🤖 AI 资讯

· · ↗
← 返回列表

Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases

Hacker News大模型,AI应用,具身智能,开源,OpenAI,Google,Anthropic,xAI,月之暗面,智谱,文生图,文生视频,代码生成,Agent智能体,搜索RAG,扩散模型,强化学习,模型评测,提示工程,招聘HR,榜单评测,论文,开发者生态原文 ↗

Introducing Real-SWE

Benchmarking frontier AI models on private, real-world, enterprise codebases.

01Introduction

Today we are releasing Real-SWE, a benchmark that evaluates frontier AI models on private, real-world, enterprise codebases. Each task comes from a private production codebase that we licensed from a real-world company. These are problems their engineers work on, with all the context and complexity that comes with an existing product.

  • Private codebases. Agents must navigate proprietary systems whose code and solutions aren’t available on the public internet.
  • Work with business consequences. Getting billing right, calculating taxes, migrating customers. Changes that affect how a business runs, often across multiple services.
  • Company-specific complexity. Every company has its own rules and ways of writing code. Agents have to understand those conventions and make changes that work with what’s already there.

Can a coding agent actually do the work of a software engineer in the real world?

  • 1Resolution rate: 38.8%Fable 5.1Claude Code
  • 2Resolution rate: 33.8%GPT-6 AstraCodex CLI
  • 3Resolution rate: 31.2%Gemini 3.8 FlashGemini CLI
  • 4Resolution rate: 28.8%GLM 5.3Claude Code
  • =5Resolution rate: 23.8%Grok 4.6Grok Build
  • =5Resolution rate: 23.8%Muse Spark 1.3Muse Code
  • 7Resolution rate: 18.8%Kimi K3Kimi Code
  • 8Resolution rate: 16.2%GPT-5.6 SolCodex CLI
#ModelHarnessResolution rate
1Claude Code

38.8%

2GPT-6 AstraCodex CLI

33.8%

3Gemini 3.8 FlashGemini CLI

31.2%

4Claude Code

28.8%

=5Grok Build

23.8%

=5Muse Code

23.8%

7Kimi Code

18.8%

8GPT-5.6 SolCodex CLI

16.2%

Expert-generated or synthetic tasks can be well designed, but they aren’t the verbatim, actual tasks that engineers in real companies need to do. Our tasks differ on two axes: the underlying coding artifact and specificity of the instruction. Both add complexities that challenge today’s frontier models.

We use native harnesses to reflect how enterprise engineers work in practice, evaluating model-and-harness combinations rather than models in isolation.

Real company tasks require company-specific context

Correct billing depends on business rules and external services

Fix invoice billing so each business charges the right tax and exempt customers aren't taxed.

View full instructionHide full instruction

Billing reopens on Monday and every invoice this service issues is coming out untaxed. Each business on the platform settles its tax a different way: some maintain a rate themselves, some want each invoice priced against the buyer's destination by our tax authority provider, and some collect nothing at all, while a customer we hold an exemption for is charged nothing whichever way its business is configured. Pricing a destination means going to the authority with both addresses, the priced lines and the product category that business sells under, on the sandbox or the production authority according to the account the business is on; an address the authority refuses must be reported without stopping the invoice. The rate, the tax and the gross belong on the issued invoice, and once an invoice is settled the sale is filed back to the authority under that invoice's number so the returns reconcile. Invoices between European parties show both sides' VAT registrations. The authority and ledger are available at TAX_JAR_URL, PROD_TAX_JAR_URL and INFLUX_URL.

  • TaxJar sandbox
  • TaxJar production
  • InfluxDB ledger
  • NestJS service
  • TypeScript

Agents work across code, infrastructure, and business tools

Tools and services across Real-SWE task environments. Each task exposes only the services its workflow needs.

