Key Takeaways
- Claude Mythos Preview, Qwen3.8-Max, and Grok 4 represent three distinct approaches to frontier scientific reasoning in 2026. Claude Mythos Preview leads on the two most discriminating academic benchmarks, GPQA Diamond at 94.6% and HLE with tools at 64.7%, but is not publicly available as a general API. Qwen3.8-Max leads on autonomous research and agentic tasks, posting the highest PaperBench score at 93.0 and the highest OSWorld-Verified score at 86.1. Grok 4 occupies the middle position, scoring 87.5% on GPQA Diamond and 44.4% on HLE, with the widest consumer access through xAI’s SuperGrok tiers.
- Claude Mythos Preview is Anthropic’s most capable model to date, announced on April 7, 2026 alongside Project Glasswing, a cross-industry initiative to secure critical software using frontier AI. Anthropic has withheld public API release because Mythos Preview’s autonomous cybersecurity capabilities are judged too powerful to ship without additional safeguards. Researchers and safety-focused teams can apply for restricted access; the model is not available for general consumer or developer use as of August 2026.
- Qwen3.8-Max, released August 3, 2026 by Alibaba’s Qwen team, is a 2.4-trillion-parameter mixture-of-experts model with 95B active parameters, a 1M-token context window, and native text, image, and video input. It prices at $2 per million input tokens and $6 per million output tokens via QwenCloud. Alibaba plans to open-source the weights alongside a smaller Qwen3.8-27B, which would make it the first open-weight Max-class model at this parameter scale.
- Grok 4 from xAI scores 87.5% on GPQA Diamond and 50.7% on the HLE text-only subset. Grok 4 Heavy is the first model reported to exceed 40% on Humanity’s Last Exam across the full multimodal dataset. The Grok 4.3 API launched April 30, 2026 at $1.25 per million input tokens and $2.50 per million output tokens with a 1M context window. Grok 4.5 is available at $2.00 per million input and $6.00 per million output. Consumer access is through SuperGrok Heavy at approximately $300 per month.
- On GPQA Diamond, the ordering is Claude Mythos Preview at 94.6%, Grok 4 at 87.5%, and Qwen3.8-Max’s score is not independently verified as of August 2026. On PaperBench, a benchmark measuring ability to replicate machine learning research papers, Qwen3.8-Max posts 93.0, higher than any published score from Claude Mythos Preview or Grok 4, suggesting Qwen has a specific advantage in structured research tasks with well-defined evaluation criteria.
- All three models support long-context reasoning, but differ in window size and multimodal scope. Qwen3.8-Max and Grok 4.3 both offer 1M-token context windows; Grok 4.20 variants extend to 2M tokens. Claude Mythos Preview’s context window is not published for the restricted-access version. Qwen3.8-Max is the only model in this comparison with native image and video input alongside text; Grok 4 supports multimodal input through its consumer interface; Claude Mythos Preview focuses on text and code reasoning.
- Pricing and access are the practical differentiators as much as benchmarks. Qwen3.8-Max at $2/$6 per MTok and Grok 4.3 at $1.25/$2.50 per MTok are the two available API options; Claude Mythos Preview has no published API pricing because general access is restricted. Teams evaluating these models for production scientific reasoning workflows can only deploy Qwen3.8-Max or Grok 4 today. Claude Mythos Preview is relevant for teams tracking the frontier ceiling on scientific accuracy rather than deploying now.
Scientific reasoning is the benchmark category that most reliably separates frontier AI models from one another in 2026. GPQA Diamond, Humanity’s Last Exam, and PaperBench require genuine graduate-level and expert-level domain knowledge rather than pattern-matching on training data distributions, and performance gaps between models are large and meaningful on these tests in a way they are not on saturated coding or general language benchmarks.
Claude Mythos Preview, Qwen3.8-Max, and Grok 4 occupy the top positions across these scientific reasoning benchmarks as of August 2026, but they are not interchangeable. Each leads on different benchmark types, targets different use cases, and has different availability status. This comparison covers benchmark performance on scientific tasks specifically, the architecture and access differences that drive those results, and which model actually fits which team’s workflow today.
