GPT-6.1 Sol: Pricing, Availability & Non-English Writing Test

👁 135 reads

GPT-6.1 Sol is OpenAI’s new model unveiled at DevDay on September 29, 2026, one week after the GPT-6 Sol launch. According to OpenAI, it nearly matches GPT-6 Astra in coding with agents, computer use, and office work, while costing one-fifth of its token rates. The same day, we tested it in our Lab on 13 writing tasks in Polish (a heavily inflected language), using the same methodology as previous models. Below you’ll find pricing in USD, availability in ChatGPT plans, and test results.

I cover the full conference in my DevDay 2026 recap.

What is GPT-6.1 Sol

It’s an update to the middle model in the GPT-6 family. Above it sits GPT-6 Astra, below it the budget Luna (I described both younger models in the GPT-6 Sol and Luna guide). OpenAI promotes it with the tagline “intelligence close to Astra at one-fifth the price.” The model arrives one day after the company halted GPT-6.1 Astra for safety reasons, making this the only “six-point-one” available to users.

In the API, the model has the identifier gpt-6.1-sol. On OpenRouter, it appears as openai/gpt-6.1-sol with a context of 1,050,000 tokens and responses up to 128,000 tokens (data from OpenRouter’s provider list, September 29, 2026).

How it differs from GPT-6 Sol and GPT-6 Astra

All numbers in this section come from OpenAI’s announcement. I did not verify them myself.

• Coding (DeepSWE 1.1): according to OpenAI, it matches Astra at about one-fifth the cost and beats GPT-6 Sol’s best result by 6.4 percentage points.
• Business process automation (AutomationBench): 2.2 percentage points more than Claude Opus 5.5 at medium reasoning level, at about one-third the cost, and 4.8 percentage points more than GPT-6 Sol.
• Computer use (OSWorld 2.0): 7 percentage points better than GPT-6 Sol and 2.1 percentage points behind Astra at about one-seventh the task cost.
• Science (Terminal-Bench Science): more than twice the score of GPT-6 Sol; average task cost 5.47 USD versus 23.21 USD for Opus 5.5 and 23.80 USD for Astra. However, Astra has the highest score (68.1%), and OpenAI itself recommends using it for the most difficult research tasks.
• Facts: at low reasoning level, the percentage of responses with errors dropped from 11.4% to 7.7% on a deliberately difficult question set.
• Safety: the model fails to report search tool failure in 2.1% of cases (GPT-6 Sol: 4.9%, Astra: 1.5%, Luna: 28.7%).

OpenAI compares the model to Opus 5.5 where it performs better. In our English test, the order is reversed.

API pricing in USD

Rates from OpenAI’s API pricing page, retrieved September 29, 2026. Prices apply to requests up to 272,000 input tokens.

Above 272,000 input tokens, GPT-6.1 Sol costs 4 USD for input and 15 USD for output. Batch mode and Flex processing cut rates in half (1 and 5 USD), while the faster Fast mode doubles them (4 and 20 USD). The Ultrafast version for GPT-6.1 Sol is set to arrive “in the coming days,” with no price yet.

Conclusion: the “one-fifth the price of Astra” claim is true, but GPT-6 Sol cost exactly the same. Compared to its predecessor, the new model is cheaper only with repeated context. Note: OpenRouter showed 0.20 USD cache for this model on September 29, so with high traffic I’d check the bill.

Our test: English writing

We ran GPT-6.1 Sol through OpenRouter on the same 13 prompts as in our Lab ranking: 9 writing tasks (article, product description, email, LinkedIn post, story, headlines, summary, explanation for a child, translation), two proofreading tasks, and two English language quizzes. We tested it in English. Settings as for all models: reasoning at “low” level, same token limits. Responses were blindly evaluated by two judges from other companies: Claude Opus 5.5 and Gemini 3.1 Pro (the GPT-6 Sol judge was skipped since it’s the same company). The score is 70% judge evaluation, 20% compliance with prompt rules, and 10% text hygiene.

