AIWant Mythos? Try Winthos!
Bringing unbiased frontier AI to people of all classes.
At Winthos, we aim to win (together)!
Our models, featuring Beta AGI, match or surpass frontier models in coding and cybersecurity while running 10x to 20x cheaper. That means saving money without laying off employees, or hiring more US-based workers instead.
AI as a tool to boost human productivity, not replace it.
Evolution = (AI: All Knowledge, Infant Thinking) + (Human: Life-Earned Critical Chain-of-Thought).
Read our missionMEET THE WINTHOS FAMILY
Specialized knowledge for different kinds of work. Every Winthos model supports vision and voice.
Multidisciplinary health knowledge, built for everyday support. Download and load Health Mini in Winthos AI Chat & Anti-Virus, a security suite with AI chat for mental wellbeing, farming, first-aid education, fitness, rehabilitation, and therapy support.
Explore the appA compact cybersecurity model with room for extensive code and context. Winthos plans to publish its weights on Hugging Face after the community reaches more than 10 million social-media followers.
Open-weight release is conditional on that milestone. Public weights are not offered on this page.
Explore the Cyber familyMedical knowledge meets specialized U.S. legal knowledge. Designed to support healthcare professionals worldwide and attorneys working across clinical information, case records, and complex evidence.
Clinical-grade performance is a stated goal. Clinical validation and regulatory status are not established here.
Explore Winthos LegalWinthos's flagship MoE model for demanding software engineering and cybersecurity work. Positioned for frontier-class tasks with a large context window and multimodal understanding.
View reported resultsModel specifications and Health Mini availability are supplied by Winthos. Benchmark and clinical claims are identified separately from independently verified results.
HIGH-FIDELITY REASONING
When an error is not an option, single-pass generation falls short. Ultra Mode runs internal debate rounds between independently trained models until both converge on the exact same verified answer.
ZERO HALLUCINATION GOAL
Ultra Mode takes slightly longer to generate an answer, but it is built to be hallucination-resistant. For critical decisions in clinical care, cybersecurity, and law, Winthos executes structured debate rounds between independently trained models, including IBM Granite 4.2 8B Cyber/Clinical DUS and Qwen 3.8 Flash Next PLE.
Each model generates independent reasoning, cross-examines citations, contraindications, and arithmetic, and emits a final answer only upon unanimous agreement.
WHAT WILL YOU MAKE POSSIBLE?
Your experience brings the context. AI helps you explore the possibilities. You stay in charge.
Understand an unfamiliar topic. Find the words you were looking for. Break a big project into a first step you can actually take.
Explore the security suiteA community garden, from first thought to first season.
Who will the garden serve, and what would a useful first season look like?
Map the space, water access, budget, and people willing to help.
Draft a one-page proposal for neighbors to discuss.
WINTHOS AI CHAT & ANTI-VIRUS
A security suite with an AI conversation at its heart. Winthos AI Chat & Anti-Virus brings device protection together with Winthos Health Mini.
Download Health Mini once, then run it locally in the app without an internet connection, according to Winthos. Its 256K context window, vision, and voice capabilities support conversations beyond a short text prompt.
Load when you need it. Eject anytime to release the model’s RAM for other work. Device, model build, and context settings can affect memory use; measurements are pending.
Health Mini download informationThe free Winthos AI AV Engine brings threat detection and remediation into the security suite, alongside Health Mini chat. Explore the AV Engine ↗
A focus spanning medical specialties, psychology, physical therapy, personal training, and everyday wellbeing. Breadth of information is not a substitute for a licensed professional's judgment.
From farmers working through an unfamiliar problem to people preparing for a health appointment, Health Mini is designed to make specialist information easier to approach.
Supported devices, installation requirements, and data-handling details belong with the app's release documentation.
WINTHOS AI AV ENGINE
Included free in Winthos AI Chat & Anti-Virus. AI-powered detection and remediation, designed for everyday devices.
Reported by Winthos
Reported by Winthos
Reported engine size, not full app size
No separate engine charge
ACROSS YOUR DIGITAL DAY
Winthos describes an engine built to detect malware and phishing in messages, email, and notifications, and to remediate supported threats.
