OpenAIPrevious version · GPT-6.1 Sol
GPT-6 Sol
Half GPT-5.6 Sol's promotional price at $2/$10 per million tokens, but no measurable gain over it on independent data: Humanity's Last Exam (47.9% vs 49.5%), LiveBench (79.3 vs 81.0) and Agents' Last Exam (32.2 vs 30.6) all land inside their noise bands. AA's own Codex-agent DeepSWE run (69.0%) and vals.ai's Terminal-Bench 2.1 run (83.15) use different harnesses from GPT-5.6 Sol's rows.
GPT-6 Sol benchmarks and pricing, every number sourced: 8 tracked GPT-6 Sol benchmark scores (7 independently run, 1 still resting on a vendor’s own claim), priced at $2.00 per million input tokens and $10.00 per million output.
- Released
- 2026-09-22
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- 2026-04-20
- Verified
- sources checked 2026-09-23–2026-10-25
- Succeeded
- GPT-5.6 Sol (2026-07-09)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
GPT-6 Sol’s verified record
GPT-6 Sol’s most-compared rival is GPT-6 Luna: 4 leads, 1 tie, and 3 not callable across their 8 shared comparisons. GPT-6 Sol is priced at $2.00/$10.00 per 1M tokens (in/out) vs GPT-6 Luna’s $0.10/$0.50.
Against the 200 head-to-head comparisons GPT-6 Sol shares with other tracked models:
A gap counts for GPT-6 Sol only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — GPT-6 Sol trails on 16 of them.
HLE · no tools Reasoning
±2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseARC-AGI-2 · max Compositional visual reasoning
±9.2 is noiseTerminal-Bench 2.1 Terminal ops
±10.6 is noiseAgents' Last Exam Professional work
±3.2 is noiseNo verdict for GPT-6 Sol anywhere on DeepSWE ( every independently confirmed comparison inside the noise band); OSWorld 2.0 · v2026 08 08 offline ( nothing independently confirmed on both sides).
- No real gap yet in any of GPT-6 Sol’s coding comparisons — the independently confirmed ones all sit inside the noise band.
What changed from GPT-5.6 Sol to GPT-6 Sol
The 8 benchmarks both models have been scored on, using the same variant each time. A raw GPT-6 Sol gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | GPT-5.6 Sol | GPT-6 Sol | Change | Verdict |
|---|---|---|---|---|
| HLE | 49.49 | 47.9 | -1.6 | Tie |
| Terminal-Bench 2.1 | 88 | 83.15 | -4.8 | Setup-dependent |
| DeepSWE | 73 | 69 | -4.0 | Setup-dependent |
| Agents' Last Exam | 30.6 | 32.2 | +1.6 | Tie |
| ARC-AGI-2 | 92.5 | 89.6 | -2.9 | Tie |
| LiveBench | 81 | 79.3 | -1.7 | Tie |
| Terminal-Bench 4.0 | 37.88 | 44.44 | +6.6 | Tie |
| OSWorld 2.0 | 64.13 | 64.43 | +0.3 | Unverified |
GPT-6 Sol API pricing
$2.00 in / $10.00 out per 1M tokens — official pricing source
What GPT-6 Sol costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.300 |
| A codebase review | 1,000K / 100K | $3.00 |
| A day of agent work | 10,000K / 1,000K | $30.00 |
Computed from GPT-6 Sol’s list rates above — cache discounts and batch tiers are not applied.