Quick Comparison: Claude Mythos vs Qwen3.8-Max vs Grok 4
| Feature | Claude Mythos Preview | Qwen3.8-Max | Grok 4 |
|---|---|---|---|
| GPQA Diamond | 94.6% | Not verified | 87.5% |
| HLE (with tools) | 64.7% | Not published | 44.4% full; 50.7% text-only |
| PaperBench | Not published | 93.0 (highest reported) | Not published |
| SWE-bench Verified | 93.9% | Not published | Not published |
| Context window | Not published (restricted) | 1M tokens | 1M tokens (4.3); 2M tokens (4.20) |
| Multimodal input | Text and code | Text, image, video | Multimodal (consumer interface) |
| API pricing (input/output per MTok) | Not available publicly | $2 / $6 | $1.25 / $2.50 (4.3); $2 / $6 (4.5) |
| Public API access | Restricted (apply only) | Yes (QwenCloud) | Yes (xAI API) |
| Open weights | No | Planned (weights pending) | No |
| Best for | Benchmark ceiling tracking, safety research | Autonomous research tasks, agentic science | Accessible frontier reasoning, real-time search |
What Is Claude Mythos Preview?
Claude Mythos Preview is Anthropic’s frontier reasoning model, announced on April 7, 2026 alongside Project Glasswing. Anthropic describes it as the highest-scoring model on record across the major frontier benchmarks: 94.6% on GPQA Diamond, 93.9% on SWE-bench Verified, 64.7% on Humanity’s Last Exam with tools, and 97.6% on USAMO 2026. On CyberGym, a benchmark measuring autonomous offensive cybersecurity capability, it scores 83.1%, which is the reason Anthropic withheld general public release.
Anthropic’s position is that Mythos Preview’s cybersecurity capability crosses a threshold where unrestricted API access poses real-world risk. The model is available only through a restricted access program for vetted researchers and organizations participating in Project Glasswing. For scientific reasoning specifically, GPQA Diamond at 94.6% means the model answers graduate-level physics, chemistry, and biology questions correctly at a rate that exceeds most human domain experts. The 97.6% USAMO score confirms that the mathematical reasoning underlying scientific problem-solving is also at or above expert human level.
What Is Qwen3.8-Max?
Qwen3.8-Max is Alibaba’s flagship model released August 3, 2026. It uses a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active parameters per forward pass, a design that allows the model to bring a large parameter count to bear on specialist tasks while keeping inference costs lower than a dense model of equivalent total size. The context window is 1 million tokens. The model accepts text, image, and video as native inputs, making it the most multimodal option in this comparison at the API level.
Qwen3.8-Max’s most notable scientific benchmark result is PaperBench at 93.0, the highest published score on a benchmark that measures a model’s ability to replicate machine learning research papers from scratch. It also posts 86.1 on OSWorld-Verified, above GPT-5.6 Sol Max at 83.2 and Claude Fable 5 at 85.0. Terminal-Bench 2.1, which measures ability to operate software through a command-line environment, is 86.6. The API is available through QwenCloud at $2 per million input tokens and $6 per million output tokens. Alibaba has committed to open-sourcing the weights, which would make it the first publicly available model at the 2.4 trillion parameter scale.
What Is Grok 4?
Grok 4 is xAI’s frontier model family, with the Grok 4.3 API version launching April 30, 2026. On GPQA Diamond, Grok 4 scores 87.5%, placing it second in this comparison behind Claude Mythos Preview. On Humanity’s Last Exam, Grok 4 Heavy achieves the first reported score above 40% on the full multimodal dataset, with the text-only subset reaching 50.7%. These results make Grok 4 the most accessible frontier model for scientific reasoning tasks that require broad domain knowledge across physics, chemistry, mathematics, and biology.
The Grok 4 family covers a range of price and capability points. The API offers Grok 4.3 at $1.25 per million input and $2.50 per million output with a 1M context window. Grok 4.5 at $2.00 input and $6.00 output and Grok 4.20 at the same price with a 2M context window are available for teams needing higher capability or longer context. On the consumer side, SuperGrok at approximately $30 per month gives access to standard Grok 4 models. SuperGrok Heavy at approximately $300 per month gives access to Grok 4 Heavy with multi-agent workflows. XAI also bundles web search, X network search, and code execution as API tools at $5 per 1,000 calls each, which makes Grok 4 well-suited for scientific research tasks that need real-time information retrieval.
Claude Mythos vs Qwen3.8-Max vs Grok 4: Feature-by-Feature Breakdown
GPQA Diamond Performance
GPQA Diamond is the most widely cited benchmark for graduate-level scientific reasoning. Questions require genuine domain expertise in physics, chemistry, and biology; human domain experts score approximately 70% on the benchmark, and scores above 90% are considered frontier-level. Claude Mythos Preview at 94.6% is the highest published score. Grok 4 at 87.5% is strong but approximately 7 percentage points below Mythos. Qwen3.8-Max’s GPQA Diamond score has not been independently verified as of August 2026; the model’s public benchmark releases prioritize agentic and research task metrics.