Explore app availabilityDetect suspicious content and support threat remediation.
Identify phishing signals across supported integrations.
A reported engine footprint below 200 MB. Runtime memory and full app size require separate measurements.
Detection and false-positive figures are developer-reported, with test data and independent evaluation pending. Coverage and remediation depend on the operating system, permissions, and integration. These figures do not establish detection of every threat or a verified “first of its kind” claim.
WINTHOS HEALTH MINI / 256K CONTEXT
Built to bring information across medical specialties, psychology, movement, and practical life into a conversation you can understand.
Doctors make mistakes, and independent reviews save lives. Fact-check prescriptions, lab values, and clinical explanations before major decisions. Like Ronald Reagan said, "Trust, but verify." Attach diagnostic imaging, including MRI scans, giving the model full visibility of your records. Persistent local memory retains your clinical background across sessions.
Structured, non-judgmental conversational support designed to help you de-escalate interpersonal conflict, process acute stress, and maintain emotional stability, available 24/7 without appointment delays.
Beta computer vision form tracking designed to monitor posture and biomechanics during workouts, whether your focus is efficient fat loss, progressive strength building, or rehabilitation from physical injury.
Scan meals to evaluate nutrient density, identify micronutrient deficiencies, and calculate dietary protocols tailored to your metabolic markers and genetic profile, unlike superficial quiz companies that sell generic supplements.
Discuss crop observations, track farm records, and analyze plant images to assess soil and disease markers. Agricultural decisions are strengthened by combining specialized models with practical local knowledge.
When an accidental injury, animal bite, or trauma occurs, Winthos delivers immediate, structured first-aid guidance to bridge the critical window until professional paramedics arrive.
Health Mini supports educational access to medical information, running directly on your local device. The model does not replace a licensed physician, psychologist, physical therapist, or emergency responder.
Emergency Notice: In any life-threatening emergency, call 911 immediately. In an emotional or mental health crisis, call or text 988. Diagnostic interpretation of medical imaging requires qualified clinical oversight.
How to evaluate health information · NIHWINTHOS LEGAL
0.5M context · Vision + VoiceMedical knowledge and specialized U.S. legal knowledge, brought together for professionals working with consequential information.
Winthos Legal matches or exceeds legacy legal models on standard benchmarks while operating at one-fifth to one-tenth the cost. It synthesizes extensive medical knowledge with indexed case law, judicial opinions, and statutory reasoning frameworks, breaking down the predatory paywalls of traditional research databases.
Our structured intake dossier evaluates case records to uncover hidden causation angles, organizing claims across the four pillars of liability: duty, breach, causation, and damages. Draft pleadings, briefs, and client summaries maintain authentic stylometric cadence, adhering strictly to Federal Rule of Civil Procedure 11(b) and ABA Formal Opinion 512.
“Clinical grade” describes Winthos's intended engineering standard, not regulatory authorization. Licensed attorneys remain responsible for legal analysis, citations, and court filings. Legal specialization is U.S.-focused.
THE WINTHOS CYBER FAMILY
Two models, each with a 1M context window and vision and voice. Built for software engineering and cybersecurity workflows.
WINTHOS CYBER MINI
A compact model with a planned public weight release on Hugging Face. Winthos's stated release milestone is more than 10 million social-media followers.
The release plan is conditional. Weight availability and license terms will be published with the release.
WINTHOS CYBER
A Mixture-of-Experts model positioned by Winthos for frontier-class software engineering and cybersecurity work.
Illustrative value for this local preview. Winthos evaluation reports are pending. See benchmark comparisons ↗
PERFORMANCE, IN CONTEXT
Explore healthcare, law, software engineering, and cybersecurity evaluations alongside a shared intelligence index.
The 95% Winthos figures below are illustrative placeholders requested for this local preview. They are not measured results. Benchmark reports and comparable evaluations are pending.
Health Mini is shown on health tasks. Winthos Legal spans healthcare and legal work.
