VISTA Leaderboard

Leaderboard

#	Model	Harness	Effort	Combined score	S	Tokens	Cost
1	Grok 4.5	cursor-agent 2026.07.01	Med		0.367	161K	$1.83
2	GPT-5.6-sol	codex 0.144 · high	High		0.309	357K	$4.66
3	fable-5	claude code 2.1.152	High		0.274	757K	$12.04
4	GPT-5.6-terra	codex 0.144	High		0.270	98K	$0.68
5	GLM-5.2	claude code 2.1.152 · z.ai	Max		0.269	304K	$3.65
6	Opus 4.8	claude code 2.1.126	High		0.263	602K	$10.45
7	GPT-5.5	codex 0.134 · high	High		0.262	279K	$3.90
8	Sonnet 4.6	claude code 2.1.152	High		0.248	252K	$3.85
9	Opus 4.7	claude code 2.1.126	High		0.246	229K	$6.54
10	Composer 2.5	cursor-agent 2026.06.15	—		0.212	120K	—
11	GPT-5.5	codex 0.134 · medium	Med		0.205	272K	$3.33
12	MAI-Code-1-Flash	copilot 1.0.68	—		0.204	7.4M†	—
13	GPT-5.4-mini	codex 0.134	Med		0.194	646K	$1.83
14	GPT-5.4	codex 0.134	Med		0.190	—	—
15	Gemini 3.5 Flash	antigravity	Med		0.145	—	—
16	Haiku 4.5	claude code 2.1.152	High		0.105	224K	$1.03

Agents are ranked by the Combined score S — DOM-grounded localization × behavior on human-annotated UI anchors, averaged over 10 apps (failures count as 0). Each model runs on its latest harness release (Codex-CLI 0.134–0.144 / Claude Code 2.1.152), with a free choice of stack. Condition C4 gives the agent the richest spec: the page's rendered Figma image (a screenshot mockup) and its pruned Figma structure (the layout tree as JSON) — but no target framework.

Each harness exposes a different reasoning-effort ladder, so levels aren't directly comparable. The highlighted dot marks the level each model ran at; the row length shows the granularity available. Notably, GLM-5.2 ran at its ceiling (max), while the Claude and Codex models still had headroom above the level used.

High · 3 / 6 Claude Code (6 levels) — fable-5, Opus 4.8, Opus 4.7, Sonnet 4.6, Haiku 4.5

High · 3 / 4 Codex (4 levels) — GPT-5.6-sol, GPT-5.6-terra, GPT-5.5 (high runs)

Medium · 2 / 4 Codex (4 levels) — GPT-5.5, GPT-5.4, GPT-5.4-mini

Max · 2 / 2 z.ai GLM-5.2 (2 levels: high · max) — ran at max

Medium · 2 / 3 cursor-agent (Grok 4.5: medium · high · xhigh) — Grok 4.5 (ran at high = "Medium")

Medium · 2 / 3 Antigravity (3 levels: low · medium · high) — Gemini 3.5 Flash

Cursor Composer 2.5 runs at its agent default; its effort ladder isn't published in a directly comparable form.

Cost & speed

Median over the 10 apps — the median is outlier-robust, so a single runaway or oversized task doesn't skew a model's value. Bars are colored by harness.

Harness: ■ Claude Code · ■ Codex · ■ Cursor · ■ z.ai GLM. Same 10 C4 tasks per model. Externally-provided entries (GPT-5.4, Gemini 3.5 Flash) have no local logs, and MAI-Code-1-Flash reports output-only tokens (not comparable), so all three are omitted here.

How tokens are counted: per task we take non-cached input + output — Claude/Cursor: input + cache-creation + output; Codex: (input − cached) + output + reasoning — then the median over the 10 tasks. Cached re-reads (cache_read) are excluded: they bill at ~10% and, summed over a multi-turn harness, mostly track turn count rather than usage (counting them made Claude look ~6× heavier). GLM-5.2's totals come from each run's final result event, since z.ai's endpoint leaves per-message usage empty. † MAI-Code-1-Flash (GitHub Copilot): Tokens shows real input+output — median 7.4M/task (input ~7.3M + output ~98K), captured via the Copilot CLI's OpenTelemetry export (gen_ai.usage.input_tokens/output_tokens). It is not comparable to the other rows and is excluded from the token sort: Copilot re-sends the full, growing context on each of ~55 LLM calls per task and its usage accounting exposes no cache-read discount, so the input side is inflated relative to the non-cached input+output figures above. (Copilot's plain JSONL stream reports only per-message output tokens — the input figure is available solely through the OTel export.)

How cost is estimated: real median API cost per task in USD = each run's actual billed tokens (input, cached, output/reasoning — cache included) × the provider's official published per-token pricing, then the median over the 10 tasks. Claude models use the CLI's own total_cost_usd (Anthropic's real bill). Codex/GPT and z.ai GLM-5.2 are computed from captured tokens × official pricing (OpenAI, z.ai, Anthropic): e.g. GPT-5.6-terra $2.50/$0.25/$15 and GLM-5.2 $1.40/$0.26/$4.40 per 1M input/cached/output. Note GLM's real z.ai bill (~$3.65) is far below the Anthropic-priced figure Claude Code would report (~$9.44). Grok 4.5 (also run through Cursor, which exposes no per-token or per-run cost) uses Cursor's actual billed total from the dashboard — $54.75 across all 3 rounds ÷ 30 task-builds = $1.83/task — rather than a per-token computation. Cost is left blank where even that isn't available: Composer 2.5 (no dashboard total captured; Cursor publishes only a $3/$15 fast-variant rate, no cached-token rate, and the variant used isn't confirmed), MAI-Code-1-Flash (GitHub Copilot bills in AI-credits, not USD, with no cache breakdown), and GPT-5.4 / Gemini 3.5 Flash (no local token logs captured).

Co-evolution

An agent is a model inside a harness, and both ship on their own cadence. Tracking the same model across harness releases shows the two co-evolving — sometimes lifting each other, sometimes regressing. Combined score by npm release date.

Mean C4 combined score (n=10, failures scored 0). Codex-CLI and Claude Code use independent version schemes, shown as separate panels.

How it's scored

Each task asks an agent to build and launch a multi-page web app from a visual spec. We bring the app up in Docker and, for every human-annotated UI anchor, match it to a rendered DOM element — scoring localization L (IoU / distance to the mockup position) and behavior B (interaction-specific browser checks). The headline Combined score is S = mean(L · B) over the critical anchors.