DeepSeek-V4.1-Flash

Leaderboard versions (for example 1.0.5 and 1.0.6) are not comparable. Some results use fallback routing for refused steps: for "Claude Fable 5.1 (with Opus 5 Fallback)", Opus 5 handled 260 of 657 tasks. Artificial Analysis runs a separate leaderboard, AutomationBench-AA, which uses a guardrail-adjusted metric; its scores are not comparable with Zapier's. The partial_credit metric is for diagnostics only. The leaderboard shows no per-row evaluation dates.

The official leaderboard uses mini-swe-agent for all models and shows whole-number percentages with margins of about ±1 to ±6 points. Providers often report scores from their own scaffolds or effort settings; DeepSeek reports 65.5% to 74.2% for the same model depending on the scaffold. DeepSWE v1 scores are not comparable with v1.1. With only 113 tasks, small differences are within noise.

FrontierMath was developed with funding from OpenAI, which has exclusive access to a subset of the problems. Scores from v1 and from Tier 4 (a separate 43-problem expansion set) are not comparable with Tiers 1-3 (v2). The best models score about 94%, so the benchmark separates mid-size models better than it separates frontier models. Epoch's numbers come from its own harness (Python tool, 1M-token limit), which can differ from providers' internal setups. Margins are about ±1.4 to ±2.6 points.

Elo is a relative rating: values change as models are added and the scale is re-anchored, so always store the date a rating was read. For example, Anthropic's launch page gives 1846 for Claude Opus 5.5, while the leaderboard on 2026-10-12 shows 1866 (Max, Default Fallback). Earlier versions (Google cites "GDPVal-AA v2") are not comparable. Judging is done by an LLM, not by human experts. Anthropic models are listed with "Default Fallback", meaning refused steps are routed to a fallback model. AA-Briefcase and GDP.pdf are separate evaluations.

Results differ a lot by setting: with tools versus without tools (Anthropic reports 45.9% without tools and 57.4% with tools for Claude Haiku 5.5), and the full set versus the text-only subset. Providers use different judge models and harnesses (Scale uses o3-mini; Alibaba states GPT-4o). Derived sets such as HLE-Verified (a third-party revision with 1,811 scored items), HLE-Rolling and HLE-Diamond are different item sets and are not comparable. The questions are public, so contamination is possible, and community reviews have reported flawed items.

Scores depend heavily on the agent harness (for example Claude Code, Codex, Grok Build or mini-SWE-agent) and on the reasoning effort, so the same model appears several times on the leaderboard. Leaderboard rows are often submitted by the organisation that builds the agent. Providers' own runs can differ from the leaderboard: Anthropic reports 66.4% for Claude Opus 5.5 at max effort, while the leaderboard lists 64.8% ± 3.1%. Versions 2.0, 2.1 and 3.0 and Artificial Analysis' "Terminal-Bench Hard" use different task sets and are not comparable with 4.0. Margins are about ±3 points.

Scores measure voter preference, which can be influenced by answer style and length. The leaderboard offers a style-control adjustment, so record which view was used. Scores shift with every update, so a value is only meaningful together with its leaderboard date. Entries marked "Preliminary" or with few votes have wide confidence intervals. The same model can appear several times with different reasoning-effort suffixes (for example -high or -max). Category views (Coding, Math, Creative Writing and others) are different rankings from Overall.