20 benchmark builds, one place to inspect them.

How Astra did it

1. 1.Data method: 18 typed records across 10 labs with four versioned benchmark tracks, per-benchmark scoreSources overrides, documentation and pricing sources, license notes, release dates, and evaluation conditions in each notes field; unverified values are absent, never zeroed.
2. 2.Interaction method: a dependency-free lib/explorer.ts handles conjunctive filters, null-last two-way sorting, the three-model selection cap, and CSV serialization, covered by six node:test cases shipped with the source.
3. 3.Dataset era: SNAPSHOT is 2026-09-04, carried consistently in the metadata, the coverage line, the model dialogs, and the CSV's verified-date column.
4. 4.Verification: the source ships its own QA report (typecheck, lint, build, six passing tests, axe 4.12.1 zero violations, 320-1440px responsive checks); Abundance reran the test suite plus its own lint, typecheck, and build before publishing.

How Fable did it

1. 1.Data method: 31 typed records across 14 labs with release dates, context and output tokens, license objects, availability states, per-million pricing plus a measured cost per task, and seven score fields with full benchmark metadata.
2. 2.Interaction method: the catalog is a server-rendered rank over URL-parsed query state; the job picker, filters, and search all serialize through lib/url-query.ts, and compare selection persists to localStorage via a store with an SSR fallback.
3. 3.Route method: five routes (catalog, models/[id], labs, benchmarks, compare) with per-model pages statically generated from the typed module and unknown ids falling through to a Missing page.
4. 4.Dataset era: a single snapshot constant, DATA_AS_OF = 2026-08-28, is rendered on the catalog header, methodology page, and footer; release dates run through 2026-08-18 and month-precision dates appear on three Qwen records.
5. 5.Verification: the source ships an eight-test catalog suite covering ranking, filtering, search, availability, and context sorting; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Muse Spark 1.3 did it

1. 1.Data method: 13 typed records across 9 labs with release dates, context, pricing (null when undisclosed), open-weights and availability flags, access notes, modalities, use-case tags, and eight benchmark fields including SWE-bench Pro and AA-AnalystAgent.
2. 2.Interaction method: one client pipeline drives search, use-case, lab, and toggle filters, eleven sort modes, view mode, the detail modal, and a three-model compare; the blended overall score is a transparent mean of best-in-lineup ratios with a neutral fill for unreported cells.
3. 3.Dataset era: explicitly dated — 'All figures vendor/lab-reported launch numbers compiled Sep 2, 2026' — the only submission besides Aperture to state a snapshot date.
4. 4.Verification: the source ships with the default create-next-app README and no QA script; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Ling 3.0 Flash did it

1. 1.Data method: 18 typed records across 8 labs with release dates, architecture, params, context, modalities, cached-input pricing, and per-benchmark entries with category and full-mark metadata.
2. 2.Interaction method: one client state pipeline drives search, lab/category filters, and five sort modes; compare is capped at four IDs and available as both a modal and a route.
3. 3.Route method: the board and compare are client surfaces; the per-model page renders from the typed module and guards unknown IDs with a not-found state.
4. 4.Verification: the source ships with the default create-next-app README and no QA script; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Muse Spark 1.2 did it

1. 1.Data method: 20 typed records across 10 labs with family, release month, params, context, modalities, access channels, pricing, arena Elo, latency, and ten benchmark fields.
2. 2.Interaction method: one client pipeline drives search, filters, sorting, view mode, benchmark focus, detail modal, and the compare tray; no routing beyond the single page.
3. 3.Dataset era: newest record is Gemini 2.5 Pro; the module carries no compiled-on date, so the gallery dates it from the release set.
4. 4.Verification: the source ships with the default create-next-app README and no QA script; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Grok 4.6 did it

1. 1.Data method: 34 typed records across 12 labs with release month, status, license, evidence, reasoning mode, context, pricing, speed, composite intelligence/coding/agentic scores, per-benchmark fields, modalities, summary, bestFor, and skipIf guidance.
2. 2.Interaction method: the field board drives search, filters, sorting, and view modes; compare state lives in a client provider mirrored to the URL; the Pareto chart plots composite intelligence against task cost.
3. 3.Route method: the field board, guide, and methodology are single surfaces; compare reads up to four IDs from searchParams; per-model and per-lab pages render statically.
4. 4.Verification: the README documents the route map, sources, and missing-data policy; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Gemini 3.7 Flash did it

1. 1.Data method: 16 typed records across 8 labs in one module, each with tagline, overview copy, category, access type, badge, pricing (including cached-input rates), context window, modalities, and ten benchmark fields.
2. 2.Benchmark set: LMArena overall and coding Elo, SWE-bench Verified, GPQA Diamond, MMLU-Pro, MATH-500, AIME 2024, LiveCodeBench, MMMU, and Tau-bench.
3. 3.Interaction method: one client-side state pipeline drives search, filters, sorting, the wizard, the comparison dock, and modals; the value-frontier scatterplot plots composite quality against blended price.
4. 4.Verification: the source ships with the default create-next-app README and no QA script; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Zcode 5.3 Flash did it