GPT-6 Sol is one of 8 OpenAI models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| GPT-6 Cyber | $20.00 | $100.00 |
| GPT-6.1 Terra | $1.00 | $5.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
| GPT-6 Sol(previous version) | $2.00 | $10.00 |
| GPT-6 Astra | $10.00 | $50.00 |
| GPT-5.6 Luna(previous version) | $0.20 | $1.20 |
| GPT-5.6 Sol(previous version) | $4.00 | $20.00 |
GPT-6 Sol benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| DeepSWE[2] Long-horizon coding · ±9.5 is noise | |
| LiveBench[3] Composite score across 7 domains · ±2.7 is noise | |
| Terminal-Bench 2.1[4] Terminal ops · ±10.6 is noise | |
| Agents' Last Exam[5] Professional work · ±3.2 is noise | |
| ARC-AGI-2(max)[6] Compositional visual reasoning · ±9.2 is noise | |
| Terminal-Bench 4.0(max)[7] Terminal ops · ±12.4 is noise | |
| OSWorld 2.0(v2026 08_08_offline)[8] Long-horizon computer use · ±9.7 is noise |
Who ran these numbers: 7 of 8 independent — artificialanalysis.ai (2), vals.ai (2), livebench.ai (1), snorkel.ai (1), arcprize.org (1); vendor self-reported (1).
- HLE: AA's own run of GPT-6 Sol at max effort, text-only 2,158-question subset.
- DeepSWE: AA's own DeepSWE v1.1 run via its Codex agent harness (max effort) — part of AA's Coding Agent Index, not deepswe.datacurve.ai's own board, which does not list this model yet. OpenAI's self-reported figure is 68.8% (also max effort), within a point of AA's independent run. OpenAI's launch chart self-runs DeepSWE at 68.8 max — within a point of AA's independent 69 stored here.
- LiveBench: Board row 'GPT-6 Sol Max Effort' on the 2026-06-25 LiveBench release.
- Terminal-Bench 2.1: vals.ai's own Terminus 2 run, board row 'GPT-6 Sol' (83.15 ± 1.30, pass@1, reasoning effort max), from its archived Terminal-Bench 2.1 table.
- Agents' Last Exam: Overall pass rate 32.2 (Codex, XHigh, the board's best effort for this model; score 56.5; est. cost $318). Per-effort results: XHigh 32.2, Max 31.6, High 31.6, Medium 28.9, Low 27.6. OpenAI's launch chart self-runs this benchmark at 56.4 max effort (linked as 'Agents\' Last Exam V1', task split unstated) — far above the official board's 32.2 stored here, and not stored as a row.
- ARC-AGI-2: ARC Prize's own leaderboard row 'GPT-6 Sol (Max)', dated 2026-09-22: 89.6% at $0.439/task, the top of the model's five populated tiers.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 44.444 at effort max. Artificial Analysis same-conditions run (Max): 43.94. Official board (grant-funded, Codex, max effort, 5-trial pass@1): 49.39 ± 3.23 CI — 5.0 points above this run.
- OSWorld 2.0: Launch chart 'OSWorld 2.0, offline set' at max effort; chart footnote: partial reward on the offline set from the v2026.08.08 release. The xhigh run (60.54) is the one the prose highlights.
Notes on the record
Pricing is tiered by prompt length, confirmed on developers.openai.com/api/docs/models/gpt-6-sol: $2 input / $10 output per million tokens under 272,000 prompt tokens (cached input $0.20), rising to $4/$15 (cached $0.40) — 2x input and cache, 1.5x output, billed on the full request — once a prompt crosses that threshold. OpenAI frames this as a standard price, not a promotion: its own comparison table reads "$4→$2 input, $20→$10 output, 50% cheaper" against GPT-5.6 Sol, but explicitly calls GPT-5.6 Sol's $4/$20 rate "promotional pricing," so the headline "50% cheaper" compares against a temporary price, not GPT-5.6 Sol's original one.
Context window 1,050,000 tokens (max input 922,000), max output 128,000. Input: text and images. Output: text only. Knowledge cutoff April 20, 2026. License not stated for GPT-6 Sol; inferred proprietary, API-only.
No OpenAI page calls this a replacement for GPT-5.6 Sol — the announcement uses "predecessor" and "counterparts" when comparing scores, never "supersede" or "replace," and GPT-5.6 Sol carries no deprecation notice.
