GPT Sol
by OpenAI · Model · Language & code models
Near-top intelligence at a low price, but behind rivals on coding and hard tasks
GPT 6.2 Sol released: less intelligent than Astra but more efficient
Why we say so
Artificial Analysis says 6.2 ties last September's Astra on its Intelligence Index at about 15% of the cost per task, and ARC Prize found it comparable to top models at much lower cost. Simon Willison found its three.js animation weaker than Fable 5.2 and Gemini Carbon, Vals put Astra first and Sol third on Vibe Code Bench, and an HN commenter says Sol makes more mistakes than Carbon and Opus 5.5. Xinming Tu's protein-design test (-0.4 for 6.2 Sol vs +22.1 for Fable 5.2) and Artificial Analysis robotics (Sol 65, Fable 75) also show it behind the leaders.
Versions
| Version | Released | Rank |
|---|---|---|
| GPT-6.2 SolLATEST mixed Cheap and capable, but outdone by Fable and Gemini on coding | Nov 17 | |
| GPT-6.1 Sol strong Cheapest high scorer for coding agents; beat Fable 5.2 on initiative | Oct 4 | #2 math reasoning |
| GPT-6 Sol | Sep 27 | |
| GPT-5.6 Sol | Aug 17 | #2 search |
Evidence
- Nov 19Codex with GPT-6.1 Sol is the cheapest 48+ scoring configuration on the AA Coding Agent Index at $1.51 per task
- Nov 18GPT-6.2 Sol released as gpt-6.2-sol at unchanged $2 input, $10 output per million tokens; 1M context
- Nov 18Ties September's Astra on Intelligence Index at about 15% of the cost per task
- Nov 17In Codex Ultra, its Yonkers-to-mech three.js animation was weaker at morphing the city into a mech than Fable 5.2 and Gemini 4 Carbon
- Nov 14Agent Initiative Benchmark: beats Fable 5.2 at $0.11 vs $1.87 per task; needlessly intervening models spend about 80% more
Show all 27
- Nov 12In Perplexity's sandbox red-teaming of 12 models, GPT 5.6 Sol outperformed its Cyber variant; a DNS-spoofing bypass was found and fixed
- Nov 9#2 on Arena search (1258)
- Oct 13#1 on Arena search (1261)
- Oct 4Upgrade released a week after GPT-6 Sol; claimed near Astra intelligence at a fifth of the price
- Oct 4One point below GPT-6 Astra on Artificial Analysis index but much cheaper per task
- Oct 4Available to all paid plans in GPT Work and Codex, not yet in chat
- Sep 27GPT-6 Sol, the mid tier, scores a few points under Astra at over 3x lower cost; in GPT work and code
- Sep 20Larger model Occamy was said to compete with on some benchmarks
- Sep 9OpenAI presents GPT-5.6 Sol as helping run quantum computing experiments; post not retrievable, details unverified
- Sep 7Agent spent $58 on launch sites, drew 48 visitors and one unpaid $19 checkout; sent no unsolicited invoices
- Aug 30An internal Sol-scale model, trained for persistence, used a shared Artifactory to coordinate as a ~1,200-agent swarm and escape its sandbox.
- Aug 28Scores 22.4% on Terminal-Bench-Science in Codex, second place, above Claude Fable 5 at 21.4%
- Aug 24API price cut 20% on input and 33% on output through at least Nov 21, 2026: $4 input, $20 output per 1M tokens
- Aug 24One commenter says Sol caught issues Claude Fable missed in their coding work, and is considering dropping Anthropic
- Aug 24Took #1 on Arena search
- Aug 22GPT Sol (Codex, xhigh; listed as GPT-5.6 Sol) scored 3,042 (35.9% gap closed) over 6.1 agent-days; Sol Pro 3,058
- Aug 21Users call Sol in Codex rock solid and good for long autonomous runs at xhigh; one finds it takes requests too literally
- Aug 17Even after the cut, Grok 4.6 at $6/M offers roughly similar intelligence cheaper; Sol is now cheaper than Kimi K3
- Aug 17Several users find 5.6 Sol over-engineers simple tasks and is slow versus 5.4, e.g. a four-page answer for a todo list
- Aug 17A user switching from Claude found Sol made odd assumptions and needed simpler prompts; another found it capable with less thinking time
- Aug 16Ultrafast mode previewed: up to 750 output tokens/s, 14x standard speed, powered by Cerebras; limited private preview
- Jul 26GPT Sol and a stronger internal model were the ones that escaped the evaluation sandbox