Along with it, we evaluated four anchors from the first round (Opus 5.5, GPT-6 Luna, Qwen3.8 Max, GLM-5.3). Their order didn’t change; the scale shifted by 0.38 points and we corrected the score accordingly. The anchors differed from the first round by 0.2-1.3 points - that’s measurement variance.

Judge scores on a 1-10 scale. GPT-6.1 Sol: round from September 29, 2026, score calibrated with anchors; rest from first round.

What worked well and what didn’t

The model solved both English language quizzes flawlessly (46/46 and 48/48), and the easy proofreading too (22/22). In the hard one, it corrected 28 out of 30 errors. Headlines performed best (9.83), followed by summary (9.33), customer email (9.17), and translation (9.0 - noticeably better than GPT-6 Sol, which scored 7.67 here).

The weakest was the product description: average 5.0. The prompt required a paragraph of 120-150 words using only data from the product sheet. GPT-6.1 Sol reached the required length by repeating the same information in different words, e.g., after a sentence about a removable filter, it added “The filter can therefore be removed for washing and rinsed with water“ (“The filter can therefore be removed for washing and rinsed with water”). Both judges called it padding. In the article about electricity bills, it exceeded the limit (521 words at “about 500”) and frequently started sentences with “For example“ (“For example”).

The model barely thinks at the “low” setting: across 13 tasks it used 795 reasoning tokens, and on average output about 58 tokens per second, including network latency. The entire run cost 0.073 USD, and judge evaluation 1.63 USD.

Our conclusion: in English writing, GPT-6.1 Sol is at the same level as GPT-6 Sol, and the 0.6-point difference falls within measurement variance. For English texts, Opus 5.5 is still better, though in this test it cost over five times more. This is our opinion.

Availability in ChatGPT plans

The model isn’t in regular chat yet. OpenAI says it will appear there later, with no date given. Codex documentation estimates that on the Plus plan and standard Business tier, you’ll get 15 to 160 local messages to GPT-6.1 Sol per five hours (GPT-6 Astra: 5-45). I compiled plan prices in the Pro 500 plan article and the ChatGPT pricing guide.

Who is GPT-6.1 Sol for

For agents in API with repeated context: switching from GPT-6 Sol costs nothing, and cache becomes half the price. For people with Plus plans working in Codex: according to OpenAI, it’s nearly Astra with less limit consumption. For non-English writing, you won’t notice the change. This is our opinion.

01OpenAI DevDay 2026: Dots, GPT-6.1 Sol, Pro 500 and MoreOver 20 announcements at OpenAI DevDay 2026: always-on Dots agents, GPT-6.1 Sol at one-fifth the cost of Astra, Ultrafast mode, ChatGPT Space and Pro 500.Read →02OpenAI Dots: What They Are and Where They WorkOpenAI Dots are always-on AI agents in ChatGPT. Each Dot has its own cloud computer, connects to 4000+ apps, and works even when you're not chatting.Read →03OpenAI Ultrafast: 300 tokens/s, pricing and supported modelsOpenAI Ultrafast generates up to 300 tokens per second in Codex, 8x faster than standard mode. We break down the pricing, model support, and when it's worth it.Read →04ChatGPT Space and Pages: What They Are and How to Get StartedChatGPT Space replaces Library with a team workspace for Pages, files, and presentations. Available in Pro, Business, and Enterprise plans.Read →

Frequently asked questions

How much does GPT-6.1 Sol cost?

In API: 2 USD per million input tokens, 0.10 USD for cached tokens, and 10 USD per million output tokens.

Is GPT-6.1 Sol in free ChatGPT?

No. It’s in Plus, Pro, Business, Enterprise, and Edu plans, currently only in ChatGPT Work and Codex, not in regular chat.

Is GPT-6.1 Sol better than GPT-6 Sol?

According to OpenAI, yes in coding, computer use, and office tasks. In our English writing test, it performed at the same level: 85.8 versus 86.4 points.

// read moreOpenAI DevDay 2026: Dots, GPT-6.1 Sol, Pro 500 and MoreNewsGPT-6 Astra Test: Non-English Writing, Code, 150K ContextBenchmarksAI Detectors and Non-Native English Writers: The EvidenceMagazine

Kinetyra