GPT-5.6 Sol
GPT-5.6 Sol is the flagship of the GPT-5.6 family, released by OpenAI on Jul 9, 2026; the company's top spot passed to GPT-6 Astra on Sep 3. It now costs US$ 4 per million input tokens and US$ 20 per million output, a promotional price valid at least through Nov 21, 2026; at launch it was US$ 5 and US$ 30. Its successor in the Sol tier is GPT-6.1 Sol, at half the price. No retirement date has been published, and it still answers in ChatGPT Chat on paid plans.
GPT-5.6 Sol in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- OpenAI
- API identifier
gpt-5.6-sol(thegpt-5.6alias points to it)- Release
- Jul 9, 2026, with GPT-5.6 Terra and GPT-5.6 Luna, after a limited preview
- Input
- US$ 4 per million tokens (promotion from Aug 21, 2026 to at least Nov 21, 2026; was US$ 5)
- Output
- US$ 20 per million tokens (was US$ 30)
- Cache read
- US$ 0.40 per million (10% of input)
- Cache write
- US$ 5 per million (1.25x input)
- Above 272 thousand input tokens
- US$ 8 input, US$ 0.80 cache and US$ 30 output, for the whole request
- Batch and Flex
- US$ 2 input and US$ 10 output
- Fast mode
- US$ 8 and US$ 40, up to 2.5x faster; Ultrafast only in limited preview
- Context window
- 1,050,000 tokens (up to 922 thousand input)
- Max output
- 128 thousand tokens
- Knowledge cutoff
- Feb 16, 2026
- Reasoning
- effort
nonetomax, defaultmedium;promode on the Responses API - Where it runs
- OpenAI API, ChatGPT (Chat and Work) and Codex
- Risk
- High in cybersecurity and biology on the Preparedness scale, below Critical
- Retirement
- no date published as of Oct 1, 2026
Where GPT-5.6 Sol does well, where it falls behind and who it is for
Where it does well
- Coding: 80 points on the Artificial Analysis Coding Agent Index at max effort, above Claude Fable 5 (77.2) and GPT-5.5 (76.4), in OpenAI's table.
- Computer use and browsing: 62.6% on OSWorld 2.0, against 54.8% for Claude Opus 4.8, and 90.4% on BrowseComp, which rises to 92.2% in ultra mode, with four agents.
- Cybersecurity: 73.5% on ExploitBench and 71.2% on SEC-Bench Pro, against 47.9% and 45.8% for GPT-5.5, in OpenAI's figures.
- Long context: 91.5% on the MRCR v2 test from 256 thousand to 512 thousand tokens, against 81.5% for GPT-5.5.
- Flexible API: accepts
noneeffort, and the model page lists Chat Completions and function calling without the tool restriction GPT-6.1 Sol has.
Where it falls behind
- Already surpassed in-house: in the GPT-6 Astra announcement table, Astra scores higher on every row where both appear, and GPT-6.1 Sol costs half for input and output.
- Price with a deadline: the US$ 4 and US$ 20 hold at least through Nov 21, 2026. OpenAI has not published the price after that.
- Hard software engineering: 64.6% on SWE-Bench Pro, against 80% for Claude Fable 5 and 80.3% for Claude Mythos 5, in OpenAI's own table.
- An agent that overreaches: in the system card, OpenAI says it goes beyond the request more than GPT-5.5, though at low rates.
- Safeguards that get in the way: Sol's cyber safeguards block roughly ten times more suspicious activity than earlier models' did, and OpenAI admits friction for legitimate use.
Who it makes sense for
- Integrations already running on
gpt-5.6-soland validated: there is no shutdown date, and OpenAI promises at least 6 months' notice. - Teams that must leave GPT-4, o1, o3 or the original GPT-5 before shutdown: the official replacement table still points to
gpt-5.6-sol. - Workflows that split work between Sol, Terra and Luna of the 5.6 family, which reuse each other's saved reasoning.