AA's own Codex-agent run of DeepSWE v1.1 (69.0%, max effort) is close to OpenAI's self-reported 68.8% (also max effort) — independent and vendor figures agree within a point. OpenAI also self-reports 56.4% on Agents' Last Exam without naming the split or metric; it sits close to Snorkel's partial-credit score for the model (56.5), not the pass rate this site tracks, so it is not recorded here. Snorkel's own board now lists GPT-6 Sol at 32.2 Overall (Codex, XHigh, checked 2026-10-01), the row recorded here. vals.ai's Terminal-Bench 2.1 run, 83.15 at max effort, was added the same day; on that board GPT-5.6 Sol reads 85.77, a tie.
Update 2026-09-30: OpenAI shipped GPT-6.1 Sol as "an upgrade to GPT-6 Sol," and this model's API page now points to it as "the newer Sol model." The deprecations page has no GPT-6 Sol entry and its price is unchanged, so it stays current here.
Compare with
FAQ
Has GPT-6 Sol been independently benchmarked?
Yes — every benchmark this page tracks a score for, seven of the fourteen this site tracks, carries an independently-run number as of 2026-10-08: Humanity's Last Exam (47.9%, Artificial Analysis, max effort), DeepSWE v1.1 via AA's own Codex-agent run (69.0%, max effort) and LiveBench (79.3 overall, max effort) from the day after launch, vals.ai's Terminal-Bench 2.1 run (83.15, max effort) and Snorkel's Agents' Last Exam (32.2 overall pass rate, Codex, XHigh), both recorded 2026-10-01, plus ARC-AGI-2 (89.6% at Max, ARC Prize's own row), and vals.ai's Terminal-Bench 4.0 (44.44), added 2026-10-08. GPQA Diamond, SWE-bench Verified, LiveCodeBench, Toolathlon-Verified and HMMT Feb 2026 have no row — this page tracks no score for them.
Is GPT-6 Sol better than GPT-5.6 Sol?
Not on anything this site can measure independently yet. On the two benchmarks where both have a score from the same evaluator at the same max effort, GPT-6 Sol comes out slightly lower — Humanity's Last Exam 47.9% versus 49.49%, LiveBench 79.3 versus 81.0 — and both gaps sit inside the benchmark's noise band, so they read as ties, not a regression. Agents' Last Exam, where Snorkel shows both at XHigh, is a tie the other way: 32.2 versus 30.6. OpenAI's improvement claims rest on evaluations this site doesn't track, including its internal factuality test. What clearly changed is price: $2/$10 versus $4/$20.
Is GPT-6 Sol's price really 50% lower than GPT-5.6 Sol's?
Only against GPT-5.6 Sol's temporary rate. OpenAI's own comparison table shows $4→$2 input and $20→$10 output, but the same announcement calls GPT-5.6 Sol's $4/$20 rate "promotional pricing." GPT-6 Sol's $2/$10 is presented as OpenAI's new standard price, not a further promotion — so the "50% cheaper" framing compares against a price that was itself already discounted.
Does GPT-6 Sol's price change with longer prompts?
Yes. The $2 input / $10 output per-million-token rate applies to prompts up to 272,000 tokens. Once a prompt exceeds that threshold, OpenAI bills the whole request at 2x the input and cached-input rate and 1.5x the output rate — $4 input / $0.40 cached / $15 output — per developers.openai.com, confirmed 2026-09-23.
Does GPT-6 Sol replace GPT-5.6 Sol?
Not according to OpenAI's own materials. The announcement calls GPT-5.6 Sol a "predecessor" when comparing scores, never a replaced model, and GPT-5.6 Sol carries no deprecation notice on OpenAI's API deprecations page. Both models remain listed and priced.
What is GPT-6 Sol's knowledge cutoff?
April 20, 2026, with the API serving a 1,050,000-token context window. Inputs past the long-context threshold move to the higher tier — the boundary is in the pricing question above.
Further reading
- Models with 10M token context windows 2026 — GPT-6 Sol is one of the 49 models it compares.
Benchmark guides
- Humanity's Last Exam — what GPT-6 Sol’s reasoning score on it does and doesn’t prove.
- DeepSWE — what GPT-6 Sol’s coding score on it does and doesn’t prove.