1. 1.Data method: 38 records compiled August 26, 2026 from vendor announcements, API docs, and independent leaderboards; the module header states null scores are never invented.
2. 2.State method: every filter, tab, sort, direction, and selection serializes through history.replaceState and rehydrates from the URL on load, so any board view is shareable.
3. 3.Rank method: overall ranks publish only for models with at least three published benchmarks; sparse sorts fall back to composite with recency tie-breaking.
4. 4.Route method: the board is one client surface; compare reads up to four model IDs from searchParams; per-model pages render statically with primary-source links, caveats, and same-lab siblings.
5. 5.Submitted verification: the source ships a headless QA script and desktop/mobile screenshots; Abundance reran lint, typecheck, build, and preview checks before publishing.

How Zcode 5.3 did it

1. 1.Data method: 32 typed records across 13 labs live in one module with release dates, context windows, in/out pricing, open-weights flags, capability chips, six benchmark fields, and LMArena Elo.
2. 2.Computation method: the overall index is the mean of the four core benchmarks; value score divides the index by blended 3:1 price raised to the 0.25 power so ultra-cheap weak models cannot automatically win.
3. 3.Interaction method: cards and table views consume one filter pipeline; comparison is capped at four IDs; detail and compare render in modals with Escape-to-close and scroll lock.
4. 4.Verification: qa/qa.mjs headless checks (filters, sorting, compare flow, table, modals) plus submitted desktop/mobile screenshots; Abundance reran lint, typecheck, build, and preview checks.

How Gemini 3.6 Flash High did it

1. 1.Build method: Next.js 16.2.11, React 19.2.4, TypeScript, Tailwind CSS v4, lucide-react, and Recharts.
2. 2.Data method: 18 typed models across nine labs include prices, context, speed estimates, modality, reasoning status, benchmark fields, strengths, and workload guidance.
3. 3.Interaction method: grid/table filtering and sorting feed a capped four-model comparison drawer; separate analytics and wizard views reuse the same static model data.
4. 4.Submitted verification: npm run build completed on Next.js 16.2.11 with no compilation or type errors. Abundance reruns release checks before publication.

How Gemini 3.5 Flash High did it

1. 1.Components found: CompareDrawer, RecommendationWizard, FilterPanel, ModelCard, Header, RadarChart.
2. 2.Original data uses Chatbot Arena ELO, MMLU, GPQA, MATH, and HumanEval.
3. 3.Work time reported: 2m 23s across the plan and build phases.

How GLM 5.2 did it

1. 1.Routes found: /, /about, /compare, /models/[id].
2. 2.Components found: Leaderboard, CompareTray, CompareButton, ProviderMark, BenchmarkBars, RadarChart.
3. 3.Work time reported: 23m 30s.

How GPT 5.5 High did it

1. 1.Source module includes sources[] and models[] with benchmark/caution fields.
2. 2.This preview preserves the source app's partial-data posture rather than filling gaps.
3. 3.Work time reported: 11m 40s.

How Composer 2.5 did it

1. 1.Source files inspected under Downloads/Archive/Composer/Grok Composer/src.
2. 2.Original source includes Tailwind 3 and Next 15 dependencies, so the preview is scoped into this site's Next 16 app.
3. 3.Work time reported: 3m.

How Grok Build did it

1. 1.Source files inspected under Downloads/Archive/Grok.
2. 2.Original app uses framer-motion and sonner; this scoped preview preserves behavior with plain React/CSS and an aria-live export status.
3. 3.Work time reported: 5m 27s initial build plus 41s to get it running.

How Codex GPT-5 did it

1. 1.Data method: Terra defines seven model records across seven labs with release, status, context, price text, open-weight state, tags, and normalized 0-100 reasoning, coding, vision, and speed values.
2. 2.Filtering method: a memoized pipeline applies case-insensitive model/lab/tag search plus capability chips for reasoning, coding, multimodal, agents, or open weights.
3. 3.Sorting method: the active reasoning, coding, vision, or speed numeric field is sorted descending; there is no hidden aggregate formula in the executable source.
4. 4.Recommendation method: fixed at-a-glance callouts name Gemini 3.1 Pro as leading overall, Claude Opus 4.7 as best for code, and Mistral Large 3 as best value.
5. 5.Comparison method: Gemini 3.1 Pro and GPT-5.5 start selected and selection is capped at three; the tray supports removal, but its Compare CTA has no handler and produces no side-by-side result.
6. 6.Source method: the submitted report references official Google DeepMind, Anthropic, OpenAI, and Meta pages, while the source records no per-model URL and no normalization procedure.
7. 7.Submitted verification: the completion report records a passing TypeScript typecheck and production build; no elapsed build time was supplied.

