Pick the job and get the verdict from 11 public benchmarks: who wins, how decisively, the best value, and the best model you can call on FullsendAI today. Every score links to where it was published.
Wins
Too close to call against Claude Fable 5 (1,503): pick either on price or speed.
#5 of 67 at 1,488, the highest-placed model live on FullsendAI.
Try it in the chat1,436 (#32) at $0.11 per million tokens: the most rating for the money.
The 12 best models by overall grade and their place on every benchmark they were tested on. Hover a square for the score.
| Arenas | Agents | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Model and overall gradeModel | Text | WebDev | Coding | SWE Verified | SWE Pro | Terminal 4.0 | Terminal 3.0 | Terminal 2.1 | Terminal 2.0 |
| Claude Opus 5.5AnthropicExcellent 99.5 | – | – | – | – | |||||
| Claude Sonnet 5.5AnthropicExcellent 91.2 | – | – | – | – | |||||
| Claude Fable 5AnthropicExcellent 90.7 | – | – | – | ||||||
| Claude Opus 5AnthropicExcellent 90.0 | – | – | – | ||||||
| Kimi K3Live on FullsendAILiveExcellent 88.7 | – | – | – | – | – | ||||
| Claude Opus 4.7AnthropicExcellent 87.3 | – | – | – | ||||||
| Qwen3.8 MaxQwenExcellent 87.1 | – | – | – | – | |||||
| MiMo V2.6 ProXiaomiExcellent 85.8 | – | – | – | – | |||||
| Seed 2.0 ProByteDanceExcellent 84.3 | – | – | – | – | |||||
| GLM 5.3Z.aiExcellent 83.8 | – | – | – | – | |||||
| DeepSeek V4.1 FlashDeepSeekExcellent 83.5 | – | – | – | ||||||
| Gemini 3.7 FlashGoogleExcellent 83.5 | – | – | – | – | |||||
What people prefer. Two anonymous models answer the same prompt and people pick the better one; ratings move as the votes come in.
3 benchmarks
LMArena’s text arena, overall: real people’s questions of every kind, from writing and research to maths and code, answered by two anonymous models side by side and judged by vote.
Best value: DeepSeek V4 Flash, 1,436 at $0.11/1M
People ask two anonymous models to build the same web app, then vote for the better one. The rating comes from thousands of those head-to-head votes.
Best value: DeepSeek V4 Flash, 1,582 at $0.11/1M
The coding category of LMArena’s text arena: real people’s programming questions, answered by two models side by side and judged by vote.
Best value: DeepSeek V4 Flash, 1,484 at $0.11/1M
Real work, end to end. The model gets a repository or a shell and a goal, and the task’s own tests decide whether it finished.
6 benchmarks
Real GitHub issues from popular Python projects, checked by people to be solvable. The model writes a patch, and the project’s own tests decide whether it worked.
Best value: DeepSeek V4 Flash, 79% at $0.11/1M
Longer, harder software tasks across several languages, from Scale AI. Closer to a day of real engineering work than a single bug fix.
Best value: Qwen3.8 Flash, 62.5% at $0.18/1M
An agent gets a shell and a goal: build software, debug a service, configure a system. Version 4.0 is the current and hardest set.
Best value: Ling 3.1 Flash, 46% at $0.45/1M
The 3.0 Terminal-Bench set: shell tasks an agent has to finish end to end. Scores can’t be compared across versions.
Best value: DeepSeek V4.1 Flash, 30% at $0.30/1M
A revision of the 2.0 set with fixed tasks. Scores can’t be compared with other versions.
Best value: DeepSeek V4 Flash 0731, 82.7% at $0.090/1M
The earlier Terminal-Bench set. Scores run much higher than on 4.0 and can’t be compared with it, but it covers more models.
Best value: DeepSeek V4 Flash, 56.9% at $0.11/1M
Correct code at the first attempt: contest problems written after the models were trained, and editing exercises in several languages.
2 benchmarks
Hard Exercism exercises in C++, Go, Java, JavaScript, Python and Rust, solved by editing files through the Aider coding assistant.
Best value: Qwen3 235B A22B Instruct 2507, 57.3% at $0.20/1M
Competitive-programming problems collected after models were trained, so answers can’t have been memorised. Scored on the first attempt.
Best value: DeepSeek V4 Flash, 91.6% at $0.11/1M
Leaderboard scores are run by each benchmark’s maintainers. Scores marked v come from the model’s maker, often with their own tools and settings, so they’re useful but not strictly comparable. Benchmarks with a weight make up the model rankings; the rest are for reference.
Each benchmark gets a call on how decided its race is, from the gap between first and second:
LMArena Text: real people’s questions of every kind, answered by two models side by side and judged by vote. It carries the most weight in the overall grade.
For agents that change real code, SWE-bench and Terminal-Bench are the closest to the job. For front-end and app building, LMArena WebDev reflects what people actually prefer.
The top two are closer than the benchmark can reliably tell apart: within 5 rating points on an arena, or 1.5 points on a percentage. Either is a fine pick, so choose on price or speed.
They’re relative: every new model and every new vote shifts them a little. We show the date of each rating and refresh them daily.
The score was published by the company that made the model rather than run by the benchmark’s maintainers. It’s shown, but labelled, because setups differ.
Each benchmark sets its own terms, shown beside it. LMArena data is CC BY 4.0: credit LMArena when you reuse it.
Sources: LMArena (CC BY 4.0), SWE-bench, Scale AI SWE-Bench Pro, Terminal-Bench, Aider (aider.chat), LiveCodeBench, and model makers’ published results where marked. Benchmark names belong to their owners.