  • AWS emulator
  • Docker
  • Kubernetes
  • GitHub
  • Linear MCP
  • PostgreSQL
  • MySQL
  • MongoDB
  • Gel
  • Redis
  • Go
  • Python
  • Node.js
  • Vitest
  • Slack
  • Intercom
  • Google Drive
  • ClickUp

Codebase Selection

We selected codebases through a rigorous screening process, focusing on real companies with substantial usage, strong engineering teams, and demanding production workloads. The sample tasks analyzed below come from these codebases, including:

  • A Luma/Partiful competitor with 200K+ users and a top 100 App Store ranking
  • A consumer fintech platform processing 100K+ bank statements
  • Enterprise AI sales platforms supporting complex business workflows

We prioritize code written to meet an actual user or business need over code written solely to create a benchmark task. Production engineering requires understanding existing architecture, preserving behavior that users rely on, and making changes within real operational constraints.

Brief instructions can require changes across many files

Our tasks describe the change needed, leaving agents to discover implementation details in the codebase and surrounding tools. Any behavior required by the verifier must be stated or reasonably discoverable. This leads to our prompts being slightly underspecified, about par with DeepSWE and Terminal Bench, but specific enough to not omit instructions.

The work is cross-functional and complex: a single change can span multiple parts of the application. Agents must understand existing business logic and company coding patterns while keeping the surrounding system working.

A typical Real-SWE instruction is 1,742 characters.

  • FrontierCode2,056 chars
  • DeepSWE1,975 chars
  • Terminal-Bench 31,584 chars
  • FrontierSWE v2992 chars
  • Real-SWE1,742 chars

11 files in Real-SWE, compared with 6 in FrontierCode and DeepSWE.

  • FrontierCode6
  • DeepSWE6
  • Real-SWE11

Cognition's published comparison; FrontierCode includes task descriptions and codebase guidelines. We measured instruction files from

Terminal-Bench 3's 74 tasks,

FrontierSWE v2's 34 tasks, and Real-SWE's eight repository-backed sample tasks. Character counts are rounded to the nearest whole character. No comparable files-edited figure is included for Terminal-Bench 3 or FrontierSWE v2.

Models fail even in short rollouts.

71.4% of rollouts under 10 minutes failed, compared with 73.4% of longer rollouts.

Triaging multiple systems and understanding requirements in codebases riddled with existing business logic and coding patterns is difficult.

70/98 failed

398/542 failed

  • Failed
  • Passed

Every task is inspired or lifted verbatim from a private, real-world codebase. We find these types of tasks super interesting for three reasons:

  • Tasks on private codebases are natively out of distribution. These types of coding tasks are not available anywhere on the internet and are unlikely to have ever been trained on by any other ai model. 99% of tokens in real-world enterprises are hidden away from the frontier models.
  • These tasks are economically viable work. Each task here has a direct relationship to spend and was assigned to an engineer earning a salary. Most benchmarks test interesting, experimental capabilities that are often unlikely to be widespread in the real-world.
  • Company-specific engineering patterns matter. Does AI code match the bar of a real-world enterprise? Our results show us that we're far from that reality. Many enterprises care about code standards and patterns. We've found that today's models are weaker at understanding company coding patterns and frequently miss requirements or don't verify their assumptions.

02Analysis

Here's an analysis of a small sample of tasks from our benchmark. If you're interested in the sample, request access here.

6 of 10 tasks have resolution rates below 15%

Select a task to view model results. Percentages show the overall resolution rate.

Multi-region sweep67.2%

  • Fable 5.17/8 passed
  • GPT-6 Astra8/8 passed
  • Gemini 3.8 Flash8/8 passed
  • GLM 5.32/8 passed
  • Grok 4.63/8 passed
  • Muse Spark 1.38/8 passed
  • Kimi K32/8 passed
  • GPT-5.6 Sol5/8 passed

API keys & environments65.6%

  • Fable 5.18/8 passed
  • GPT-6 Astra5/8 passed
  • Gemini 3.8 Flash7/8 passed
  • GLM 5.35/8 passed
  • Grok 4.64/8 passed
  • Muse Spark 1.36/8 passed
  • Kimi K30/8 passed
  • GPT-5.6 Sol7/8 passed

