# GPT-6.1 Sol Web Scraping: 134 of 134 Right

> GPT-6.1 Sol, released 29 September 2026, judged all 134 scraped pages correctly at $2.74 per 1,000. GPT-6 Luna matched it at $0.147. Jev: 131 in 0.3 s.

[Home](https://quanticdata.io/)/[Blog](https://quanticdata.io/blog/)/GPT-6.1 Sol Web Scraping: 134 of 134 Right

# GPT-6.1 Sol Web Scraping: 134 of 134 Right

AI scrapingSep 30, 2026·10 min read·By [Aldo Morese](https://quanticdata.io/about/), founder of QuanticData

Five models asked whether 134 scraped pages are the requested content, measured on 30 September 2026: bar length is the cost per 1,000 pages, from GPT-6 Sol at $2.75 down to Jev 1.13 at $0.07, and the box on the right is how many of the 134 pages each model judged correctly

On this page [GPT-6.1 Sol got all 134 right, and so did two other OpenAI models](/blog/gpt-6-1-sol-block-pages/#gpt-6-1-sol-got-all-134-right-and-so-did-two-other-openai-mo) [The upgrade is invisible on a yes-or-no question](/blog/gpt-6-1-sol-block-pages/#the-upgrade-is-invisible-on-a-yes-or-no-question) [GPT-6 Luna does the same job for one eighteenth of the price](/blog/gpt-6-1-sol-block-pages/#gpt-6-luna-does-the-same-job-for-one-eighteenth-of-the-price) [Jev missed three pages, and flagged all three as unsure](/blog/gpt-6-1-sol-block-pages/#jev-missed-three-pages-and-flagged-all-three-as-unsure) [Gemini 3.5 Flash Lite is fast and lets empty pages through](/blog/gpt-6-1-sol-block-pages/#gemini-3-5-flash-lite-is-fast-and-lets-empty-pages-through) [The setting that works on block-page detection](/blog/gpt-6-1-sol-block-pages/#the-setting-that-works-on-block-page-detection)

OpenAI released GPT-6.1 Sol on 29 September 2026, and the next morning we had a GPT-6.1 Sol web scraping number: on 134 real scraper responses labelled by hand, it judged all 134 correctly, for $0.3666 in total or $2.74 per 1,000 pages. GPT-6 Sol, the model it replaces, also judged 134 of 134 at $2.75 per 1,000. GPT-6 Luna judged 134 of 134 at $0.147. If a model behind your scraper only has to say whether a page is the page you asked for, the upgrade changes nothing: use GPT-6 Luna, or Jev with Luna as the second opinion, and keep Sol for work that needs its reasoning.

## GPT-6.1 Sol got all 134 right, and so did two other OpenAI models

The corpus is the one we built for ../jev-block-page-detection/: 70 heavily protected sites fetched twice each, once by a plain Python client and once through our [web scraping API](https://quanticdata.io/web-scraping-api/) over residential exits, reduced to 134 responses and labelled by hand. 55 are usable, meaning the page is the listing, article, product or search the URL asks for. 79 are not: an empty shell, a sign-in page, a consent page, an error page, or a real page that is the wrong page.

Each LLM saw the URL, the title and up to 12,000 characters of visible text, and one question: is this the content the URL asks for? It answered as JSON with a true or false and a probability, at temperature 0, with 200 output tokens allowed, through OpenRouter from Europe on 30 September 2026 with two requests in flight. Jev 1.13 got the same URL, title and text through the OpenRouter decisions endpoint, with the one "usable" question we published in the Jev test, on the same day. Cost is what OpenRouter billed per call.

| Model | Correct of 134 | Bad pages accepted | Good pages rejected | Cost per 1,000 | Median latency | 90th percentile |
| --- | --- | --- | --- | --- | --- | --- |
| GPT-6.1 Sol | 134 | 0 | 0 | $2.736 | 1.98 s | 2.89 s |
| GPT-6 Sol | 134 | 0 | 0 | $2.749 | 1.64 s | 2.96 s |
| GPT-6 Luna | 134 | 0 | 0 | $0.147 | 2.17 s | 3.18 s |
| Jev 1.13, one question | 131 | 3 | 0 | $0.073 | 0.32 s | 0.38 s |
| Gemini 3.5 Flash Lite | 120 | 14 | 0 | $0.380 | 0.75 s | 0.89 s |
| Claude Sonnet 5.5 (29 Sep, same rig) | 133 | 1 | 0 | $4.170 | 1.46 s | 9.74 s |

Every model answered every page; no call failed after retries. The Sonnet 5.5 row is from the same rig and prompt one day earlier, published in ../claude-sonnet-5-5-block-pages/; everything else ran on 30 September. GPT-6 Luna also scored 134 of 134 on 29 September, so its result held across two days.

