How close is AI to replacing programmers?

As of October 2026, GPT-6 Astra and Gemini 4 Argon score 100% on this year's programming olympiad problems, where the best student scored 83%, and Claude Opus 5.5 gets 6 in 10 real code changes to pass every required check.

100%

GPT-6 Astra and Gemini 4 Argon on this year's programming olympiad problems

GPT-6 Astra and Gemini 4 Argon100%
Best of 375 students83%

Vals AI, Oct 2026 · IOI 2026 results, Aug 2026

3.1

workdays of AI agent time for every workday of OpenAI's researchers

OpenAI, Sep 2026

6 in 10

real code changes where Claude Opus 5.5 passes every required check

Claude Opus 5.5 · FrontierCode, Sep 2026

How much code does AI write?

90%

of the code in Claude Code, Anthropic's coding app, is written by Claude.

Anthropic · Oct 2026

Share of new code written by AI

Claude CodeAnthropic's coding app · Oct 2026about 90%
Anthropicall code merged · May 2026over 80%
Googleall new code · Apr 202675%

Anthropic merges over 10 times as much code a day as in May

May 2026code changes merged a day2,500
Sep 2026code changes merged a day30,000+

Claude Code's own team: merged changes more than doubled from June to September, and the share later undone fell by more than 40%.

Whole programs rebuilt without seeing their code

77%
Claude Opus 5.577%

100% = all 30 tasks passed (15 programs, each in 2 languages) · Claude Opus 5.5 · Sep 2026 · MirrorCode

#ModelShare of tasks rebuilt
1Claude Opus 5.577
2Claude Fable 5.173
3Claude Fable 564
4GPT-6 Astra47
5Claude Opus 4.731

METR time horizon: the longest coding job AI finishes alone

Since 2023, the coding jobs AI finishes alone have doubled in length about every 4 months

1 minute10 minutes2023202420252026Length of job, in expert time1 hour1 work day (8 hours)GPT-4 · under 1 minute · Mar 2023GPT-4 Turbo (Nov 2023) · under 1 minute · Nov 2023Claude 3 Opus · under 1 minute · Mar 2024GPT-4 Turbo · under 1 minute · Apr 2024GPT-4o · 1.3 min · May 2024Claude 3.5 Sonnet (Jun 2024) · 1.7 min · Jun 2024o1-preview · 4.4 min · Sep 2024Claude 3.5 Sonnet (Oct 2024) · 2.6 min · Oct 2024o1 · 7.1 min · Dec 2024Claude 3.7 Sonnet · 12 min · Feb 2025o3 · 30 min · Apr 2025Claude Opus 4 · 20 min · May 2025Claude Opus 4.1 · 23 min · Aug 2025GPT-5 · 38 min · Aug 2025Gemini 3 Pro · 54 min · Nov 2025GPT-5.1-Codex-Max · 51 min · Nov 2025Claude Opus 4.5 · 49 min · Nov 2025GPT-5.2 · 1.1 hours · Dec 2025Claude Opus 4.6 · 1.2 hours · Feb 2026GPT-5.3-Codex · 55 min · Feb 2026Gemini 3.1 Pro · 1.5 hours · Feb 2026GPT-5.4 · 54 min · Mar 2026Claude Mythos Preview (early) · 3 hours · Apr 2026GPT-4 · under 1 minute (Mar 2023)Claude Mythos Preview (early) · 3 hours (Apr 2026)AGI Meridian1 minute10 minutes2023202420252026Length of job, in expert time1 hour1 work day (8 hours)GPT-4 · under 1 minute · Mar 2023GPT-4 Turbo (Nov 2023) · under 1 minute · Nov 2023Claude 3 Opus · under 1 minute · Mar 2024GPT-4 Turbo · under 1 minute · Apr 2024GPT-4o · 1.3 min · May 2024Claude 3.5 Sonnet (Jun 2024) · 1.7 min · Jun 2024o1-preview · 4.4 min · Sep 2024Claude 3.5 Sonnet (Oct 2024) · 2.6 min · Oct 2024o1 · 7.1 min · Dec 2024Claude 3.7 Sonnet · 12 min · Feb 2025o3 · 30 min · Apr 2025Claude Opus 4 · 20 min · May 2025Claude Opus 4.1 · 23 min · Aug 2025GPT-5 · 38 min · Aug 2025Gemini 3 Pro · 54 min · Nov 2025GPT-5.1-Codex-Max · 51 min · Nov 2025Claude Opus 4.5 · 49 min · Nov 2025GPT-5.2 · 1.1 hours · Dec 2025Claude Opus 4.6 · 1.2 hours · Feb 2026GPT-5.3-Codex · 55 min · Feb 2026Gemini 3.1 Pro · 1.5 hours · Feb 2026GPT-5.4 · 54 min · Mar 2026Claude Mythos Preview (early) · 3 hours · Apr 2026under 1 minute3 hoursAGI Meridian

A straight line through each new record since 2023 doubles about every 4 months.

METR stopped updating this chart on 8 Sep 2026; its newest model, added May 2026, is Claude Mythos Preview (early).

How close are we to AGI? →

Sources (9)
  1. Vals AI, vals.ai/benchmarks/ioi
  2. International Olympiad in Informatics, stats.ioinformatics.org/results/2026
  3. OpenAI, Research acceleration: The view inside OpenAI (2026-09-06), Aug 2026
  4. Mercor, mercor.com/apex/partner-benchmarks/frontiercode-leaderboard
  5. claude.dev, claude.dev/blog/solving-code-review-at-anthropic
  6. Anthropic, When AI builds itself: Our progress toward recursive self-improvement, and its implications, Mar 2024
  7. Google, Sundar Pichai shares news from Google Cloud Next 2026, Apr 2026
  8. Anthropic, anthropic.com/engineering/redesigning-source-control
  9. Epoch AI, epoch.ai/benchmarks/mirrorcode