Fieldwork
Which AI model to use for real engineering work
Fieldwork ranks AI models for real software engineering. It combines eight engineering benchmarks with what developers say in 1.8 million votes and real coding sessions, and keeps every number linked to its source.
The overall score below is a formula you can see and change. Every benchmark is still here on its own, with its exact configuration, date and caveats, so nothing is hidden inside the average.
Current tables were checked on 23 October 2026. This update includes Kimi K3.1 Pro and Highspeed, DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Muse Spark Open. Kimi K3.1 Pro appears in the overall ranking with a provisional score based on limited engineering coverage and a preliminary WebDev rating.
Fieldwork ran no model evaluations of its own. You can read the sources, score your own runs, or press ⌘Ktap the island to jump to any model.
Overall Fieldwork Score
By job
The overall score hides how different the jobs are. Each board measures one kind of engineering work, under its own harness, budget and grading. Pick one.
Agents Coding Agent Index 2.0
The same model behaves differently inside different agents. Artificial Analysis averages three jobs into one index; pulled apart, the specialists show. The tick is the index, the shapes are the three jobs.
Index against cost per task
Effort Intelligence Index 4.3.2
Turning up reasoning effort buys points, then stops buying them. Each line is one model climbing from its lowest setting to max. This is Artificial Analysis's general index, not an engineering score, so it never feeds the overall ranking.
What each step buys
Price
Show the numbers
Opinion Checked 23 Oct 2026
What developers say is a separate measure from what tests say. Arena has two kinds: head-to-head votes on website builds, and signals from real coding-agent sessions. Bands are the published uncertainty; where they overlap, the order is not settled.
Blind review, in miniature
History Aider Polyglot · archive
Same agent, different languages
Claude 3.7 Sonnet with SWE-agent on SWE-bench Multilingual. Whiskers are 95% Wilson intervals; the dashed line is the all-language rate. Each language has different repositories, so gaps do not isolate language ability.
Catalog Vercel AI Gateway · 480 IDs
Workbench runs in your browser
Import task-level runs as JSON or JSONL and score them here: first-attempt success, 95% Wilson intervals, cost per success and coverage. Nothing is uploaded. Until you import, it shows six fictional models with synthetic runs.
Where each configuration is strong
Method
Evaluate a configuration, not a name
Model, agent harness, effort, budget and task snapshot are pinned. The overall score keeps each model's best configuration and shows which one it used.
Never hide the parts
Every component of the score is visible per model, missing data is shown as missing, and each board is still ranked on its own.
Show uncertainty
Published intervals stay on the charts. The workbench uses 95% Wilson intervals over unique tasks; repeated attempts are not new tasks.
Count the full cost
Cost per successful task includes the spend on failures. Retries are excluded from the headline, not free.
Keep opinion honest
Votes and the praise-to-complaint balance come from people who chose to use Arena. They measure preference, not correctness, so they form a separate, weighted part. The one session signal that reports an outcome, users confirming the agent finished the job, counts toward engineering and is labelled there.
What a real engineering benchmark should cover
Seven task families, each graded on an outcome you can observe. Three have public Python calibration tasks in the starter pack.