Skip to content
← AI by field
Software engineers and engineering teams

AI for software engineering, right now

Writing, fixing and reviewing code, and running as an agent on whole projects for hours.

Moving: fastUpdated Nov 12, 2026 from 11 sources
Open software in Spezia Free account. No card.
Where AI is in software now

Claude Fable 5.2 leads software engineering benchmarks, with models now solving 93% of SWE-Bench Pro tasks.

What models can do now

  • Claude Fable 5.2 achieves 93% on SWE-Bench Pro, up from 56.4% ten months ago.[1]
  • Claude Fable 5.2 scores 69.9% on FrontierSWE, a benchmark of open-ended technical projects with a 20-hour budget.[2]
  • Claude Sonnet 5.5 reaches 70.6% on Terminal-Bench 4.0, which tests hard terminal-based software tasks.[3]
  • Muse Wonder Full achieves 79% on DeepSWE 1.1, a benchmark using a simple agent setup for software engineering tasks.[4]
  • Qwen3.8 Flash scores 91.9% on LiveCodeBench v6, a benchmark of programming contest problems collected after model training.[5]
  • Claude Fable 5.2 leads LMArena WebDev with an Elo of 1,836, where users vote on web pages built by anonymous models.[6]

Where they still fall short

  • On Alibaba's Co-Design Bench, the best coding agent meets only 69.9% of acceptance criteria and fully completes just 26.4% of tasks.[10]
  • AI coding agents leak secret tokens in 15% of runs and execute untrusted scripts in 24% under swarm pressure, according to a study.[8]
  • External prompt injection is ranked as the top security risk for autonomous AI coding agents.[9]

What is expected next

  • Google DeepMind is moving toward orchestrating specialized agent teams with the launch of /swarm in Google Antigravity.[11]
  • OpenAI says a developer used Codex to build sites and games that drew over 30,000 visitors in under four days.[7]

How fast it is moving

The record on this field's benchmarks over the last year, from Spezia's model leaderboard.

SWE-Bench Pro

Long changes to real codebases in several languages, often across many files, including code from private and copyleft projects that models are unlikely to have seen.

Record: 93%, up from 56.4% 10 months ago (1.6×) (Claude Fable 5.2)

0%25%50%75%100%Claude Fable 5.2Claude Fable 5
Apr 26Jul 26Oct 26

the record so far set a record other models

FrontierSWE

Open-ended technical projects that take hours to tens of hours, from making systems faster to building large programs and applied machine learning research. Each run gets a 20-hour budget.

Record: 69.9%, up from 47.0% 5 months ago (1.5×) (Claude Fable 5.2)

0%20%40%60%80%Claude Fable 5.2GPT-6 Astra
Jul 26Oct 26

the record so far set a record other models

The best models for software now

On this field's benchmarks alone.

  1. 1Claude Fable 5.2
  2. 2Claude Opus 5.5
  3. 3Gemini 4 Carbon
  4. 4GPT-6 Astra
  5. 5Muse Wonder Full
  6. 6GPT-6.1 Sol

Milestones

What became possible, newest first, each marked by how firm it is. Open one for the words it rests on.

Showndemonstrated with published results, a paper or an independent checkClaimedsaid by the company or people who built it, not yet checked by othersReportedreported by others, not confirmed by the company or a published resultNot publicdone by a model that is not released to the public

  1. Nov 10, 2026ReportedGemini 4 Carbon
    Google DeepMind released Gemini 4 Carbon for general usage with gains in software engineering.
    “Gemini 4 Carbon is now available for general usage.” Google DeepMind releases Gemini 4 Carbon for general usage with gains in softwar
  2. Nov 9, 2026ClaimedClaude Fable 5.2
    Anthropic released Claude Fable 5.2, which can coordinate over 1,000 AI agents to build a prototype operating system in 8 days.
    “Fable 5.2 đã phối hợp hơn 1.000 tác nhân AI làm việc liên tục trong 8 ngày, tạo ra hệ điều hành có thể khởi động trên máy tính xách tay” Claude Fable 5.2 có thể tự viết phần mềm trong 8 ngày, lập trình viên đối mặt th
  3. Nov 8, 2026ClaimedRipple-65B Preview
    Ricursive released Ripple-65B Preview, claiming 80.52 on TerminalBench 2.1.
    “Ricursive Intelligence said it released Ripple-65B Preview, reporting 80.52 on TerminalBench 2.1” Ricursive released Ripple-65B Preview, claiming 80.52 on TerminalBench 2.1 and $
Every milestone, story and model Pick a model to see what it can do in software.

Sources

  1. [1]SWE-Bench Pro: best result 93%, up from 56.4% 10 months ago (1.6×)Leaderboard · 2026-11-09
  2. [2]FrontierSWE: best result 69.9%, up from 47.0% 5 months ago (1.5×)Leaderboard · 2026-11-09
  3. [3]Terminal-Bench 4.0: best result 70.6%, up from 55.8% 1 months agoLeaderboard · 2026-09-28
  4. [4]DeepSWE 1.1: best result 79%, up from 68.8% 3 months agoLeaderboard · 2026-10-12
  5. [5]LiveCodeBench v6: best result 91.9%, up from 83.1% a year agoLeaderboard · 2026-08-26
  6. [6]LMArena WebDev: best result 1,836, up from 1,417 a year agoLeaderboard · 2026-11-09
  7. [7]OpenAI says a developer used Codex to build sites and games that drew over 30,000 visitors in under four days.Spezia news · 2026-11-12
  8. [8]Study finds AI coding agents leak secret tokens and run untrusted scripts under swarm pressure.Spezia news · 2026-11-10
  9. [9]Threat model ranks external prompt injection as top security risk for autonomous AI coding agents.Spezia news · 2026-11-10
  10. [10]Alibaba's Co-Design Bench finds best coding agent meets 69.9% of criteria but fully completes only 26.4% of tasks.Spezia news · 2026-11-08
  11. [11]@clmt on XX · 2026-11-10

Other fields