Entitlement overage lines50.0%

  • Fable 5.18/8 passed
  • GPT-6 Astra7/8 passed
  • Gemini 3.8 Flash5/8 passed
  • GLM 5.33/8 passed
  • Grok 4.61/8 passed
  • Muse Spark 1.31/8 passed
  • Kimi K36/8 passed
  • GPT-5.6 Sol1/8 passed

Customer identity migration40.6%

  • Fable 5.13/8 passed
  • GPT-6 Astra1/8 passed
  • Gemini 3.8 Flash3/8 passed
  • GLM 5.34/8 passed
  • Grok 4.68/8 passed
  • Muse Spark 1.33/8 passed
  • Kimi K34/8 passed
  • GPT-5.6 Sol0/8 passed

Billing schedule migration14.1%

  • Fable 5.13/8 passed
  • GPT-6 Astra1/8 passed
  • Gemini 3.8 Flash2/8 passed
  • GLM 5.32/8 passed
  • Grok 4.60/8 passed
  • Muse Spark 1.30/8 passed
  • Kimi K31/8 passed
  • GPT-5.6 Sol0/8 passed

API token metering12.5%

  • Fable 5.11/8 passed
  • GPT-6 Astra5/8 passed
  • Gemini 3.8 Flash0/8 passed
  • GLM 5.31/8 passed
  • Grok 4.60/8 passed
  • Muse Spark 1.30/8 passed
  • Kimi K31/8 passed
  • GPT-5.6 Sol0/8 passed

S3 datastore measurement10.9%

  • Fable 5.10/8 passed
  • GPT-6 Astra0/8 passed
  • Gemini 3.8 Flash0/8 passed
  • GLM 5.33/8 passed
  • Grok 4.62/8 passed
  • Muse Spark 1.31/8 passed
  • Kimi K31/8 passed
  • GPT-5.6 Sol0/8 passed

Linearizable scan4.7%

  • Fable 5.10/8 passed
  • GPT-6 Astra0/8 passed
  • Gemini 3.8 Flash0/8 passed
  • GLM 5.32/8 passed
  • Grok 4.61/8 passed
  • Muse Spark 1.30/8 passed
  • Kimi K30/8 passed
  • GPT-5.6 Sol0/8 passed

Tax jurisdiction3.1%

  • Fable 5.11/8 passed
  • GPT-6 Astra0/8 passed
  • Gemini 3.8 Flash0/8 passed
  • GLM 5.31/8 passed
  • Grok 4.60/8 passed
  • Muse Spark 1.30/8 passed
  • Kimi K30/8 passed
  • GPT-5.6 Sol0/8 passed

Analytics stream reducer0.0%

  • Fable 5.10/8 passed
  • GPT-6 Astra0/8 passed
  • Gemini 3.8 Flash0/8 passed
  • GLM 5.30/8 passed
  • Grok 4.60/8 passed
  • Muse Spark 1.30/8 passed
  • Kimi K30/8 passed
  • GPT-5.6 Sol0/8 passed

Missed requirements are the most common failure

Failures are grouped by observed submission behavior using the same taxonomy across models, following DeepSWE.

GPT-6 AstraGemini 3.8 FlashGPT-5.6 Sol

No model solves every task

One square per rollout: each row is a task, each column a trial, eight trials per task for every model.

GPT-6 AstraGemini 3.8 FlashGPT-5.6 Sol

Different models fail in different ways

Percentages are out of each model's failed runs, not all runs.

Unverified assumption

Builds on a guess about the system instead of checking it in the workspace.

