GPT-6 Sol is the middle tier of OpenAI's GPT-6 family: below the flagship GPT-6 Astra and above the fast, cheap GPT-6 Luna. It is aimed at complex coding and agentic workflows — the work where a model has to plan, call tools, read results and keep going, rather than answer once and stop.
All three share the same 1,050,000-token context window. What separates them is depth and price: Sol costs a fifth of Astra and roughly twenty times Luna, which makes the choice between them a question about how hard the task actually is.
A chat-sized exchange of 1,000 input and 300 output tokens runs about $0.0065 — roughly 150 calls per dollar. A code-review request with 50,000 tokens of repository context and a 2,000-token reply costs about $0.156, so a dollar covers around six of them.
Prompt caching is where the real money is on agentic work. Cached input is $0.26 against $2.6 — a tenth — while writing to the cache costs $3.25 once. A twenty-step agent replaying a 30,000-token preamble pays about $1.56 without caching and about $0.25 with it, after a one-time write of roughly $0.10. The break-even lands at the second step, so for anything that loops it is close to free money.
Sol runs from none to max, defaulting to medium. That control matters more than it sounds: reasoning tokens bill at the output rate, so effort is the main lever on both latency and cost. Routine edits, classification and formatting rarely need more than low; architecture decisions, tricky debugging and multi-file refactors are where high or max pays for itself. Leaving everything at max is the most common way to overspend on a model like this.
1,050,000 tokens is on the order of 3,000 pages, or a substantial codebase with its history and documentation. For agents, the practical win is that a long transcript does not need to be summarised away every few steps — the point where long runs usually start contradicting themselves. For code work, it means the model can hold the actual call sites rather than a retrieved guess at them.
Output is capped at 128,000 tokens, which is enough for a large diff or a long structured document in a single reply.
Complex coding. Multi-file changes, migrations and debugging where the model needs both the code and the surrounding context to be right rather than plausible.
Agentic pipelines. Planning, tool calls, parallel tool calls and structured outputs, with web search available when the answer is not in the prompt.
Vision-adjacent work. Screenshots, diagrams and design mockups alongside the prompt — UI work that would otherwise have to be described in words first.
Long-context analysis. Research, due diligence and audits across document sets that would need a retrieval layer elsewhere.
For high-volume, well-specified work — classification, extraction, short rewrites — GPT-6 Luna at $0.13 / $0.65 does the same job at a twentieth of the price. For the hardest problems, where one correct answer is worth more than the difference in bill, GPT-6 Astra at $13 / $65 is the better trade. Sol is the middle: strong enough for real engineering, cheap enough to run inside a loop.
All three carry the same 1,050,000-token window, so the comparison is about capability per dollar rather than capacity. Astra is the flagship at $13 / $65 per 1M tokens; Sol sits at $2.6 / $13; Luna is the fast tier at $0.13 / $0.65. A common pattern is to route by task: Luna for the mechanical steps, Sol for the ones that need judgement, Astra reserved for the few that genuinely stall the other two.
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
Complex coding. Multi-file refactors, migrations and debugging where the model needs the real call sites, not a retrieved approximation. The 1,050,000-token window holds the code and its history; the 128,000-token output ceiling leaves room for a large diff in one reply.
Agentic workflows. Planning loops with function calling, parallel tool calls and structured outputs, plus web search when the answer is not in the prompt. Prompt caching at a tenth of input price is what keeps a twenty-step loop affordable.
Multimodal engineering work. Screenshots, diagrams and mockups in the same request as the prompt — UI fixes and design-to-code work without describing the picture first.
Long-context research. Audits, due diligence and literature review across document sets large enough that a retrieval pipeline would otherwise be mandatory.
import requests r = requests.post( "https://api.aimlapi.com/v1/chat/completions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "openai/gpt-6-sol", "messages": [ { "role": "user", "content": "Hello!" } ] }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "openai/gpt-6-sol", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"openai/gpt-6-sol","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Cached input |
Cache writes are billed at $3.25 / 1M tokens.
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| Intelligence | 47.5 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | September 24, 2026 |
Complex coding. Multi-file refactors, migrations and debugging where the model needs the real call sites, not a retrieved approximation. The 1,050,000-token window holds the code and its history; the 128,000-token output ceiling leaves room for a large diff in one reply.
Agentic workflows. Planning loops with function calling, parallel tool calls and structured outputs, plus web search when the answer is not in the prompt. Prompt caching at a tenth of input price is what keeps a twenty-step loop affordable.
Multimodal engineering work. Screenshots, diagrams and mockups in the same request as the prompt — UI fixes and design-to-code work without describing the picture first.
Long-context research. Audits, due diligence and literature review across document sets large enough that a retrieval pipeline would otherwise be mandatory.
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
GPT-6 Sol This page | Everyday coding and agents | |||
| Agentic workflows and computer use | ||||
| Frontier reasoning + agents | ||||
| Everyday coding + agents |
Yes. GPT-6 Sol is available on AI/ML API through the OpenAI-compatible chat completions endpoint under the model ID openai/gpt-6-sol.
On AI/ML API, GPT-6 Sol costs $2.6 per million input tokens, $13 per million output tokens and $0.26 per million cached input tokens. Cache writes are billed at $3.25 per million.
GPT-6 Astra is the flagship of the GPT-6 family, at $13 / $65 per million tokens. GPT-6 Sol sits below it at $2.6 / $13 with the same 1,050,000-token context window, for coding, reasoning and agentic work that does not need the heaviest tier.
Send a request to https://api.aimlapi.com/v1/chat/completions with openai/gpt-6-sol as the model id. It is the same OpenAI-compatible endpoint every other GPT model on AI/ML API uses.
