Compare AI models in 3D.

Sources: Each axis draws from one named source at a time, never a blend. Artificial Analysis runs one harness over all 88 configurations, so its evaluations are internally comparable; since 8 October that includes CritPt, 71 unpublished research-level physics problems on which every model scores low (32% at best). AA now times every configuration on the atlas. An untimed one would be placed by a separate, labelled visual source at 0 below the positive log scale; that zero is not a measurement and is excluded from Pareto comparisons and the price cloud. The Terminal-Bench 4.0 official board and DeepSWE v1.1 are vendor-submitted runs on each model’s native agent, published with a 95% interval, and reach 4 and 11 configurations respectively. Epoch AI contributes five feeds: the Capabilities Index for 60 configurations, and its own runs of GPQA Diamond (15), Mock AIME (17), FrontierMath tiers 1-3 (15) and SimpleQA Verified (14), each with the standard error of that run. LMArena contributes ten leaderboards: the style-controlled text arena (26), coding prompts (26), maths prompts (24), expert prompts (26), votes scored for factuality (26), the WebDev arena (27), and the agent arena (24) with three of its five signals: invented tools (24), bash recovery (21) and steerability (24). BridgeBench contributes ten editorial ratings of 30 coding models — overall and nine axes: reasoning, front end, back end, one-shot completeness, security, trustworthiness, efficiency, speed and cost — one rating per model, so every reasoning variant of a model shares it (51 configurations; rows run without reasoning are left out); its Speed and Cost are judgements of time and spending to a useful result, not tokens per second or a price. Vals AI contributes six boards from domains no other source here reaches, each one its own run on its own harness: legal research (29 of the 88 configurations), legal agent work over real documents (24), LegalBench (16), tax (14), medical billing codes (23) and the work of a financial analyst (29). A standard error is published beside every value except on the legal agent and medical billing boards, whose pages stopped serving it in October 2026. Price is carried by three independent sources — Artificial Analysis for 88, models.dev for 72 and LiteLLM for 65 — and the context window by models.dev for 74 and LiteLLM for 67. Swapping a source can therefore remove models the new source has never measured, and the coverage figure next to each picker says how many remain. Where a submitted run used a different reasoning effort from the row it lands on, the selected-model panel states the mismatch instead of hiding it. Full terms are on the sources and licences page.

Axes: Output speed always holds the Y axis, and price per million output tokens is the cloud and the colour ramp; you choose X and Z. Three large buttons above the chart turn it to one kind of evidence at a time: Artificial Analysis (its software-engineering knowledge index against its Terminal-Bench 4.0 run), BridgeBench (one-shot completeness against overall, with no price limit dimming the field) and EvalMapBench (solved against Terminal-Bench 4.0). The pickers stay free, and either coding benchmark — Terminal-Bench 4.0 or DeepSWE v1.1 — can take an axis. The three law axes are Vals AI’s own runs. No legal figure from Artificial Analysis appears anywhere here: the AA-Omniscience law domain and the AA Legal Index are served only as an encrypted client-side payload whose extraction their terms forbid, and that payload is still untouched.

Two things worth reading twice: the agent arena publishes an IPS score from LMArena’s causal tracing of live sessions rather than a Bradley-Terry rating: how much a session’s outcome changes when this model runs the agent, against a baseline LMArena describes as typically a uniform mix of models. Zero is that baseline. The four agent scores are differences in outcome, not votes, so compare them only within one agent board. And models.dev lists the same model at many providers, including free and resold ones; the price shown is the model owner’s own listing where there is one, with the full spread of listings recorded beside it.

Green plane: A bar on the Z axis, and yours to move. It starts at 40 % on Terminal-Bench 4.0 and on DeepSWE, at 150 on Epoch’s scale because that is where GPT-5 sits, at 0 on the knowledge indices because that is where errors cancel correct answers, and on each Vals AI board at a round number near the middle of that board’s field, because each of them has a ceiling of its own. The slider says how many measured configurations reach it. It is decoration only: no value, no colour and no frontier changes when it moves.

Pareto frontier: No other configuration is at least as good on all three axes and at least as cheap, and better on at least one. The frontier is recomputed whenever you change an axis, so the diamonds always describe what is on screen.

Epoch Capabilities Index: Not a percentage and not comparable to the pass rates on this page. It is a fitted latent scale anchored so that Claude 3.5 Sonnet sits at 130 and GPT-5 at 150, with no upper bound. Its interval is a 90 % bootstrap over 500 resamples, narrower in kind than the 95 % intervals the benchmark boards publish. Epoch fits one score per model rather than one per reasoning effort, so every effort variant of a model carries the same figure — 45 of the 60 rows inherit a score that way, and each says so. It is also not independent of the rest of this page: Epoch’s fit consumes DeepSWE and Terminal-Bench results for many models, so reading ECI against those axes is partly circular.