## The upgrade is invisible on a yes-or-no question

OpenAI's launch page describes GPT-6.1 Sol as nearly matching GPT-6 Astra on agentic coding, computer use and professional work, with gains over GPT-6 Sol of 6.4 points on a coding benchmark and 7 points on computer use at maximum reasoning. Those are multi-step tasks. Deciding whether one page is the page you asked for is a single step, and GPT-6 Sol already made no mistakes on it. There was nothing left to improve.

The two Sols were billed the same 147,824 input tokens for the whole corpus, which means the same tokenizer, and a median of 19 output tokens per page. GPT-6.1 Sol wrote slightly less in total, 2,721 output tokens against 3,607, and the bill moved by $0.0017 for 134 pages. The list price did not change: $2 per million input tokens and $10 per million output for both.

The only price change in the launch is cached input, now $0.10 per million tokens, half of GPT-6 Sol's. That helps an agent that sends the same long instruction block on every call. It does not help a page classifier, where the page text is different every time and the fixed instruction is about 40 tokens. One detail on the bill: OpenRouter charged $0.3666 where 147,824 input and 2,721 output tokens at list price come to $0.3229, 13.5% more. Its model page says prompts of 1,024 tokens or more are cached automatically and billed a cache write at $2.50 per million; 52 of our 134 prompts were over that line. If you run Sol on page text that is never repeated, send the request with caching set to explicit and no breakpoints, as that page describes, and the write is not billed.

## GPT-6 Luna does the same job for one eighteenth of the price

On this task GPT-6 Luna and GPT-6.1 Sol are indistinguishable in accuracy and close in speed: 134 of 134 each, medians of 2.17 and 1.98 seconds, 90th percentiles of 3.18 and 2.89 seconds. The difference is the bill, $0.0197 against $0.3666 for the corpus. Per 1,000 pages that is $0.147 against $2.736, a factor of 18.6.

Our web scraping API charges from $0.0002 per page over plain HTTP and $0.001 with JavaScript rendering, and bills only pages that come back. Per 1,000 pages, with the check added:

| Classifier | Model per 1,000 | Plain fetch + model | Rendered fetch + model | Model as a multiple of the plain fetch |
| --- | --- | --- | --- | --- |
| Jev 1.13 | $0.073 | $0.273 | $1.073 | 0.4 |
| Jev, then GPT-6 Luna on hesitant scores | $0.087 | $0.287 | $1.087 | 0.4 |
| GPT-6 Luna | $0.147 | $0.347 | $1.147 | 0.7 |
| Gemini 3.5 Flash Lite | $0.380 | $0.580 | $1.380 | 1.9 |
| GPT-6.1 Sol | $2.736 | $2.936 | $3.736 | 13.7 |

A GPT-6.1 Sol verdict costs 13.7 plain fetches, or 2.7 rendered ones. A GPT-6 Luna verdict adds 70% to a plain fetch and 15% to a rendered one. Both caught every one of the 79 unusable pages, so the extra $2.59 per 1,000 buys nothing on this job.

## Jev missed three pages, and flagged all three as unsure

Jev 1.13 judged 131 of 134 on 30 September, the same 131 it scored in the first Jev test two days earlier, for $0.0098 in total and a median of 316 milliseconds. Its three errors were the three hardest pages in the corpus, all accommodation searches for Austin, Texas. One travel site answered with a complete 1,621,876-byte hotel listing for a different city; two others answered with their homepage, 576,101 and 529,633 bytes of real content that is not the search the URL asked for. Claude Sonnet 5.5 accepted the first one on 29 September; Gemini 3.5 Flash Lite accepted all three today.