Scroll horizontally to see every column.
| Model | HealthBench ProfessionalLength-adjusted score | Harvey Legal Agent BenchmarkFinal task score, not criteria pass rate |
|---|---|---|
| Winthos draft values Illustrative only. Not measured or ranked. | ||
| Winthos Health Mini256K context | 95% Draft only | Outside scope |
| Winthos Legal0.5M context | 95% Draft only | 95% Draft only |
| Published reference results Follow a score to its source. | ||
| Claude Fable 5.1 | 58.1%Opus 5 fallback for provider refusals. | No result sourced |
| GPT-6 Astra | 63.4% | No result sourced |
| Claude Opus 5 | 56.4% | 6.7% |
| GPT-5.6 Sol | 60.5% | 2.5% |
| Grok 4.7 | No result sourced | No result sourced |
| Qwen3.8 Max | No result sourced | No result sourced |
| GLM-5.3 | No result sourced | No result sourced |
| Kimi K3 | No result sourced | 10.83% |
| Gemini 3.8 Flash | 52.1% | 10% |
| DeepSeek V4 Pro | No result sourced | No result sourced |
Sources checked 21 September 2026: OpenAI evaluations · Google legal results · Vals legal results.
Settings and evaluation environments differ. These results are not a controlled comparison with Winthos. “No result sourced” means an exact-model result was not verified for this snapshot; it does not mean zero.
Compare Cyber Mini and Cyber across coding and cybersecurity evaluations.
Scroll horizontally to see every column.
| Model | Terminal-Bench 4.0OpenAI-reported evaluation | SEC-Bench ProOpenAI-reported evaluation |
|---|---|---|
| Winthos draft values Illustrative only. Not measured or ranked. | ||
| Winthos Cyber Mini11B · 1M context | 95% Draft only | 95% Draft only |
| Winthos CyberMoE · 1M context | 95% Draft only | 95% Draft only |
| Published reference results Follow a score to its source. | ||
| Claude Fable 5.1 | 55.8% | No result sourced |
| GPT-6 Astra | 57.9% | 85.4% |
| Claude Opus 5 | 52.6% | No result sourced |
| GPT-5.6 Sol | 37.3% | 79.1% |
| Grok 4.7 | No result sourced | No result sourced |
| Qwen3.8 Max | No result sourced | No result sourced |
| GLM-5.3 | No result sourced | No result sourced |
| Kimi K3 | No result sourced | No result sourced |
| Gemini 3.8 Flash | 19.1% | No result sourced |
| DeepSeek V4 Pro | No result sourced | No result sourced |
Sources checked 21 September 2026: OpenAI software and cybersecurity evaluations.
Settings and evaluation environments differ. These results are not a controlled comparison with Winthos. “No result sourced” means an exact-model result was not verified for this snapshot; it does not mean zero.
Artificial Analysis Intelligence Index v4.3.2. One sourced snapshot of the selected frontier models.
Scroll horizontally to see every column.
| Selection rank | Model and configuration | Index score |
|---|---|---|
| 1= | Claude Fable 5.1max with fallback | 53 |
| 1= | GPT-6 Astramax | 53 |
| 3 | Claude Opus 5max | 51 |
| 4 | GPT-5.6 Solmax | 47 |
| 5 | Grok 4.7xhigh | 46 |
| 6= | Qwen3.8 Max0902 | 45 |
| 6= | GLM-5.3max | 45 |
| 8 | Kimi K3max | 44 |
| 9 | Gemini 3.8 Flashhigh | 41 |
| 10 | DeepSeek V4 Pro0813 · max | 36 |
| Unranked | Winthos Health Mini256K context | Not yet evaluated |
| Unranked | Winthos Legal0.5M context | Not yet evaluated |
| Unranked | Winthos Cyber Mini11B · 1M context | Not yet evaluated |
| Unranked | Winthos CyberMoE · 1M context | Not yet evaluated |
Source: Artificial Analysis leaderboard · Methodology v4.3.2. Checked 21 September 2026.
This is a selected-model comparison, not the complete global leaderboard. Winthos has no verified index score or rank. Draft benchmark percentages cannot be converted into Intelligence Index points. This index does not certify clinical performance, antivirus effectiveness, or vision and voice quality.
Winthos results will need a model version, benchmark version, test configuration, scoring method, and supporting report before receiving a measured score or rank.