- Anyone on a paid ChatGPT plan: in Chat, it is the model behind Instant and the Medium, High and Extra High levels.
When to pick another model
- New API projects: GPT-6.1 Sol costs half for input and output, a quarter for cache reads, and knows facts up to Apr 30, 2026.
- High volume of simple tasks: GPT-6 Luna costs US$ 0.10 and US$ 0.50, one fortieth of Sol's price per token.
- Scientific research in the terminal: 22.4% on Terminal-Bench Science, against 64.6% for GPT-6 Astra in the Astra announcement table.
- Unsupervised agents in sensitive environments: OpenAI itself recommends supervising its work on long coding tasks.
What OpenAI published at launch
The GPT-5.6 family arrived on Jul 9, 2026, after a limited preview, with three tiers: Sol as the flagship, Terra for everyday work and Luna for volume. It brought max effort, Programmatic Tool Calling and the multi-agent beta on the Responses API. In ChatGPT Work and Codex, ultra mode runs four agents in parallel.
In the announcement table, Sol leads five of the nine tests below and loses four to Claude Fable 5, all chosen by OpenAI. On Agents' Last Exam, the text cites 53.6, the score at xhigh effort; the table shows 52.7% at max.
| GPT-5.6 Sol OpenAI | GPT-5.5 OpenAI | Claude Fable 5 Anthropic | Claude Opus 4.8 Anthropic | |
|---|---|---|---|---|
| Agents' Last Exam professional agent work |
52.7% | 46.9% | 40.5% | 45.2% |
| GDPval-AA v2 office work, Elo |
1,747.8 | 1,493.7 | 1,759.6 | 1,600.1 |
| AA Coding Agent Index v1.1 coding agent |
80 | 76.4 | 77.2 | 72.5 |
| SWE-Bench Pro software engineering |
64.6% | 59.4% | 80% | 69.2% |
| DeepSWE v1.1 long-horizon engineering |
72.7% | 67% | 69.7% | 59% |
| Terminal-Bench 2.1 terminal work |
88.8% | 85.6% | 83.1% | 78.9% |
| OSWorld 2.0 computer use |
62.6% | 47.5% | not published | 54.8% |
| FrontierMath Tier 4 hard math |
83% | 72.5% | 87.8% | 56.1% |
| Toolathlon tool use |
58% | 55.6% | 61.7% | 59.9% |
Table from the GPT-5.6 announcement, Jul 9, 2026. The dot marks the best in each row. "Not published" cells are tests where OpenAI did not show Fable 5. In later announcements, some Sol scores change: OSWorld 2.0 (offline set) appears at 65.7% in the GPT-6 Astra announcement.
Is GPT-5.6 Sol still worth using in 2026?
For a new API integration, no. On Sep 22 OpenAI released GPT-6 Sol at a price 50% below 5.6 Sol's promotional rate and, a week later, GPT-6.1 Sol, which the documentation calls the newer Sol model.
For teams already in production with it, there is no rush. OpenAI's deprecations page shows no date for GPT-5.6 Sol and still lists it as the replacement for older models that shut down on Oct 23, 2026 and Dec 11, 2026.
The thing to watch is price. The US$ 4 and US$ 20 are promotional, and OpenAI has not said what comes after Nov 21, 2026. Even with the promotion, 5.6 Sol costs twice as much as 6.1 Sol per token and four times as much on cache reads.
| GPT-5.6 Sol | GPT-6.1 Sol | GPT-6 Astra | GPT-6 Luna | |
|---|---|---|---|---|
| Input | US$ 4 | US$ 2 | US$ 10 | US$ 0.10 |
| Output | US$ 20 | US$ 10 | US$ 50 | US$ 0.50 |
| Cache read | US$ 0.40 | US$ 0.10 | US$ 1 | US$ 0.01 |
| Cache write | US$ 5 | US$ 2.50 | US$ 12.50 | US$ 0.125 |
| Above 272 thousand tokens input and output |
US$ 8 and US$ 30 | US$ 4 and US$ 15 | US$ 20 and US$ 75 | US$ 0.20 and US$ 0.75 |
| Knowledge cutoff | Feb 16, 2026 | Apr 30, 2026 | Apr 30, 2026 | May 18, 2026 |
| Accepts effort none | yes | no | no | yes |
| Role today | previous generation | newer Sol | highest intelligence | fastest and most cost-effective |
OpenAI documentation, pricing page and model pages, checked on Oct 1, 2026. GPT-5.6 Sol price is the promotional one.