The difference is how Jev said it. Its scores on those three were 0.54, 0.76 and 0.68: barely over the line. Every page Jev scored below 0.3 or above 0.8 was judged correctly. That turns Jev's weakness into a routing rule we already published in the Sonnet 5.5 post: send anything between 0.3 and 0.8 to a second model. On this corpus that band held 10 pages, 7.5% of the total, and all three errors were inside it. With GPT-6 Luna judging those 10, the pipeline scored 134 of 134 for $0.087 per 1,000 pages, and 124 of the 134 answers came back in about a third of a second.

The band was not tuned on today's run; it is the 0.3 to 0.8 rule from the day before. The narrower band 0.5 to 0.8 would have held 6 pages and still caught all three, but a band set that tight on 134 pages is fitted to them. Keep it wide.

The New Stack reported on 29 September that OpenAI announced a Decisions API built on Luna, in limited preview, returning fixed answers with confidence scores in a claimed 150 milliseconds, with no published price. It is not available to test. When it is, it goes through this same corpus the day it opens.

## Gemini 3.5 Flash Lite is fast and lets empty pages through

Gemini 3.5 Flash Lite was the fastest LLM in the run, a median of 0.75 seconds and a 90th percentile of 0.89, at $0.38 per 1,000 pages. It also accepted 14 bad pages and rejected none, which makes it the only model today that would have written empty rows into a dataset. Eleven of the 14 were shells whose whole visible text is under 300 characters, between 0 and 286 characters, served on search, category, listing and route pages of retail, review and travel sites. The other three were the Austin pages that fooled Jev.

Those eleven are the pages a free length check removes before any model is paid: a visible body under 300 characters is not the listing the URL asked for. With that check in front, Gemini's errors shrink to the same three as Jev's, at five times Jev's price and twice its latency. A false accept is the expensive error: a wrong page stored as data returns empty or plausible fields and nothing downstream flags it, while a false reject costs one refetch at $0.0002.

## The setting that works on block-page detection

- **Model**: Jev 1.13 first, at $0.073 per 1,000 pages and 0.32 s median, with GPT-6 Luna judging every page Jev scores between 0.3 and 0.8. On this corpus: 134 of 134, $0.087 per 1,000, 10 of 134 pages handed off. If you want one model, GPT-6 Luna alone: 134 of 134, $0.147 per 1,000, 2.17 s median.

- **GPT-6.1 Sol**: 134 of 134 at $2.74 per 1,000, the same score as GPT-6 Sol and GPT-6 Luna. Use it where the same call also extracts fields, reads a long document or plans the next step; as a yes-or-no gate it is 18.6 times the price of Luna for the same answers.

- **Free checks first**: a visible body under 300 characters, a network error, a missing title. They remove 11 of Gemini 3.5 Flash Lite's 14 false accepts before any model is paid.

- **Fetch**: our [web scraping API](https://quanticdata.io/web-scraping-api/) over plain HTTP at $0.0002 per page, rendered at $0.001 only for the pages the check marks as empty shells, over [residential proxies](https://quanticdata.io/residential-proxies/) from $0.80/GB when you run your own client. The API retries unusable responses and bills only pages that come back, so the model sees the hard cases: the right site answering with the wrong page.

- **When the classifier is not enough**: a rejected page on a listing site is usually the wrong page for the place in the URL; refetch it with the country pinned to that place, and work through ../proxy-not-working-checklist/ for the fetch-side settings that change the answer. For whole listings, the [collectors](https://quanticdata.io/collectors/) return rows instead of pages to judge.

- Every account gets $2 of free API usage per month: at $0.0002 a page that is 10,000 plain fetches, and the Jev-then-Luna verdicts on all 10,000 cost $0.87 more.

### Sources & further reading

- [Introducing GPT-6.1 Sol, OpenAI (29 September 2026)](https://openai.com/index/introducing-gpt-6-1-sol/)

- [OpenRouter model page: OpenAI GPT-6.1 Sol (pricing, caching, release date)](https://openrouter.ai/openai/gpt-6.1-sol)

- [OpenAI's new GPT-6.1 Sol undercuts its own Astra flagship, The New Stack](https://thenewstack.io/openai-gpt-6-1-sol/)

- [OpenAI answers TypeSafe's Jev with a Decision API built on Luna, The New Stack](https://thenewstack.io/openai-decision-api-luna/)

## FAQ

Quick answers on gpt-6.1 sol web scraping.