START WITH WINTHOS HEALTH MINI
Download the model and load it in Winthos AI Chat & Anti-Virus. Health Mini availability is confirmed by the Winthos team.
The app and model download URLs have not yet been supplied for this website. This panel will link directly to the official release when those destinations are available.
View Health Mini specificationsAMBITION, WITH ACCOUNTABILITY
Trust should come from how a model performs and how openly its limits are explained.
Our training approach prioritizes authoritative material, including resources from NIST and NIH, with expert review as a goal for other sources. Dataset provenance and review coverage need to be documented, not assumed.
We aim to publish model versions, evaluation methods, hardware, and limitations alongside benchmark results. Coding, defensive security, medical, and legal evaluations must be assessed separately.
Lower inference costs, faster responses, and fewer unsupported answers are research priorities. Comparative results will follow measured, repeatable tests.
No AI is free from bias or mistakes. Our goal is balanced reasoning, clear uncertainty, and answers people can verify. Human judgment remains part of the process.
ENGINEERING PERSPECTIVE
Two centuries ago, suggesting that heavy machines would fly through the air invited ridicule; talking about landing on the moon sounded like sheer madness. Today, both are established engineering achievements.
We draw that exact parallel when applying disciplined training to cancer research and stubborn scientific challenges. We do not claim one model or one GPU cures cancer overnight. But by pairing high-efficiency AI compute with the life-earned critical reasoning of human researchers, breakthrough discoveries become achievable.
Operating on a budget under $30,000 proves that breakthrough AI does not require hundreds of millions of dollars in waste. Through Win's VPN Inc., an IRS 501(c)(3) nonprofit, public support expands open access: if you give someone a fish, you feed them for a day; teach them how to fish, and they eat for a lifetime.
Winthos explores new ways of combining AI's ability to work across information with the judgment people develop through lived experience. The founder's long-term ambition includes contributing to difficult scientific problems, including cancer research and more personalized health knowledge.
These are long-term research aspirations. Neither AGI nor a medical breakthrough is established by a model description or benchmark percentage. A proprietary training approach must be evaluated through transparent, reproducible evidence.
RESEARCH THAT INSPIRES US
RECURSIVE SELF-IMPROVEMENTThe Last AI Built by HumansMODEL SCALINGSOLAR 10.7B: Depth Up-ScalingINFERENCE EFFICIENCYDeepSeek-V4.1-Flash: KV Cache CompressionIndependent research acknowledgments. These papers do not validate Winthos's performance or imply an endorsement.
WHY WE'RE HERE
Technology is at its best when it gives people more agency. We believe AI should help a small team do ambitious work, help someone understand a difficult subject, and make knowledge less dependent on income.
Winthos means winning together. We want the gains from AI to create opportunity for people, including the people doing the work.
Explore our principlesHuman intelligence and artificial intelligence come together for a brighter tomorrow.
A symbol of guidance, hope, and new beginnings.
A brighter tomorrow for everyone.
A greater future. Together.
PROGRESS, WITH PEOPLE IN MIND
Technology does not exist in isolation. These are the values and economic policy ideas behind our work.
Many perspectives shape better decisions.
AI can inform the discussion.
People remain responsible for the choice.
Lower technology costs should give organizations room to invest in their people. We favor AI that expands human productivity, with thoughtful retraining and fair employment practices.
Our founder's experience in large organizations shaped concerns about contracting, unstable employment, and opaque outsourcing. We advocate clear compensation, transparent employment relationships, and accountability for the effect of automation on workers.
Public discussions about H-1B visas frequently confuse the statutory mechanics. The often-cited $60,000 threshold comes from the 1998 American Competitiveness and Workforce Improvement Act (8 U.S.C. 1182(n)(3)(B)), where it defined exempt workers to relieve employers of certain recruitment attestations. That threshold was set nearly three decades ago and was never indexed for inflation.
In major tech hubs, $60,000 is barely entry-level compensation. Raising this statutory floor to $110,000 ensures that foreign sponsorship is reserved for genuine specialized talent rather than suppressing local wages.
