GPT-6 Sol
Released Sep 22, 2026 · knowledge cutoff 2026-04
- Status
- Unknown
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1.1M
- Max output
- 128K
- Speed (medium effort)
- 114 tok/s · 1.96s to first answer token
- Price ($/MTok in / out)
- $2 / $10
- Cost per task (medium effort)
- $0.25
- Open weights
- No
Benchmarks1 of 10 reported
LMArena Elo
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Stays on the task it was given instead of spawning side projects — developers who leave overnight agent runs prefer it to Astra's scope creep
- Terse by default and stays out of the way, inheriting 5.6 Sol's concise prose with Astra's plainer wording layered on top
- Roughly half the factual errors of 5.6 Sol on OpenAI's conversation-derived evals, and far fewer misleading claims about coding work it hasn't actually finished
- Streams faster and reaches its first answer sooner than 5.6 Sol at the same effort, so the price cut isn't paid for in latency
Weaknesses
- Burns subscription quota far faster than 5.6 Sol did — Plus and Codex users report exhausting a five-hour window in minutes on demanding prompts
- Function calling in Chat Completions only works with reasoning effort set to none — tool use with any reasoning means moving to the Responses API
- Knowledge-work deliverables regressed against 5.6 Sol — Artificial Analysis traces its GDPval and Briefcase drops to thinner presentation and omitted rubric elements
Middle tier of the GPT-6 family, priced where GPT-5.6 Terra sat (Terra has no GPT-6 counterpart): OpenAI's model docs describe it as a balanced model for interactive and agentic coding, hence tier "balanced" although the GPT-5.6 model of the same name was the flagship. Context 1,050,000 tokens (922K max input), 128K output, cutoff 2026-04-20, and $2/$10 per MTok ($0.20 cached input, $2.50 cache writes) are from OpenAI's model docs; prompts over 272K input tokens bill at 2x input/cache and 1.5x output, a fast tier costs 2x, and OpenAI calls the rates permanent rather than promotional. predecessorId follows OpenAI's own launch wording, "GPT-6 Sol makes about half as many mistakes as its predecessor", set against GPT-5.6 Sol throughout — but GPT-5.6 Sol stays on sale at its promotional $4/$20 and no retirement is scheduled. OpenAI's launch table uses only untracked suites (AutomationBench 1.0.6 33.2% xhigh, Agents' Last Exam 56.4% max, FrontierCode 1.1 49.3% max, DeepSWE v1.1 68.8% max, OSWorld 2.0 offline 60.5% xhigh / 64.4% max, factual error rate 4.5%) and it shipped without a system card or Preparedness rating; OpenAI's stress tests put its deception rate at 1.3% (5.6 Sol 10.4%) and failure-to-disclose at 5.4% (77.8%). hle is null: OpenAI published no HLE figure, and Artificial Analysis' 48% (max; 5.6 Sol reads 49) comes from its own 2,158-question text-only subset whose tool setting its methodology doesn't state, so it can't be pinned to a primary source. AA's Terminal-Bench run is 4.0 (43–44% max, not the tracked 2.1), and its GDPval-AA v2.1 ratings — 1487 max / 1320 medium, down from 5.6 Sol's 1588 — are on a different scale from the v2 figure this key records, so gdpvalAA stays null. AA Intelligence Index 48 at max ($1.06 per task) and 40 at medium, the effort the costPerTask and speed figures are taken at; AA also notes hallucination rate fell from 92% to 60% and 31K output tokens per index run against 5.6 Sol's 29K. LMArena Elo 1456 (±7, 7,574 votes) is the "gpt-6-sol-max" text-arena listing (arena.ai, 2026-09-30 update, rank 66); it is absent from Scale's SWE-bench Pro board, and ARC Prize has published Luna but no Sol results page. Generally available in the API, Bedrock, GitHub Copilot and ChatGPT Work/Codex for paid plans, though not yet in ChatGPT's Chat mode at launch.
For developers
API model strings
- openai
gpt-6-sol
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
Replaces GPT-5.6 Sol