Human preference: Six LMArena boards come from blind paired votes: the style-controlled text arena, coding prompts, maths prompts, expert prompts, factuality and the WebDev arena. They are the only numbers here that measure what people preferred rather than whether a task passed. The text arena reaches 26 of the 88 configurations. Arena identifiers are matched in three deterministic steps — the identifier is a row’s display name without its effort bracket, or else its slug, and the name wins because a slug can hide a checkpoint (Qwen3.8 Max and Qwen3.8 Max (0902) are two rows, and the arena lists them apart); names are compared word by word, and failing that as letters and digits alone, so “Qwen 3.8 27B” is qwen3.8-27b; where several rows share a name, one per effort, the entry goes to the one that states no effort, else to the row whose slug it is, and a slug ending in -reasoning, -thinking or -non-reasoning counts as stating one; or it splits into a base plus an effort level the row states; or it names the only effort variant the arena publishes for a row whose effort Artificial Analysis leaves unstated — with no fuzzy matching and no alias table. Each row takes its best match, so the order of a board changes nothing, and a row with two equally good matches gets neither. Where only one side names an effort — LMArena lists qwen3.8-27b with none, Artificial Analysis measured that row at xhigh — the value says so in a caveat built from the dataset alone: 26 values on 3 rows. The interval is the half-width of the bounds LMArena publishes, and it is a 95 % one: the agent arena states that level, and in every Bradley-Terry table each bound sits exactly 1.96 standard errors from the rating.

Vals AI: Six boards in professional domains: two agentic legal ones — research across eight areas of US law against rubrics written by practising firms, and Harvey’s benchmark of legal work done over documents, spreadsheets and presentations with file tools — plus LegalBench, tax, medical billing codes and an agent doing the work of a financial analyst. Vals runs every evaluation itself and publishes the standard error and the reasoning effort of each run, which is how these values land on the right row. Four things to know before reading them. The benchmarks are not all theirs: the agent board and its scoring are Harvey’s, LegalBench is Stanford CodeX’s dataset and the coding set is Protege’s, with Vals as the one who ran the models. Legal research, tax, medical coding and the finance board are graded on held-out task sets Vals does not publish, so those four cannot be reproduced from outside. The two agentic legal boards sit near their floor — half the field scores under 4 % on the Harvey board and a refusal counts as a failure — while LegalBench is saturated at the other end, the whole frontier within four points of 89 %, so read the first for spread and the last for a floor. And Vals scores one reasoning effort per model, so a board reaches at most one row of each model family: that is the practical ceiling of its coverage, not a shortfall. Their cost per task and latency per task are not carried here, because neither is the same measurement as price per million tokens or tokens per second. Their coding and knowledge boards are left out for the same reason: they would put a second, differently harnessed number under a benchmark name this page already carries.

Terminal-Bench 4.0: The two sources are separate runs of the same 66 tasks, not copies of each other. Artificial Analysis scores pass@1 over 198 trials on its own harness; the official board scores 330 trials on the vendor’s own agent. At maximum effort they agree within a few points, but at lower efforts they can differ by ten points or more, and the board only carries configurations a vendor chose to submit.

Speed: Median output speed across the providers serving each model, as AA’s Data API gives it on the snapshot’s day, measured with approximately 10,000 input tokens. Intelligence Index v5 is a composite benchmark score, not human IQ; it already includes Terminal-Bench and Omniscience, so plotting them against each other is partly self-referential.

Knowledge: The AA-Omniscience Index runs from −100 to 100: correct answers increase the score, errors lower it, abstentions incur no penalty. A negative score means more errors than correct answers. The all-domains index covers six domains including humanities, law, business and software engineering; the SWE index is the software-engineering domain alone, and accuracy is the raw share of correct answers before the error penalty. The all-domains index and accuracy are available for every configuration, and the SWE index for all 88 as well: since 6 October the snapshot comes from AA’s Data API, whose per-model detail carries the software-engineering domain of AA-Omniscience directly, so no value is carried from an older extract or read off a chart any more.

Price limit: Price is not an axis. The slider inflates a translucent cloud around every configuration at or below the limit, so the shape itself is the answer to “what can I afford here”. Models above the limit stay visible, drawn darker and greyer. The default limit is $6 per million output tokens.

Price: USD per million output tokens. Input costs and the number of reasoning tokens used are not compared here. Artificial Analysis reports one price per configuration; five of them are $0, which is why the axis is symmetric-log rather than log. LiteLLM instead lists every provider it knows resells a model, so its value is the cheapest of those and the panel reports how many providers and how wide the spread was. The two disagree for open-weight models in particular, where AA often records $0 and a hosted provider charges for serving.

Attribution: Source: Artificial Analysis (artificialanalysis.ai), Terminal-Bench 4.0 (Apache-2.0), DeepSWE v1.1 (Apache-2.0) and LiteLLM (MIT), Epoch AI · Epoch Capabilities Index (CC BY 4.0) and LMArena (CC BY 4.0), Vals AI (vals.ai) and BridgeBench (bridgebench.ai), which publish no licence and no terms of use — their values are republished with attribution as a matter of practice, and the benchmark owners named above keep their own. Benchmark questions and answers remain the property of their creators, and externally sourced results keep their original terms. None of these projects endorses EvalMap, is affiliated with it or has reviewed it.