In addition, the statutory cap of 65,000 plus 20,000 advanced degree exemptions (8 U.S.C. 1184(g)) operates as a lottery with multi-month delays. The real defect is third-party staffing agencies and consulting intermediaries gaming the registration lottery to undercut domestic university graduates.
Wage requirements · U.S. Department of Labor ↗Across modern corporate America, many organizations maintain only executive leadership on permanent payroll while outsourcing engineering, IT, and customer support to staffing vendors. Intermediary consulting firms routinely take 20% to 40% of the worker's billed rate while leaving workers vulnerable to sudden mass layoffs with minimal notice.
Global outsourcing takes this dynamic further. Large enterprises frequently route Tier 1 customer calls to overseas contractors who lack context on fundamental US systems, such as Personal Injury Protection (PIP) in health and auto insurance. Companies avoid direct domestic employment by contracting through intermediary entities whose frontline workers sit overseas. Meaningful reform requires closing these contracting loopholes, enforcing prevailing-wage transparency, and prioritizing domestic workforce development.
The EB-3 other workers category for unskilled permanent labor is legally capped at only 10,000 visas per year worldwide (8 U.S.C. 1153(b)(3)(B)), creating severe labor shortages in agriculture, meatpacking, and seasonal facilities where domestic applicants are scarce.
A workable bipartisan solution is establishing a five-year conditional legal status tied to verified agricultural and seasonal work. Completing five years of satisfactory service in essential food supply or seasonal infrastructure would earn permanent legal residency. This provides agricultural producers with a stable, legal workforce, keeps family farms solvent, lowers grocery store prices for consumers, and provides immigrants with an honest, protected path to legal status.
Building datacenters in already strained metropolitan regions, such as Northern Virginia, overburdens local electrical grids and water utilities. Siting facilities in areas with surplus renewable power and balanced municipal resources protects residential utility rates while sustaining technological progress. Winthos prioritizes energy-efficient, environmentally conscious computing infrastructure.
Data-center cooling and water efficiency · DOE ↗Winthos is being built on a founder-reported budget of under $30,000, excluding our team's unpaid work and time. We believe useful AI should reach beyond the people who can afford the most expensive tools.
As a disabled creator overcoming technical barriers, the founder credits AI assistants with helping bridge technical knowledge gaps and physical limitations. Their identity is withheld here for personal safety. Trust must be built through documented work, clear governance, and honest limits.
Through our IRS 501(c)(3) nonprofit framework, community support and donations directly fund compute and open access for families and small businesses who need it most. If you give someone a fish, you feed them for a day; teach them how to fish, and they eat for a lifetime.
ALREADY HAVE AN AI SUBSCRIPTION?
Put Winthos alongside the tools you already use. Check an answer against sources, or work through an initial code review before bringing in a more expensive model.
Bring an answer from another AI and use Winthos to check it against primary sources. Follow the evidence and resolve disagreements before relying on the answer.
Hallucination-free answers are our goal. That claim awaits evaluation; a second model’s agreement is not proof that an answer is correct.
Our approach to evidenceUse Winthos for an initial review of your code, address its findings, then run a final audit with your preferred premium AI. Keep tests and human judgment in the process.
Our target is to identify 50% to 100% of the confirmed bugs and vulnerabilities that the final AI audit would find. Comparative evaluations are pending; this is a target, not a measured result or guaranteed coverage.
Potential savings come from reducing metered premium usage or avoiding a higher-priced tier. Moving work between fixed-price subscriptions does not automatically lower either bill.
Explore Winthos SpecialWINTHOS SPECIAL
Three monthly plans, half the planned standard price. Keep your special rate while subscribed, or save another 20% with annual billing.
Half the monthly price + double the usage = 4× usage per dollar.Compared with Winthos’s planned standard monthly price and allowance.
An accessible place to start
Planned standard price $20 / month
$10USD / month
12 special monthly payments $120
$96USD / year
Equivalent to $8 per month, billed annually.
For a more demanding workflow
Planned standard price $100 / month
$50USD / month
12 special monthly payments $600
$480USD / year
Equivalent to $40 per month, billed annually.
Our highest planned usage tier
Planned standard price $200 / month
$100USD / month
12 special monthly payments $1200
$960USD / year
Equivalent to $80 per month, billed annually.