GPT-5.6 Sol against its direct successor, on the same chart
In the GPT-6 Sol announcement, OpenAI put the two side by side. The successor costs less per task on every test, but does not win them all: on DeepSWE and OSWorld 2.0, 5.6 Sol at max effort still scores higher, while spending more than twice as much.
OpenAI says GPT-6 Sol makes about half as many factual mistakes as its predecessor. At medium effort, the rate of answers with errors on hard prompts falls from 15.0% to 6.9%.
| GPT-5.6 Sol OpenAI | GPT-6 Sol OpenAI | |
|---|---|---|
| AutomationBench automation |
28.8% (US$ 0.67) | 33.2% (US$ 0.27) |
| Agents' Last Exam agent tasks |
53.6% (US$ 5.08) | 56.4% (US$ 2.93) |
| DeepSWE v1.1 software engineering |
72.7% (US$ 6.46) | 68.8% (US$ 2.74) |
| FrontierCode 1.1 Main coding |
47.5% (US$ 5.19) | 49.3% (US$ 2.14) |
| OSWorld 2.0 offline computer use |
66.2% (US$ 7.71) | 64.4% (US$ 3.25) |
| Factual errors hard prompts, lower is better |
8.4% (US$ 0.39) | 4.5% (US$ 0.13) |
Charts from the GPT-6 Sol and Luna announcement, Sep 22, 2026. Each model's best score at any effort; in brackets, the average cost per task. Some scores differ from the GPT-5.6 launch, such as 5.6 Sol's AutomationBench (28.8% here, 18.1% there).
What changes when migrating to GPT-6.1 Sol
The switch is one line, but OpenAI's migration guide lists points that break integrations:
- 6.1 Sol does not accept
noneeffort; OpenAI says to uselow. GPT-6 Sol does, if your integration depends on it. - Tool calling only through the Responses API. On GPT-6 Sol, Chat Completions only calls functions at
noneeffort. - With effort other than
none, thetemperature,top_pandtop_logprobsparameters go away. - Reasoning saved on 5.6 is not reused, because OpenAI only allows reuse within the same model family.
- 6.1 Sol is rated Critical in cybersecurity and uses GPT-6 Astra's safeguards; 5.6 Sol is High. Test defensive security workflows first.
Examples of using GPT-5.6 Sol
A pipeline validated on gpt-5.6-sol can keep running, because there is no shutdown date. Before Nov 21, 2026, it is worth running ten real tasks on GPT-6.1 Sol and comparing accuracy and cost per task.
In ChatGPT Work and Codex, ultra mode splits the search across four agents; on BrowseComp, that took Sol from 90.4% to 92.2%. On the API, the same setup is built with the Responses API multi-agent beta, paying for the tokens of every agent.
Sol scored 91.5% on OpenAI's needle test between 256 thousand and 512 thousand tokens. The bill changes in that range: above 272 thousand input tokens, the whole request moves to US$ 8 input and US$ 30 output.
