores, release dates, and newly added papers and people could show up a day or more late, or wait for the next night that happened to finish. The slowest steps are now minutes faster, and the parts that record scores and update search no longer get dropped when an earlier step runs long.

June 26, 2026

  • AddedModels that are announced but not generally available - like OpenAI's new GPT-5.6 Sol, Terra, and Luna, or the rolled-back Claude Fable 5 - now carry a "Not available" badge across the catalog, and their score bars render with a checkerboard fill on the charts, the same way Artificial Analysis flags them. Numbers that are still provisional (independent evaluation forthcoming) show a diagonal hatch.

June 10, 2026

  • FixedPasting a Papers link into an AI assistant (or any reader that doesn't run JavaScript, like crawlers) now shows the actual paper list with the link's filters applied - week/month/year window, domain, search. Previously such tools got an empty page shell with no papers at all.

June 9, 2026

  • FixedNew models, scores, and papers land in the feed every night again. The nightly data sync had been timing out and getting cut off partway for a week (Jun 2-8), so several days of updates surfaced late or not at all.
  • AddedClaude Fable 5 and Claude Mythos 5 (released today) are in the catalog with their full launch benchmark table - 13 evals including seven new ones (FrontierCode, ExploitBench, BioMysteryBench, and more), pricing, and specs. Each benchmark shows Fable 5 against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, with scores marked by who reported them.

June 8, 2026

  • AddedAI models now have a "Mentions on X" tab on their page, showing what people are saying about the model on Twitter/X - filtered for spam, ranked by signal, and tagged by what kind of mention it is (comparison, result, news, opinion). Sort by signal, most recent, or most-liked.
  • FixedAn eval's README tab now shows the actual README again, instead of a block of raw configuration text (title, tasks, tags, and so on).
  • ImprovedA model page now loads its benchmark-score charts in batches as you scroll, instead of rendering all of them at once, so pages for models measured on dozens of benchmarks no longer stall.
  • ImprovedThe Papers "Trending" tab now defaults to the last month instead of all time, so it reflects what's actually trending now; switch to this week or this year anytime.

June 6, 2026

  • AddedSearch results in the API and `sophon` CLI can now be sorted: pass `sort=title` or `sort=recent` (prefix `-` to reverse) to `/api/v1/search`, or `sophon search … --sort title`. The default stays relevance.

June 3, 2026

  • ImprovedOpening a paper's first-page preview now expands it into the reader with a quick animation, and collapses back when you close it, instead of opening abruptly.

June 2, 2026

  • AddedBranded social preview cards for every page - the image shown when a Sophon link is shared on Slack, X, Discord, or LinkedIn. Previously only paper pages had one; the homepage card shows live catalog counts.
  • AddedA public API and a `sophon` command-line tool for querying the catalog (evals, models, tools, papers) from code, agents, or the terminal - find them on the new API & CLI page, linked in the sidebar.
  • FixedCollapsing the sidebar no longer briefly flashes its floating version before it closes.
  • FixedThe feed's New tab no longer shows older papers as if they were just published; papers are ordered by their real publication date again.
  • FixedWhen a paper's preview image fails to load, the feed and paper page now render the paper's real first page instead of a broken image or blank placeholder.
  • FixedPaper abstracts no longer show raw LaTeX code (like `\textsc{}` or `\href{}{}`); the markup is cleaned up so the text and links read normally.
  • ImprovedThe Capabilities page is now grouped into categories (reasoning, language, safety, agents, coding, and more) you can filter by, instead of one long flat list.
  • ImprovedOn list pages, moving between the filter menus now glides the dropdown across to the next filter (or the Add filter button) in place, instead of closing one and opening the next.
  • ImprovedHovering a benchmark bar on a feed model card now explains what the benchmark measures, instead of repeating its name and score.
  • ImprovedIn a paper's PDF, hovering a highlight you saved now shows its note inline; click it to edit.
  • ImprovedThe OSWorld, DABstep, Arena-Hard, and AlpacaEval leaderboards now refresh automatically every day, instead of only when updated by hand.

June 1, 2026

  • AddedFresh arXiv papers are discovered and added automatically every day.
  • AddedNew benchmarks from AfterQuery (agentic IDE coding, quant trading, web-app generation, finance) and Vals AI (tax, finance, legal, and medical domain evals), with model scores.
  • FixedThe feed could freeze on stale content and stop showing new items; it now updates as soon as new evals, models, tools, or papers land.
  • FixedModel cards no longer show "$0/M" or "0 tok/s" before a model's price and speed have been measured.
  • ImprovedThe feed's paper stream now shows only notable papers.
  • ImprovedPapers in the feed now display their own generated preview cards instead of external thumbnails.

May 30, 2026

  • AddedSubscribe to any feed view by email digest (weekly or monthly) or RSS.
  • AddedRename saved feed views inline.
  • FixedLeaderboard pages no longer scroll sideways on mobile.
  • FixedLeaderboard pages no longer freeze on mobile when a board tracks many charts.
  • FixedPrivate notes no longer zoom the page on iOS or flash a phantom scrollbar.
  • ImprovedModel list shows price, prompt caching, speed, latency, and context as optional inline tags you toggle per field.
  • ImprovedFeed model rows explain each spec (context, price, speed, license) on hover.
  • ImprovedEval pages group similar evals into their own tab.
  • ImprovedLarge leaderboards load faster, revealing more models as you scroll.
  • ImprovedLeaderboards show when their scores were last refreshed, next to how often they update.

May 29, 2026

  • AddedBento landing homepage with per-feature previews.
  • AddedFilterable feed header with saved views.
  • FixedCmd-K search shows the most relevant results first.
  • FixedFeed "New" orders by real release date.
  • FixedHorizontal tab strips no longer scroll vertically.
  • ImprovedModel titles are cleaner and more consistent.
  • ImprovedOrganization pages rank lab models by intelligence and release date.