THE FOUNDER’S RATE
Get 120 months of the $10 plan for one $500 payment. Your subscription ends after ten years, with no automatic renewal.
$500USD, paid once
About $4.17 per month over ten yearsCompared with $20 × 120 months = $2,400 at the planned standard monthly price.
Also saves $700 versus 120 payments at the $10 special monthly rate. The 2× usage bonus still ends on December 31, 2026.Designed for usage comparable to similarly positioned ChatGPT, Claude, and Gemini plans. Exact model access, allowances, and fair-use limits will be published before checkout opens; competitor limits vary.
Prices are in USD. Monthly and annual plans renew until canceled, with special rates retained while active. The usage bonus ends on January 1, 2027. Standard prices are planned reference prices, not prices previously charged.
Annual prices are 20% below twelve special monthly payments: $120 becomes $96, $600 becomes $480, and $1,200 becomes $960. The founder offer is a separate ten-year term.
PayPal billing is planned through the same Win’s VPN Inc. account. This website does not take payment or create an account; subscription access and checkout are being prepared.
WINTHOS API
API key creation is planned for accounts with an active Winthos Plus, Pro, or Max subscription. No API keys or live API calls are available from this website yet.
Contributor price, without requiring you to contribute your data to qualify. Final product privacy and data-use terms will be published before API access opens.
API usage is billed separately from subscription fees. Model availability, supported features, billing details, and rate limits will be published before API access opens.
When account creation opens, subscribers must review and explicitly accept the Terms and liability limits and Acceptable Use Policy before receiving a key. Keep keys private and use the models only within authorized work.
Published standard, uncached text API list rates, checked September 24, 2026. Winthos is a planned rate. Prices alone do not compare model capability, token efficiency, cache discounts, long-context surcharges, tools, taxes, or data practices.
Source prices: GPT-6 Sol, GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, and Kimi K3. Multiples divide each listed total by Winthos's proposed $6 total.
HUMAN REVIEW MATTERS
AI output can be wrong or incomplete, including in security, code, health, and legal work. We put our best effort into the models within a limited budget, but no result is guaranteed. Have a qualified person verify consequential answers, test generated code, and obtain authorization before security testing. You are responsible for how you use the output and for harm caused by your use, subject to applicable law.
Read the Terms and liability limits and Acceptable Use Policy. These documents are drafts for legal review before any account or payment flow opens.
HELP MAKE ACCESS POSSIBLE
Support the work of making useful technology accessible to more people. Contributions are directed to Win's VPN Inc. through its existing donation channel.
The donation page identifies the recipient and its terms. For tax status and documentation, visit Win's VPN's published records.
ACCESS & PRICING
Health Mini brings specialist knowledge into the Winthos security suite. Sustainable pricing for the wider model family is part of our access mission.
Explore Winthos Special App and model download informationBUILT WITH GRATITUDE
To the researchers who share their work, and the tools that help people build beyond their limitations: thank you.
Acknowledgments reflect the founder's account of the tools and research that supported this work. No partnership or endorsement is implied.
Winthos Health Mini is available to download and load in Winthos AI Chat & Anti-Virus, according to the Winthos team. The official download URLs are still being added to this website. Cyber Mini public weights are planned after Winthos reaches more than 10 million social-media followers. The workflow examples here are illustrations, not live model responses.
The benchmark section shows published competitor results with their sources. Winthos 95% values are draft placeholders, not measured results. All four Winthos models remain unranked in the Intelligence Index until comparable evaluations are available. Explore the comparisons ↗
API key access is planned for active Plus, Pro, and Max subscribers after account creation launches. The planned rates are $1 per 1 million input tokens and $5 per 1 million output tokens, billed separately from the subscription. Keys cannot be created on this marketing site today.
No. This marketing website does not accept uploads or run an AI model. Sensitive documents belong only in services with verified protections and appropriate professional oversight.
Donations go through Win's VPN Inc.'s existing channel. Cryptocurrency addresses match those published on its website. A donation is separate from a Winthos subscription; a transfer alone does not activate a model account. Confirm any subscription amount, network, and payment reference with support first.