Where GPT-5.6 Sol stands out and where it falls behind, by the numbers
GPT-5.6 Sol has published results on 39 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Oct 12, 2025 and Oct 7, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 39 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 93.5% | #15 / 81 | #15 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 82 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 96.0% | #12 / 70 (tied with 3) | #12 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 59.21 | #6 / 70 | #6 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 32.3% | #1 / 79 | #1 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 89.4% | #5 / 55 | #5 / 112 | 93.6% GPT-6 Astra |
| SciCode | 57.8% | #14 / 76 (tied with 1) | #15 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 89.1% | #6 / 66 | #6 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 92.5% | #4 / 51 | #4 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 2,177 | #4 / 65 | #4 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,617 | #16 / 76 | #16 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 97.5% | #7 / 51 (tied with 2) | #7 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 69.7% | #8 / 64 (tied with 1) | #8 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 68.0% | #2 / 65 (tied with 1) | #2 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 58.0% | #7 / 61 (tied with 1) | #7 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 83.0% | #11 / 69 (tied with 1) | #12 / 74 | 100.0% Claude Fable 5.1 |
| SimpleBench | 71.7% | #10 / 38 | #10 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 82.9% | #7 / 56 | #7 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 9,619 | #7 / 49 | #7 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 51.4% | #25 / 49 | #25 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 47.5% | #13 / 43 | #13 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 30.7% | #4 / 36 | #5 / 37 | 35.0% GPT-6.1 Sol |
| EnigmaEval | 37.1% | #2 / 14 | #2 / 36 | 39.3% Claude Fable 5 |
| Furniture Assembly | 56.7% | #9 / 33 | #9 / 34 | 83.3% Claude Opus 5.5 |
| BALROG | 60.0% | #3 / 10 | #3 / 32 | 68.3% GPT-6 Astra |
| Blueprint Bench 2 | 33.6% | #9 / 31 (tied with 1) | #9 / 31 | 54.4% Gemini 4 Argon |
| Gso | 76.5% | #4 / 17 | #4 / 28 | 88.2% Claude Fable 5.1 |
| Ebr Bench | 44.8% | #8 / 23 | #8 / 27 | 76.2% GPT-6 Astra |
| DeepSWE | 72.7% | #4 / 26 | #4 / 26 | 74.1% GPT-6 Astra |
| Surface Evolver Bench | 93.1% | #3 / 23 | #3 / 24 | 95.0% Claude Fable 5 |
| Gbaeval | 52.6% | #7 / 22 | #7 / 22 | 79.6% Claude Opus 5 |
| FrontierSWE | 32.2% | #9 / 20 | #9 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 41.7% | #9 / 16 | #9 / 16 | 57.8% Claude Opus 5.5 |
| Posttrainbench | 36.2% | #2 / 11 | #2 / 11 | 41.8% Claude Fable 5 |
| Weirdml V3 | 15.0% | #6 / 11 | #6 / 11 | 42.2% GPT-6 Astra |
| MirrorCode | 20.0% | #6 / 9 | #6 / 9 | 77.4% Claude Opus 5.5 |
| Osworld 2 | 27.3% | #2 / 9 | #2 / 9 | 31.4% Claude Opus 5 |
| Btf3 | 13.7% | #7 / 8 | #7 / 8 | 14.9% GLM-5.3 |
| Frontiermath Erdos | 0.0% | #4 / 7 (tied with 3) | #4 / 7 | 2.9% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare GPT-5.6 Sol with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 2026 |
| Claude Opus 5 | Anthropic | 162.60 | 1,573 | 80.1 | Jul 24, 2026 |
| Claude Fable 5 | Anthropic | 162.13 | 1,542 | 83.0 | Jun 9, 2026 |
| GPT-5.6 Sol | OpenAI | 161.73 | 1,527 | 81.1 | Jul 9, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
GPT-5.6 Sol or Claude Fable 5?
Claude Fable 5 is the model from another lab closest to GPT-5.6 Sol in capability. Across the 36 tests both took, this is the score:
| Test | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| MirrorCode | 20.0% | 63.9% |
| Vending Bench 2 | 9,619 | 5,680 |
| Furniture Assembly | 56.7% | 35.8% |
| FrontierSWE | 32.2% | 47.0% |
| Chess puzzles | 68.0% | 47.0% |
| Gbaeval | 52.6% | 74.5% |
| Apex Agents | 51.4% | 63.6% |
| Posttrainbench | 36.2% | 41.8% |
| Blueprint Bench 2 | 33.6% | 38.6% |
| ProofBench | 83.0% | 95.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What GPT-5.6 Sol costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
ARC-AGI 2
The score scale does not start at zero.