Humanity’s Last Exam Performance
Humanity’s Last Exam (HLE) is a 3,000-question benchmark designed to be unsolvable by current AI, covering mathematics, natural sciences, and humanities at the expert level. Claude Mythos Preview at 64.7% with tools and 56.8% without tools is the highest published score across both conditions. Grok 4 at 44.4% on the full dataset and 50.7% on the text-only subset is the second-highest reported score and the first to exceed 40% on the complete multimodal version. Qwen3.8-Max has not published an HLE score. The gap between Mythos and Grok 4 on HLE is approximately 20 percentage points on the full benchmark with tools, which is a meaningful margin on a benchmark where every percentage point represents genuinely difficult expert-level questions.
Autonomous Research and Agentic Scientific Tasks
PaperBench and OSWorld-Verified measure a different capability than GPQA or HLE: the ability to autonomously execute multi-step research workflows rather than answer individual questions correctly. Qwen3.8-Max leads on both. Its PaperBench score of 93.0 exceeds any published score from Claude Mythos Preview or Grok 4. Its OSWorld-Verified score of 86.1 outperforms Claude Fable 5 at 85.0 and GPT-5.6 Sol Max at 83.2. This pattern suggests Qwen3.8-Max’s mixture-of-experts architecture is particularly well-suited to sustained multi-step task execution, where different expert modules can be routed to the specific subtask demands of a longer research workflow.
Pricing and API Access
Grok 4.3 at $1.25 per million input tokens is the lowest-cost API option in this comparison for teams that can work within Grok 4.3’s capability profile. Qwen3.8-Max and Grok 4.5 are both priced at $2.00/$6.00 per MTok and are directly comparable on price. Claude Mythos Preview has no published API pricing because general API access is not available. Teams building production scientific reasoning pipelines today have two options: Qwen3.8-Max through QwenCloud or Grok 4 through xAI’s API. Both are available without a waitlist. The Qwen3.8-Max open-weight release, when it arrives, will add a third option for teams that need to run inference on their own infrastructure.
Context Window and Multimodal Scope
For long-document scientific reasoning tasks such as analyzing a full research paper, reviewing a clinical trial report, or processing multi-paper literature reviews, context length matters. Qwen3.8-Max and Grok 4.3 both offer 1M tokens; Grok 4.20 extends to 2M tokens. Qwen3.8-Max is the only model in this comparison with verified native image and video input at the API level, which is relevant for scientific tasks that involve charts, microscopy images, or experimental data visualizations. Grok 4 supports multimodal input through xAI’s consumer interface, with API multimodal support varying by version. Claude Mythos Preview’s specifications under the restricted access program are not publicly disclosed.
Who Should Use Which?
Teams that need the highest benchmark ceiling for scientific accuracy and can obtain restricted access should prioritize Claude Mythos Preview. At 94.6% on GPQA Diamond and 64.7% on HLE with tools: it is demonstrably the most accurate model on structured scientific question-answering tasks. The relevant use cases are safety research, pharmaceutical research validation, and academic benchmarking where accuracy on expert-level scientific questions is the primary constraint and deployment at API scale is not the goal.
Teams building autonomous research workflows, multi-step paper analysis pipelines, or agentic scientific tools that need general API availability should evaluate Qwen3.8-Max. Its PaperBench score of 93.0 is the highest published score on a task that directly mirrors autonomous research execution. The $2/$6 pricing is competitive for API-scale deployment, the 1M-token context window handles full research papers and extended document sets, and the native video and image input is relevant for multimodal scientific data. The planned open-weight release will make it the only frontier-class model in this comparison that teams can deploy on their own infrastructure.
Teams that need accessible frontier reasoning with real-time web and X network search integration should evaluate Grok 4. The Grok 4.3 API at $1.25/$2.50 per MTok is the lowest-cost entry point in this comparison for frontier-level scientific reasoning. The SuperGrok Heavy tier at $300 per month gives consumer access to Grok 4 Heavy with multi-agent capability and the highest HLE score on the full multimodal benchmark. The built-in web search and X search tools at $5 per 1,000 calls make Grok 4 specifically well-suited for scientific research that requires combining trained reasoning with live information retrieval.