  • GPT-5.6 Sol43.3%: 29 of 67 failed runs
  • GPT-6 Astra34.0%: 18 of 53 failed runs
  • GLM 5.328.1%: 16 of 57 failed runs
  • Grok 4.624.6%: 15 of 61 failed runs
  • Fable 5.124.5%: 12 of 49 failed runs
  • Muse Spark 1.319.7%: 12 of 61 failed runs
  • Kimi K315.4%: 10 of 65 failed runs
  • Gemini 3.8 Flash10.9%: 6 of 55 failed runs
Missed requirement

Leaves out behavior the instruction requires.

  • Grok 4.667.2%: 41 of 61 failed runs
  • Kimi K353.8%: 35 of 65 failed runs
  • GLM 5.338.6%: 22 of 57 failed runs
  • Fable 5.136.7%: 18 of 49 failed runs
  • Muse Spark 1.336.1%: 22 of 61 failed runs
  • GPT-5.6 Sol31.3%: 21 of 67 failed runs
  • Gemini 3.8 Flash29.1%: 16 of 55 failed runs
  • GPT-6 Astra28.3%: 15 of 53 failed runs
Integration error

Right idea, wired into the surrounding system incorrectly.

  • Gemini 3.8 Flash49.1%: 27 of 55 failed runs
  • Muse Spark 1.341.0%: 25 of 61 failed runs
  • Fable 5.134.7%: 17 of 49 failed runs
  • GPT-6 Astra34.0%: 18 of 53 failed runs
  • Kimi K327.7%: 18 of 65 failed runs
  • GLM 5.326.3%: 15 of 57 failed runs
  • GPT-5.6 Sol16.4%: 11 of 67 failed runs
  • Grok 4.68.2%: 5 of 61 failed runs
Regression

Breaks existing behavior while making the change.

  • Gemini 3.8 Flash10.9%: 6 of 55 failed runs
  • GPT-5.6 Sol9.0%: 6 of 67 failed runs
  • Fable 5.14.1%: 2 of 49 failed runs
  • GPT-6 Astra3.8%: 2 of 53 failed runs
  • Muse Spark 1.33.3%: 2 of 61 failed runs
  • GLM 5.30%: 0 of 57 failed runs
  • Grok 4.60%: 0 of 61 failed runs
  • Kimi K30%: 0 of 65 failed runs
Wrong file

Delivers the change somewhere the running application never calls, such as a one-off script.

  • GLM 5.37.0%: 4 of 57 failed runs
  • Kimi K33.1%: 2 of 65 failed runs
  • Fable 5.10%: 0 of 49 failed runs
  • GPT-6 Astra0%: 0 of 53 failed runs
  • Gemini 3.8 Flash0%: 0 of 55 failed runs
  • Grok 4.60%: 0 of 61 failed runs
  • Muse Spark 1.30%: 0 of 61 failed runs
  • GPT-5.6 Sol0%: 0 of 67 failed runs

03Effort & the Frontier

Higher cost does not guarantee a higher resolution rate

  • 1Fable 5.138.8% · $6.96
  • 2GPT-6 Astra33.8% · $4.67
  • 3Gemini 3.8 Flash31.2% · $2.50
  • 4GLM 5.328.8% · $5.12
  • 5Grok 4.623.8% · $3.44
  • 6Muse Spark 1.323.8% · $2.74
  • 7Kimi K318.8% · $3.90
  • 8GPT-5.6 Sol16.2% · $2.65

Estimated rollout costs range from $2.50 to $6.96

RankModelEstimated cost (USD)
1Gemini 3.8 Flash$2.50
2GPT-5.6 Sol$2.65
3Muse Spark 1.3$2.74
4Grok 4.6$3.44
5Kimi K3$3.90
6GPT-6 Astra$4.67
7GLM 5.3$5.12
8Fable 5.1$6.96

Swipe the chart to see all tasks.

View task values

  • Fable 5.134k
  • GPT-6 Astra13k
  • Gemini 3.8 Flash78k
  • GLM 5.368k
  • Grok 4.67k
  • Muse Spark 1.336k
  • Kimi K330k
  • GPT-5.6 Sol12k

04Evaluation Setup

Each agent w