ARC-AGI
The score scale does not start at zero.
CursorBench
The score scale does not start at zero.
Other models from OpenAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| GPT-6 Sol | Sep 22, 2026 | 162.55 | n/a | 79.3 | API access |
| GPT-6 Luna | Sep 22, 2026 | 156.31 | n/a | 72.0 | API access |
| GPT-6 Astra | Sep 3, 2026 | 166.41 | n/a | 82.2 | API access |
| GPT-5.6 Sol | Jul 9, 2026 | 161.73 | 1,527 | 81.1 | API access |
| GPT-5.6 Terra | Jul 9, 2026 | 159.43 | 1,515 | 77.9 | API access |
| GPT-5.6 Luna | Jul 9, 2026 | 156.33 | 1,478 | 73.6 | API access |
| GPT-5.5 Instant | May 5, 2026 | 142.45 | 1,456 | n/a | API access |
| GPT-5.5 Pro | Apr 23, 2026 | 161.97 | n/a | n/a | API access |
| GPT-5.5 | Apr 23, 2026 | 158.95 | 1,533 | 80.2 | API access |
Have you used GPT-5.6 Sol? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about GPT-5.6 Sol
How much does GPT-5.6 Sol cost?
US$ 4 per million input tokens and US$ 20 per million output, a promotional price since Aug 21, 2026 and valid at least through Nov 21, 2026. At launch it was US$ 5 and US$ 30. Cache reads cost US$ 0.40 and cache writes US$ 5. Above 272 thousand input tokens, the whole request moves to US$ 8 and US$ 30. Batch and Flex are half price, and Fast mode costs double.
Will GPT-5.6 Sol be retired?
No date has been published. As of Oct 1, 2026, OpenAI's deprecations page does not list the model and still names it as the replacement for older models that shut down on Oct 23, 2026 and Dec 11, 2026. OpenAI promises at least 6 months' notice for generally available models.
Which model replaced GPT-5.6 Sol?
In the Sol tier, GPT-6 Sol, released on Sep 22, 2026 at half the promotional price, and then GPT-6.1 Sol, from Sep 29, 2026, which OpenAI calls the newer Sol model. The role of OpenAI's most capable model passed to GPT-6 Astra on Sep 3, 2026.
Is the GPT-5.6 Sol in ChatGPT the same as on the API?
Not exactly. OpenAI says the Chat version is slightly different from the one used on the API, in Codex and in ChatGPT Work. On Aug 6, 2026, Plus and Pro plans got an updated version in Chat, while Codex and Work kept the July one.
Where does GPT-5.6 Sol rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, GPT-5.6 Sol ranks #11 out of 267 models, with an ECI of 161.73. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is GPT-5.6 Sol better than Claude Fable 5?
Across the 36 tests both models took and Epoch AI compiled, GPT-5.6 Sol comes out ahead on 13, behind on 21 and tied on 2. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is GPT-5.6 Sol best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of GPT-5.6 Sol are: CritPt (#1 of 79); Chess puzzles (#2 of 65); ALE-Bench (#4 of 65). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- OpenAI, GPT-5.6 announcement
- OpenAI, GPT-5.6 Sol model page (documentation)
- OpenAI, API pricing
- OpenAI, API changelog (launch, Fast mode and promotional price)
- OpenAI, GPT-5.6 system card
- OpenAI, GPT-5.6 August update system card
- OpenAI, API deprecations
- OpenAI, GPT-6 Sol and GPT-6 Luna announcement (charts including GPT-5.6 Sol)
- OpenAI, GPT-6 Astra announcement (table including GPT-5.6 Sol)
- OpenAI, latest model guide (migration)
- OpenAI Help Center, GPT-5.6 and GPT-6 Pro in ChatGPT
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking