GPT-5.1-Codex-Max: Long-Horizon Tasks, Native Compaction and 24-Hour Sessions

gpt-5-1-codex-maxcompactionlong-horizonxhigh-reasoningswe-lanceragentic-pod

Sketchnote diagram for: GPT-5.1-Codex-Max: Long-Horizon Tasks, Native Compaction and 24-Hour Sessions

GPT-5.1-Codex-Max: Long-Horizon Tasks, Native Compaction and 24-Hour Sessions

Most Codex tasks complete in minutes. But some tasks — a full microservice migration, a large-scale refactor across 50+ files, a days-long debugging session — run into context limits before they finish. GPT-5.1-Codex-Max was built specifically for these long-horizon scenarios.

What Makes GPT-5.1-Codex-Max Different

GPT-5.1-Codex-Max is the first Codex model trained to operate natively across multiple context windows. This matters because earlier compaction strategies used external scaffolding (the harness summarising older turns and injecting a summary) — the model wasn’t trained to expect those summaries, leading to coherence failures on very long tasks.

With Codex-Max, compaction is trained into the model itself:

  • As the session approaches its context limit, the model summarises essential state: variable names, architectural decisions, outstanding bugs, pending steps
  • It carries that summary into a fresh context window and continues seamlessly
  • This process repeats as many times as needed

Result: OpenAI has observed GPT-5.1-Codex-Max working autonomously for more than 24 hours on a single task — persistently iterating on implementation, fixing test failures, and ultimately delivering a working result.

When to Use Codex-Max

GPT-5.1-Codex-Max is the right choice when:

  • The task is likely to exceed a single context window (large refactors, migrations, multi-file audits)
  • You want to fire-and-forget an overnight or weekend-length job
  • The task requires sustained coherence across many decision points
  • You’re working in a regulated environment where you want complete audit continuity

For everyday tasks — fix this bug, write this feature, review this PR — gpt-5-codex or gpt-5.3-codex are faster and more cost-effective.

Benchmarks: The SWE-Lancer IC SWE Result

GPT-5.1-Codex-Max scores 79.9% on SWE-Lancer IC SWE — a significant jump from GPT-5.1-Codex’s 66.3%.

What is SWE-Lancer?

SWE-Lancer is an OpenAI benchmark of 1,488 real freelance software engineering tasks sourced from Upwork and Expensify, worth $1 million USD in real-world payouts. Tasks range from $50 bug fixes to $32,000 feature implementations. End-to-end tests are triple-verified by experienced engineers.

Why it’s more meaningful than SWE-Bench Verified: it uses real economic value as the scoring unit. A 79.9% result on SWE-Lancer IC SWE means the model can complete the equivalent of ~$798K worth of freelance IC software engineering tasks — compared to 26.2% (Claude 3.5 Sonnet) and 8.6% (GPT-4o) when the benchmark launched in early 2025.

Other benchmark scores (all with compaction + xhigh reasoning):

Benchmark GPT-5.1-Codex-Max GPT-5.1-Codex Notes
SWE-Lancer IC SWE 79.9% 66.3% Real freelance tasks
SWE-Bench Verified (n=500) 77.9% 73.7% Standard synthetic benchmark
Terminal-Bench 2.0 (n=89) 58.1% 52.8% Terminal-native task sequences

The xhigh Reasoning Effort Level

GPT-5.1-Codex-Max introduced a new reasoning effort level: xhigh. This is above high — the model “thinks for an even longer period” before responding.

# .codex/config.toml — for long-horizon overnight tasks
model = "gpt-5.1-codex-max"
model_reasoning_effort = "xhigh"

Key data point: GPT-5.1-Codex-Max at medium reasoning effort already outperforms GPT-5.1-Codex at medium, while using 30% fewer thinking tokens. Use xhigh for:

  • Hard algorithmic problems
  • Security audits requiring deep analysis
  • Tasks where quality matters more than speed
  • Long-horizon runs (the extra thinking pays off over a 24h session)

Cost note: xhigh carries a significant token multiplier (~8–15× vs minimal). For long-horizon tasks, this is usually justified — the model is less likely to go off-track, reducing expensive restarts.

30% Token Efficiency Gain

GPT-5.1-Codex-Max completes equivalent tasks using roughly 30% fewer tokens than its predecessor. For long-horizon runs where you’re burning credits over hours, this is significant.

Combined with cached input pricing (the model reuses previous context where possible), the effective cost of a 24-hour session can be lower than you’d expect from a frontier model.

Windows Support

GPT-5.1-Codex-Max is the first Codex model explicitly trained for Windows environments. If your team includes Windows developers or your CI runs on Windows-hosted runners, this matters for reliability.

Practical Configuration for Long-Horizon Jobs

# .codex/agents/overnight-worker.toml
name = "overnight-worker"
model = "gpt-5.1-codex-max"
reasoning_effort = "xhigh"
developer_instructions = """
You are a long-horizon agent. Your task may span many context windows.
At each compaction boundary:
1. Summarise the current state in a structured note
2. List: completed steps, current blocker, next 3 steps
3. Preserve all variable names, file paths, and architectural decisions verbatim

Never truncate work. If you hit a context limit, compact and continue.
"""

Fire it off via codex exec:

codex exec \
  --agent overnight-worker \
  --dangerously-bypass-approvals-and-sandbox \
  "Migrate the payments service from Stripe v7 to v8 API. All tests must pass. File a PR when complete."

Monitor via the Codex app sidebar — the thread will show compaction events as the session progresses.

When NOT to Use Codex-Max

  • Short, bounded tasks: The model’s compaction overhead adds latency on tasks that would finish in a single context window anyway. Use gpt-5-codex or gpt-5.3-codex-spark for these.
  • Subagent workers: Use gpt-5.4-mini or gpt-5-codex for parallel worker subagents in your agentic pod. Reserve Codex-Max for the orchestrator or for standalone long-running jobs. See Subagent Model Routing: Mini for Workers.
  • Interactive sessions: If you’re steering the agent in real time, shorter context windows keep you in the loop. Codex-Max’s strength is autonomous long-horizon runs.

Model Lineage Context

GPT-5.1-Codex-Max’s place in the lineage:

gpt-5-codex (current recommended)
    ↑ gpt-5.3-codex (better tool reliability, Tau2-bench)
        ↑ gpt-5.2-codex (native compaction, context rollup)
            ↑ gpt-5.1-codex-max (24h sessions, SWE-Lancer 79.9%)

It has been succeeded by gpt-5.2-codex (better native compaction) and gpt-5.3-codex (current recommended for most tasks). But for the specific use case of fire-and-forget long-horizon jobs, Codex-Max remains a valid choice — particularly for teams already optimised around its benchmark characteristics.

Citations