GPT-6 Sol Is Now in Felo Search: Agent Loops at $2/$10
GPT-6 Sol is live in Felo Search: 1.05M context at $2/$10 per million tokens, the mid-tier built for the agent loops AI search depends on.
OpenAI released GPT-6 Sol on September 22, alongside GPT-6 Luna and two and a half weeks after the Astra flagship. Sol is the middle tier: $2 per million input tokens and $10 per million output, one-fifth of Astra's price and half of what GPT-5.6 Sol cost.
It is live in Felo Search now. The case for Sol is not that it is the best model available. Astra is. It is that a search answer is never one model call, and the tier you pick decides how many calls the answer is allowed to take.
What GPT-6 Sol Actually Is
Sol sits between Astra and Luna in OpenAI's GPT-6 generation. The position sounds unremarkable until you look at the price ladder it belongs to.
| Spec | GPT-6 Sol |
|---|---|
| Model ID | gpt-6-sol |
| Maker | OpenAI |
| Released | September 22, 2026 |
| Context window | 1,050,000 tokens |
| Max input | 922,000 tokens |
| Max output | 128,000 tokens |
| Input | Text, image |
| Output | Text |
| Price | $2 per million input, $10 per million output |
| Cached input | $0.20 per million tokens |
| Cache writes | $2.50 per million tokens |
| Knowledge cutoff | April 20, 2026 |
The GPT-6 generation inverts the naming from GPT-5.6. In the previous generation Sol was the flagship, with Terra and Luna beneath it. In GPT-6, Astra leads at $10/$50, Sol is the middle tier, and Luna anchors the family at $0.10/$0.50. Same three names, opposite ranking. The version number is the only thing separating the two Sols.
Sol shares its request surface with Astra. Same request format, same tool-calling schema, same 1,050,000-token context window, same reasoning-effort scale. Moving a workload between the flagship and the mid-tier is a model-ID change, not a rewrite. That is a bigger deal than it sounds: it means you can prototype on the cheap tier and escalate the hard cases without touching your integration.
How This Generation Became Public
Sol's release is unusually well documented, because the model was visible before it was announced.
September 3. OpenAI ships GPT-6 Astra, its first model rated Critical for cyber capability, establishing the generation.
September 6. The first public claim that Sol is coming, describing Astra as the flagship with "Sol, Terra, and Luna still to come." No screenshot, no changelog. Just a name circulating without evidence.
September 16. gpt-6-sol starts appearing in live API traffic. Developers routing requests to the previous generation's model ID report responses coming back tagged gpt-6-sol. A string in a response header is not an announcement, but it is hard evidence that something exists.
September 22. OpenAI ships Sol and Luna on the API and confirms the identifier that had already been answering requests.
The middle two entries are worth keeping in the record because they are how most model releases now surface: not a press release, but a string in someone's logs. It is also why Felo pages label vendor-reported benchmark numbers as claims rather than results.
Why the Middle Tier Is the Interesting One
A single search question in Felo is not one call to one model. It is a loop: decide what to look up, read what comes back, decide whether that was enough, look again, reconcile the sources, then write. Planning, tool selection, and retries all burn tokens, and a long investigation can run that loop a dozen times before producing an answer.
At Astra's $10/$50, a deep research pass is a real line item. At Sol's $2/$10, it is a fifth of that. Nothing about the loop changed; the ceiling on how many times you can afford to go around it moved.
That is the specific thing OpenAI built Sol for. It is positioned for complex coding and agentic workflows, where the per-token price decides whether the job is worth running at all, and it keeps Astra for the plans that need the deepest reasoning.
If you are choosing between the two on the same question, the rule of thumb is simple: Astra when being wrong is expensive, Sol when being slow is expensive.
How to Use GPT-6 Sol in Felo Search
1. Open the model picker and select GPT-6 Sol. Same search box, same sources, same workspace.
2. Ask questions that require several rounds. Sol's advantage shows up on questions that need a loop: compare, verify, reconcile. Single lookups do not need a mid-tier model at all.
3. Use the full context window instead of trimming. 922,000 tokens of input is enough for a repository, a filing set, or a long research thread with every source still attached. Retrieval layers are optional at this size.
4. Escalate deliberately. When a question turns out to need the deepest reasoning, switch to a stronger model with the evidence already in context. The shared request surface means nothing breaks in the handoff.
5. Keep the loop short where you can. Sol is cheap, not free. Cached input at $0.20 per million tokens rewards reusing the same source set across questions rather than re-reading it.
The Other Three Models Live in Felo Search
Sol arrived with three peers, and Felo Search runs all four.
| Model | Released by | Price per million tokens | Context | Built for |
|---|---|---|---|---|
| Grok 4.7 | SpaceXAI | $2 in / $6 out | 500K | Long-horizon reasoning and knowledge work |
| Claude Opus 5.5 | Anthropic | $4 in / $20 out | 1M | Answers that survive verification |
| GPT-6 Luna | OpenAI | $0.10 in / $0.50 out | 1.05M | High-volume search at the lowest cost per call |
| GPT-6 Sol | OpenAI | $2 in / $10 out | 1.05M | Agent loops and repeat calls at mid-tier price |
Sol and Grok 4.7 land at the same input price with different strengths: Sol carries twice the context, Grok 4.7 charges less per output token. For a research pass that reads a lot and writes a little, that difference is small. For one that writes a long synthesis, it is not.
What to Try First
A multi-step comparison. Ask for three vendors, their pricing, and the caveat each one buries. That question needs several retrieval rounds, which is exactly where Sol's price matters.
A repository-scale question. Point Sol at a large body of documentation and ask how a specific behavior is implemented, with citations back to the source files.
A first pass you intend to escalate. Run the broad investigation on Sol, then move the conclusion to a stronger model for a final check. Two tiers, one thread.
A long document you have to argue with. Paste a contract and ask what it does not cover. The context window is what makes "read all of it" a realistic instruction.
Try GPT-6 Sol in Felo Search
GPT-6 Sol is live in Felo Search with a 1,050,000-token context window, image input, and $2/$10 pricing, a fifth of the flagship tier it sits beneath.
Open GPT-6 Sol in Felo Search →
This post is also available in 简体中文, 日本語, 한국어, 繁體中文, हिन्दी, Français, العربية, Русский, اردو, Bahasa Indonesia, Deutsch, Tiếng Việt, Türkçe, Italiano, ไทย, Español, বাংলা and Português.