Benchmarks · Coding & software engineering
FrontierSWE
also: FrontierSWE 2, Frontier-SWE
Can a coding agent finish an engineering problem at the edge of what skilled humans can do — a performance optimisation, a research implementation, a system built to an exact specification — given up to 20 hours to work on it?
ProximalReleased April 2026Live
Most coding benchmarks ask for a fix that takes a human minutes or hours. FrontierSWE, from the startup Proximal, is built from tasks that would take a strong engineer days: porting a code formatter between languages, optimising a video-scaling library, writing a filesystem to an exact byte layout. An agent gets up to 20 hours per attempt. Proximal describes the aim as measuring software engineering “at the edge of human ability”.
The first version, in April 2026, had 17 tasks and ran each model in its own company’s agent tool. It ranked models by how often they beat one another rather than by a score. Version 2, released around September, doubled the task set and moved every model into a single harness. It scores each task from 0 to 1 and averages five attempts, which gives one comparable percentage per model. Proximal’s write-up said the shared harness raised scores compared with the vendors’ own tools on a sample of tasks.
On the v2 board, GPT-6 Astra led at 65.5% when Gemini 4 Carbon was added at 61.2% on its launch day. Claude Opus 5.5 and Claude Sonnet 5.5 were between them. It was one of the few rows in Google’s own launch table where Carbon was not first.
The set
Version 2: 34 long-horizon tasks across implementation, performance optimisation, AI research, scientific computing and visual reasoning. Every model runs in the same harness (Proximal's own, called Proximus) at maximum reasoning effort, with a 20-hour budget per attempt and five attempts per task, graded by a clean-room verifier in a separate container. Each task is scored 0–1 and the headline figure is the mean over five attempts (mean@5), averaged across all 34 tasks; the board also shows worst-of-five to best-of-five ranges, cost and time.
Example
Implement a crash-proof NOR flash filesystem in Zig, down to the exact on-disk byte layout.frontierswe.com
Where it stands
A blog-published benchmark rather than a peer-reviewed one, run by Proximal on every model itself. The v1 board ranked models by relative wins in each vendor's own harness, with no percentage score, so it cannot be compared with v2.
How the top score changed hands
FrontierSWE 2current
Current best: GPT-6 Astra — 65.5% Proximal's board, read 29 October 2026: GPT-6 Astra 65.5%, Claude Opus 5.5 62.3%, Claude Sonnet 5.5 61.9%, Gemini 4 Carbon 61.2%, Claude Fable 5.1 56.3%, Gemini 4 Argon 55.0%. Astra is also the most expensive run on the board.
In the timeline · 3 entries
Google releases Gemini 4 Carbon to cyber defenders a month after Argon
Artificial Analysis ranked it first on its new v5.0 index at 59, a point above Claude Opus 5.5; introductory pricing matches Argon at $2/$10 per million tokens.
Models & capabilities · Security & misuse
Alibaba unveils Qwen3.8-Max, its largest model, ahead of open-weight release
2.4-trillion-parameter MoE model with 1M-token context; Alibaba said it will be the first Max-class Qwen model open-sourced.
Open weights & ecosystem · Models & capabilities
Zhipu AI releases GLM-5.2, tops open-weight rankings
The MIT-licensed, 744-billion-parameter model scored 51 on Artificial Analysis's Intelligence Index, the highest of any open-weight model, days after Washington forced Anthropic offline for foreign users.
Open weights & ecosystem · Models & capabilities