How Codex GPT-5 did it

1. 1.Data method: Luna defines seven typed model records across seven labs with release, context, weights, modalities, fit tags, one headline signal, and nullable SWE-Pro, Terminal 2.0, GPQA, BrowseComp, and ARC-AGI 2 fields.
2. 2.Filtering method: a memoized pipeline applies case-insensitive model/lab/family/fit search, exact lab selection, and capability tags for coding, agents, reasoning, multimodal, or open weights.
3. 3.Sorting method: users sort by GPQA/signal, SWE-Pro, context window, or release date; null scores are treated as zero for ordering while still displayed as missing.
4. 4.Comparison method: GPT-5.5 and Gemini 3.1 Pro start selected; users may keep up to three IDs and open a modal comparing published GPQA, SWE-Pro, context, weights, and best-fit notes.
5. 5.Methodology method: the expandable note says values are lab-reported and not normalized, SWE-Pro is not SWE-bench Verified, tool settings matter, and missing values stay blank on purpose.
6. 6.Source method: each record links one official provider release or model page; the footer dates the available-source snapshot to July 15, 2026.
7. 7.Submitted verification: the completion report records a successful production build and responsive browser QA; no elapsed build time was supplied.

How Cursor Grok 4.5 High did it

1. 1.Data method: Grok curated a July 15, 2026 snapshot of 19 models across 11 labs with nullable Arena, AA Index, SWE-bench, GPQA, MMLU-Pro, and HLE fields plus price, speed, context, access, modalities, and use cases.
2. 2.Recommendation method: six presets set the relevant use-case constraint and sort key for frontier, coding, science, value, open weights, or speed; no hidden composite determines the result.
3. 3.Filtering method: ExplorerProvider applies search, lab, use-case, access, open-weight, minimum Arena, maximum price, sort key, and direction to the shared model array.
4. 4.Comparison method: users select up to four models; the tray serializes their IDs into the compare route, which displays score bars and practical specifications side by side.
5. 5.Featured result: Cursor Grok 4.5 is published at AA Index 54 with rank hint #4, $2 input / $6 output per million tokens, 500K context, and null values for Arena, SWE-bench, GPQA, MMLU-Pro, HLE, and speed.
6. 6.Source method: the methodology page links LMArena, Artificial Analysis, SWE-bench, and the GPQA repository and warns that vendor runs differ, Arena moves daily, and critical decisions need primary-source checks.
7. 7.July 15 verification: Cursor and SpaceXAI confirm the joint-training and $2/$6 launch facts; Artificial Analysis confirms score 54 and 500K context. Its live rank has moved since the submitted #4 snapshot, so Abundance publishes #4 only as the artifact's dated result.
8. 8.Submitted verification: the local source contains a completed Next.js production build and route artifacts; Abundance reruns lint, type, tests, build, browser QA, gateway QA, and the production audit before publication.

How GLM 5.2 did it

1. 1.Data method: GLM placed 20 models, nine labs, 10 benchmark keys, price, context, license, modalities, strengths, and release state in typed data modules; null remains the missing-data state.
2. 2.Ranking method: six named benchmark leaderboards compute the top three directly from model fields; a separate best-value list divides the reported benchmark average by a 3:1 blended input/output price.
3. 3.Interaction method: toolbar state filters by search, lab, tier, modality, open weights, and price; users can sort by recency, name, price, context, or any of six surfaced benchmarks.
4. 4.Comparison method: PinProvider persists up to four model IDs in localStorage; CompareView draws a hand-rolled SVG radar and highlights row winners while keeping the metric column sticky on mobile.
5. 5.Build method: Next.js 15 App Router, React 19, strict TypeScript, Tailwind CSS v4, static prerendering, and zero runtime dependencies beyond React/Next in the submitted package.
6. 6.Submitted verification: the completion report records a clean TypeScript check, clean production build, 27 prerendered pages, and 200 responses for all routes. Abundance reruns release checks before publication.

How GPT 5.6 Sol High did it

1. 1.Data method: Sol put nine models from seven labs in a typed module, attached a first-party HTTPS source to every record, preserved missing values, and documented provider-run differences.
2. 2.Decision method: published benchmark fields remain separate from editorial fit scores derived from capability, price, modalities, openness, and deployment tradeoffs.
3. 3.Interaction method: useDeferredValue powers search; pure render-time filters and immutable sorting drive lab, openness, scenario, price, context, and recency views; comparison is capped at three models.
4. 4.Test method: three Vitest checks cover unique IDs and seven-lab coverage, usable price/context/fit ranges, and first-party HTTPS source metadata.
5. 5.Submitted verification: ESLint, TypeScript, three Vitest tests, a production build, responsive Playwright QA, gateway QA/postflight, and a zero-vulnerability production audit. Abundance reruns release checks before publication.