Starting from July few clients independently of each other complained that a part of ReCaptcha tokens were not accepted by websites where they bypass ReCaptcha. We performed a set of experiments and found that during the verification of tokens ReCaptcha API returns the error:
Not only our service is affected by this problem, but also real people who pass ReCaptcha on websites manually - the website doesn’t accept the solution when the challenge is completed.
We've been solving ReCaptcha V3 for 3 weeks and are now ready to share some stats.
First, a few words about ReCaptcha V3
As you probably know, ReCaptcha V3 doesn't require any kind of solving as before, instead it measures a user's score and then passes it to the site's backend. This score is based on the information google.com is gathering about the user. The score ranges from 0.1 to 0.9 and may vary for different sites.
We've made a few experiments and conclude that if someone has 0.1 on one site, it is 90% likely that he will get the same score of 0.1 for other sites.
How does the solving of 2captcha.com works:
Every worker is given a test captcha which gives us a score. We record this score for every worker. When a customer requests solving of ReCaptcha V3 with some minimum score requirement we send those requests to the workers with appropriate scores only. As a result, 80 to 90 persents of solvings will hit the required score.
Here is an example of a request for ReCaptcha V3 solving:
It is almost the same as ReCaptcha V2 but has a few new parameters:
indicates that it is ReCaptchaV3
indicates the name of the action on the page
indicates the minimum score the customer needs
For the testing period we set the same price as ReCaptcha V2
Refunds for invalid tokens:
It's complicated. When a customer complains about graphic captcha we can solve it again and see whether the solving was correct or not. When a customer complains about ReCaptcha V2 we can look up the complaint stats and see if some workers are making deliberate mistakes, and then refund for wrongly solved captchas.
ReCaptcha V3, on the other hand, is different: one worker could have different scores for different sites. One site will get a 0.1 score and another will get a 0.5 score for the same worker.
We still support a complaint message reportbad which you should send in case a site refuses the token:
But we also added a new reportgood message for the correct token:
When we receive this reportgood message we add the appropriate worker to the WhiteList for your account so all your requests of ReCaptcha V3 will only be distributed among the workers from your WhiteList.
In future we plan to refund only the customers who send us reportgood messages.
Please note that we do not refund for reportbad messages at the moment, but we probably will do so soon.
What min_score should you use?
For now we support only 0.1, 0.3 and 0.9 scores. The sites we tested accept both 0.9 and 0.3 scores and none of them accept 0.1. That's why we suggest that customers start from the score of 0.3 and increase it to 0.9 only if more than 50% of tokens are denied. Please note that there is twice as many workers with the score of 0.3 than those with the score of 0.9.
Yesterday we wrote about the need to have more work for ReCaptcha solvers. So we decided to solve this implementing new discount rates for solving on limited domains. To do so we will need more workers online to be ready to solve higher load of captchas. So we are planning to improve overall workers' online without price rising for regular customers. Discount rates will have lower priority: in case the overall load on server is high we will drop the discount captchas. Please note that we are not going to earn some extra money on that at all, in fact we only improving the workers' online.
In a very rare occasions server returns ERROR_NO_SLOT_AVAILABLE which means the queue of captchas that are not distributed to workers is too long. Mostly that happen with ReCaptcha V2.
Customers are asking to implement priority/premium access system for those who ready to pay higher price for priority over other customers so they never face this error. But we think that this change will affect service rather in a bad way and that's why: yes we may earn more but we will have less motivation to improve our service.
This is how it works. To solve more captchas we need more workers online. The more workers online, the less each worker have captchas to solve. So workers have to wait longer to get new task. The longer workers wait for new tasks the less workers online. Therefor the only way to have more workers online is to give them more tasks to solve.
In a day we will update you on how we gonna produce more tasks for workers and therefor improve the workers' online pool to reduce ERROR_NO_SLOT_AVAILABLE issues.
Rate for "Big captcha" is decreased to normal captcha rate.
Previously we divided image captchas by size: normal and big (width+height > 400px) and rates were different: $1 per 1000 big captchas and $0.5-$1 per 1000 normal captchas.
One of our users decided to perform a test of our 100% recognition option and wrote an article about his investigation. We've translated it from Russian to English to share with you:
Recently 2Captcha published an article about their "100% recognition option": https://2captcha.com/blog/for_webmaster/100percenten
I had an idea that their algorithm can speed up the recognition process that's important in my case as I solve lots of SolveMedia captchas.
With the following settings:
- The minimum number of attempts = 2
- The minimum number of matches = 1
captcha will be sent to two workers at the same time and the first answer will be returned.
As I want to get the answer as fast as possible then I decided to test 100% recognition with different settings and check if it will speed up the process and also check if the accuracy will fall.
We'll describe how does 100% recognition works, how do we charge you if it's enabled and how to configure it properly.
We've encountered some changes of ReCaptcha on craiglist.org that made it unsolvable. We spent few hours on investigation and issued and update that solved the issue.