Show what you have. We’ll pick.

Compare with a reference model

Pick a view, a reference and the reasoning levels you care about. The shaded Pareto zone holds variants that beat the reference on both axes, and the dashed line is the Pareto frontier of what is shown. Drag to zoom, click a point to compare all reasoning variants of that model. The task tabs above choose the evidence being compared. The Orchestrator row applies to orchestration evidence. The page address keeps every setting, so you can share the exact view.

The benchmark data could not be loaded. Refresh the page to try again.

Sources: Arena released dataset, Terminal-Bench and models.dev. First source captured . Individual capture times are listed below.

• selected reference
• Pareto zone: better than the reference on both axes
• Pareto frontier
• variant on the frontier
• estimated value (dashed)
• point size = reasoning level, colour = creator

What beats the reference

Key findings

All variants

How to read the numbers

• One selected task uses one source metric. Arena WebDev, factuality, text ratings, agent improvements and Terminal-Bench success rates are separate scales. Scores never become a shared intelligence index.
• Access policy, updated 5 October 2026: OpenAI models use a selected OpenAI/Codex path first; Claude models use a selected Claude path first. Qualified models on direct plans precede Copilot and metered gateways as fallbacks. A selected plan does not prove live availability or remaining quota.
• Default execution pool: Opus, Fable and Astra require explicit escalation through the main agent. Luna is reserved for bounded worker tasks. A curated successor policy keeps older families in reserve and does not transfer scores to new models.
• New model lifecycle: detect in the model catalog, verify exact model ID, effort and plan access, then admit only supported task evidence. Unmeasured models stay in the watchlist. This page refreshes on a verified release; no background refresh schedule is configured.
• The quality pick uses the lower confidence bound. Value admits candidates within 15% of the source cohort's score span below the best available lower bound, then selects the lowest task cost. The tolerance is our selection rule.
• Terminal-Bench cost per task is total published API spend divided by all attempts. Efficiency also considers cost per successful task, including spend on failures. Elapsed time belongs to the named source harness and includes its tools. These trials do not measure NoBrainer.Tech Flow or your subscription.
• Cost per task marked * is an estimate for a shared workload of 75k total input and 25k total output tokens, including reasoning, at models.dev API prices. It is not measured token demand for that model or reasoning effort. Retries, context and tool calls change real costs. Identical reference cost across efforts does not prove identical task cost.
• Unknown time stays unknown. A fastest choice or a triangle point requires a known time value. The per-role triangle (good, cheap, fast; pick one or two) decides how the main agent and the subagents are chosen; a fast choice never promotes an unknown time. There is no invented timing or predecessor score.
• Orchestration uses Arena task outcome improvement as a proxy. Following corrections uses steerability. Neither directly measures prompt writing for subagents or completion of a long project.
• Source results must match the exact model and reasoning variant. Unspecified source effort stays unspecified. Results older than 90 days are excluded; routing exports expire 14 days after the snapshot's first capture. Your current MAIN stays selected.
• Arena Leaderboard Dataset by Arena Intelligence, CC BY 4.0. Changes: task filtering, exact variant joins, percentage display and derived routing rules. Terminal-Bench submissions: Apache 2.0; 2.1 license. models.dev: MIT. Source publishers do not endorse these recommendations.
• NoBrainer created the selection rules. There are no new model test runs or paid benchmark feeds behind this page. Download the dated source evidence.
• Own agent telemetry (marked with a dagger): aggregated medians of seconds and list-price USD per agent action from NoBrainer's own Codex and opencodex logs since 1 September 2026, with no prompts, identifiers or paths. Costs and times use a reference task of 12 worker or 17 orchestration actions, so they compare pace in one workflow, not the same tasks across models. Download the rates.
• Exact source capture times (Europe/Warsaw):