gpt-5.5
Meant to excel at
-
Described by heavy Codex users as fast, consistent, close to instructions and low-rework for routine development, preferred over 5.6 Sol for those tasks.
Practitioner report community.openai.com
Watch out for
-
Retired from ChatGPT-signed-in Codex on 2026-10-14; still available via API.
Vendor claim, relayed vincentschmalbach.com
Benchmarks
Agentic coding percentile 0.42
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 58.6 | 0.42 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-04-24 |
Computer use percentile 0.75
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| osworld-verified | 78.7 | 0.75 | Vendor claim, relayed rits.shanghai.nyu.edu | 2026-04-24 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
A heavy Codex user kept GPT-5.5 as his all-day main model and went back to it after a week with Astra; in his logs Sol used about 2.25x GPT-5.5's tokens per session and ran far longer on the same work. Single user's impressions, no task benchmarks.
Task: agentic-long-running | Compared with gpt-5.6-sol | 2026-09-21
Practitioner report vincentschmalbach.com - +
Several Codex Pro users asked to keep GPT-5.5 as a legacy option, saying Sol over-reasons, widens scope, spawns needless sub-agents and burns quota on ordinary tasks, while 5.5 needs little rework. One poster cites a controlled GitHub run where 5.5 used web search and finished 3 of 3 trials, unverified. Posters admit harness effects may be the cause.
Task: chore | Compared with gpt-5.6-sol | 2026-09-15
Practitioner report community.openai.com - +
Several forum users treated 5.5 as the stable coding model after 5.6 Sol felt unstable, wanting it kept available; one recommends 5.5 High for bounded implementation alongside 5.6 Sol as planner/reviewer.
Task: feature | Compared with gpt-5.6-sol | 2026-07-15 older than 90 days
Practitioner report community.openai.com - -
Users comparing against 5.6 say Luna medium is about on par with 5.5 medium in intelligence at lower cost, and Terra Ultra gave roughly twice 5.5's quality and speed (anecdotal, relayed).
Task: feature | Compared with gpt-5.6-luna undated
Practitioner report go.tabbit.ai