[Something else? Ask us](mailto:hello@quanticdata.io)

### How accurate is GPT-6.1 Sol at telling a block page from real content?

On 134 hand-labelled scraper responses from 70 protected sites, tested on 30 September 2026, GPT-6.1 Sol judged all 134 correctly: no bad page accepted, no good page rejected. GPT-6 Sol and GPT-6 Luna also judged 134 of 134 on the same pages the same day; Jev 1.13 judged 131 and Gemini 3.5 Flash Lite 120.

### Is GPT-6.1 Sol better than GPT-6 Sol for web scraping checks?

Not on this task. Both judged 134 of 134, were billed the same 147,824 input tokens, and cost $2.736 and $2.749 per 1,000 pages. Median latency was 1.98 s for GPT-6.1 Sol and 1.64 s for GPT-6 Sol. The gains OpenAI reports are on multi-step coding and computer-use work, not on a single yes-or-no question.

### What does GPT-6.1 Sol cost per 1,000 pages classified?

$2.736 per 1,000 pages as billed by OpenRouter on our corpus, or $0.3666 for 134 pages. The list price is $2 per million input tokens, $0.10 cached and $10 output. The bill was 13.5% above list-price arithmetic because 52 of the 134 prompts were over 1,024 tokens, where automatic caching bills a cache write at $2.50 per million.

### Which model should I put behind a scraper to detect block pages?

Jev 1.13 with GPT-6 Luna as the second opinion: Jev on every page at $0.073 per 1,000 and 0.32 s, Luna on the pages Jev scores between 0.3 and 0.8. On our corpus that was 10 of 134 pages and the pipeline scored 134 of 134 for $0.087 per 1,000. GPT-6 Luna alone scores the same at $0.147 per 1,000.

### Where did Jev go wrong, and can you tell in advance?

Jev accepted 3 of 134 pages it should have rejected: a hotel listing for the wrong city and two homepages served for a city search. Its scores on them were 0.54, 0.76 and 0.68. Every page it scored below 0.3 or above 0.8 was right, so a score in that middle band is the signal to ask a second model.

### Is Gemini 3.5 Flash Lite good enough for page classification?

Only behind a free length check. It was the fastest LLM in the test at 0.75 s median and $0.38 per 1,000, but it accepted 14 of the 79 bad pages. Eleven of those had under 300 characters of visible text, which a length rule removes for free; the other 3 are the same wrong-place pages that Jev scored as unsure.

## Feed the classifier pages worth judging

Every page in this test came from our web scraping API: plain HTTP or rendered, residential exits, unusable responses retried and never billed, so the model only sees the cases a fetcher cannot judge. Every account gets $2 of free API usage per month.

[Start free — $2/month included](https://quanticdata.io/signup/)[Explore Web Scraping API](https://quanticdata.io/web-scraping-api/)

## Related reading

[AI scraping Jev vs GPT, Claude, Gemini on Real Web Data Same 317 hand-labelled cases, same question, six models. Accuracy was a tie: 306 to 309 correct. Jev was 2 to 42 times cheaper and answered in a third of a second. The difference that matters is where the mistakes were: all 11 of Jev's errors came with an uncertain score, while the LLMs made theirs sounding sure. Read more](https://quanticdata.io/blog/jev-vs-llm-benchmark/) [AI scraping Claude Sonnet 5.5 for Web Scraping: 133 of 134 Hours after Anthropic released Claude Sonnet 5.5 we sent it the same 134 hand-labelled scraper responses we used for the Jev tests: is this the requested page, or a stub, a sign-in page, an error, the wrong page? It got 133 right for $0.56, about $4.17 per 1,000 pages. GPT-6 Luna and DeepSeek V4.1 Flash got all 134 for $0.147 and $0.236 per 1,000. Claude Haiku 4.5 got 125. The one page Sonnet 5.5 missed was a real one. Read more](https://quanticdata.io/blog/claude-sonnet-5-5-block-pages/) [AI scraping OpenAI Scraping Lawsuit: 0 of 30 Block Bingbot Unsealed filings in the New York Times-led case say Microsoft and OpenAI built training sets from news, including data gathered for Bing. We read the robots.txt of 30 news sites on 29 September 2026: 20 block GPTBot, 27 block ClaudeBot, and not one blocks Bingbot. Read more](https://quanticdata.io/blog/openai-scraping-lawsuit/)

---

Source: https://quanticdata.io/blog/gpt-6-1-sol-block-pages/ · Site index for AI: https://quanticdata.io/llms.txt