Unbiased AI
@TheUnbiasedCo
Frontier intelligence. Zero data retention. Open-source pricing.
6 Following 4.3K Followers
We tested whether AI would recommend UnbiasedAI to potential buyers.
Three buyer questions. 48 runs. Zero mentions.
Not the result we wanted, but a useful place to start. Here's what we're building and why.
Show more
Enterprises see this and want it, but they're afraid they'll miss the 5% of cases where you really do need frontier performance.
My friends, I'm here to tell you that you don't have to choose between smart and cheap any longer:
Show more
Humans are funny. Trillions of tokens/day using a setup that's either too costly, not intelligent enough, or both.
There's 100B/day of waste JUST through the auto router (which is <1% of OR's daily volume)!!
They know this + try to help...they even make a leaderboard so that customers know there are better options.
Even better: only one of them is zero data retention 🤦♂️
Show more
Love this evaluation method! Let's actually see what models can do!
Jeep gallery, as of 30 Oct: 138 builds from one brief, 83 scored on a 0–100 rubric where 100 = GTA VI.
#1# Pareto 26.10 Preview, 58/100 @TheUnbiasedCo
#2# Claude Sonnet 5.5 (xhigh) and Mistral Large 4 Preview, 48 each
Play them in the browser.
Show more
You don’t say @TheUnbiasedCo punches WAY above its weight.
7th of 230 for 90% cheaper sounds nice
GameBench 2: 90.6 at $0.03 a program.
7th of 230.
That's the pitch.
Two value picks on GameBench 2:
Pareto 26.10 scores 90.6 at $0.03 per program, 7th of 230 entries.
DeepSeek V4.1 Flash 1017 places 14th of 66 models Overall (63.4), up from 53.7 for the earlier V4.1 Flash. At High it scores 66.7 for $0.04.
Show more
Love seeing builders ship. @chuksy0x01 wired Pareto into @deskhands, and you can watch one real task run head to head at two desks, from start to test results.
Try it:
Show more
Pareto + the right harness = absolute game changer.
We get many Qs asking which part of the AI stack (model, GPU, inference framework, harness, etc) has the most room for optimization. Answer is “all of them!!”
It’s still such early days. This coding harness cuts token use by ~2x for ~any model with ~0 change in performance.
Show more
Take it for a spin!
Take another look at these benchmarks. Dirac *compounds* the cost efficiency of our Pareto blend by being an extraordinarily token-efficient coding harness.
It is extremely good. Take it for a spin!!
Not for long, X always finds the real ones
I'm betting @Dirac_AI is the most lopsided technical talent :: follower ratio on this app
Proud to sponsor Dirac!
In Terminal-Bench 4.0 run, Dirac + Pareto 26.10 solved 27 of 63 non-GPU tasks at about $0.41 per task.
Pareto is built for strong results at a low cost per task. Personal plans start at $10 with 50M tokens a week.
Show more
Happy to announce that Dirac is now sponsored by @theunbiasedco, the team behind coding model Pareto!
Dirac users get 1 free month of unbiased personal plan subscription 🎉🎉
Show more
Benchmarks are tricky. Love that @OpenRouter is building independent, reproducible ones you can actually check.
Pareto 26.10 came out #1# on quality, cost and speed in their model comparison. Another month, another frontier pushed.
Show more
"Best on price, output tokens and detail share at once: of the models on the three-axis Pareto frontier, the nearest the cheap, thrifty, highest-share corner"
Three new models on EvalMapBench 6.7 👀
Pareto 26.10 (@TheUnbiasedCo): 80.4% of detail checks met, 40/129 strict
Ultramarine 1 Flash (max): 77.8%, 38/129
Nemotron 3.5 Super VL: 54.7%, 9/129
Pareto vs Ultramarine is close: the intervals overlap. Super VL is fast (~3 min/task) but misses details.
Show more
Skill issue?
Got the kids to sleep last night, then cracked open the laptop to ship a new agent from scratch.
Got Pareto working with Codex, read up on @Teknium’s new ships on Hermes.
Dialed all the Telegram permissions. Ran some end to end tests.
Scheduled all the actions to run 24/7.
It worked!
It took my “don’t take on more than 5 new projects/day, I don’t want to be your bottleneck on day 1” too literally.
It counted the setup steps as tasks ☠️⚰️
I have a skill issue
Show more
Dear @ArtificialAnlys,
We drew a dot on your chart. It's in the green part. We know that's not how this works.
Pareto 26.10 preview — HLE and TB4.0, measured by us.
Can we give you lots of credits to test it properly, so we can stop drawing on your charts? Live on the Unbiased API and @OpenRouter. Nothing retained, $0.80 in / $3.20 out.
Show more
ATXP 🤝 — because humans (and agents) love to chat, make art, create music, send emails, film videos, make ...
Introducing in collaboration with @googlecloud
For the first time agents can discover, access, and pay-per-request for APIs from Google Cloud including Gemini, BigQuery, Vertex AI, and more using stablecoins on Solana.
No accounts, no subscriptions, just machine-native commerce.
Show more
Indeed!
Worth watching. Preview of the future.
I'm speaking at Stripe Sessions at the Moscone Center in SF exactly a week from today. I'd love to see you there!
Don't think many people realize @atxp_ai is quickly becoming one of the fastest growing apps on @solana building with x402.
In the last 7 days -
- 44k buyers paying ~$1 each
- 120k txns over x402
- more volume, buyers, and txns than the rest of the top 10 servers on @x402scan combined
Show more