Our Verdict
Claude Mythos Preview sets the benchmark ceiling on scientific reasoning in 2026, but the restricted access makes it irrelevant for most production deployments today. Between the two available models, Qwen3.8-Max and Grok 4 serve different strengths: Qwen3.8-Max leads on autonomous research task execution and agentic workflows, while Grok 4 leads on accessible academic question-answering benchmarks (GPQA, HLE) with the widest range of consumer and API pricing options. For teams that cannot wait for Mythos access and need verified strong performance on scientific reasoning at the API level, Grok 4 is the clearer benchmark performer today; for teams that need autonomous research execution and are willing to work with newer, less independently validated scores, Qwen3.8-Max is the better fit.
Frequently Asked Questions
Which model scores highest on GPQA Diamond in 2026?
Claude Mythos Preview scores highest on GPQA Diamond with 94.6% as of August 2026, followed by Grok 4 at 87.5%. GPQA Diamond tests graduate-level scientific reasoning across physics, chemistry, and biology at a difficulty level where human domain experts score approximately 70%. Qwen3.8-Max has not published a verified GPQA Diamond score. Claude Mythos Preview is not available through a public API; Grok 4 is available through xAI’s API and is the highest-scoring publicly accessible model on GPQA Diamond.
Is Claude Mythos available to use?
Claude Mythos Preview is available only through a restricted access program as of August 2026. Anthropic announced it on April 7, 2026 but withheld general release because the model’s autonomous cybersecurity capabilities scored 83.1% on CyberGym, a threshold Anthropic judged too powerful for unrestricted deployment. Vetted researchers and organizations participating in Project Glasswing can apply for access. A general consumer or API release has not been announced. For practical use today, Qwen3.8-Max and Grok 4 are the available frontier alternatives.
What is Qwen3.8-Max and who makes it?
Qwen3.8-Max is Alibaba’s flagship AI model, released August 3, 2026 by Alibaba’s Qwen research team. It uses a mixture-of-experts architecture with 2.4 trillion total parameters and 95 billion active parameters, making it one of the largest MoE models released to date. It accepts text, image, and video input natively and offers a 1M-token context window. The API is available through QwenCloud at $2 per million input tokens and $6 per million output tokens. Alibaba has committed to open-sourcing the weights, which would make it the first publicly available model at this parameter scale.
How does Grok 4 perform on Humanity’s Last Exam?
Grok 4 Heavy scores 44.4% on the full Humanity’s Last Exam dataset and 50.7% on the text-only subset, making it the first model to exceed 40% on the complete multimodal version of the benchmark. Claude Mythos Preview scores higher at 64.7% with tools, but is not publicly available. HLE is a 3,000-question expert-level benchmark designed to resist AI systems; any score above 30% is considered a significant frontier capability demonstration. Grok 4 Heavy is available through the SuperGrok Heavy consumer subscription at approximately $300 per month, which includes multi-agent workflow access.
What is PaperBench and why does Qwen3.8-Max lead on it?
PaperBench is a benchmark that measures a model’s ability to replicate the results of machine learning research papers autonomously, including understanding the experimental setup, implementing the code, running experiments, and producing results that match the original paper’s findings. Qwen3.8-Max scores 93.0 on PaperBench, the highest published score as of August 2026. The lead is likely due to Qwen3.8-Max’s MoE architecture, which can route subtasks (code understanding, mathematical reasoning, structured execution planning) to specialist parameter subsets within the model, producing more coherent multi-step research task execution than a single-dense-model architecture.
How much does Grok 4 cost?
Grok 4 is available at multiple price points through xAI’s API and consumer tiers. The Grok 4.3 API costs $1.25 per million input tokens and $2.50 per million output tokens with a 1M context window, making it the lowest-cost frontier scientific reasoning API in this comparison. Grok 4.5 and Grok 4.20 cost $2.00 input and $6.00 output per million tokens; Grok 4.20 adds a 2M-token context window. Consumer subscriptions range from SuperGrok at approximately $30 per month for standard Grok 4 access to SuperGrok Heavy at approximately $300 per month for Grok 4 Heavy with multi-agent workflows and the highest benchmark scores.
Which model is best for scientific research tasks in 2026?
The answer depends on the task type. For structured expert-level question-answering (GPQA Diamond, HLE), Claude Mythos Preview is the most accurate model but is not publicly available; Grok 4 is the strongest available option. For autonomous research task execution, such as replicating experiments, analyzing research papers at scale, or running multi-step literature workflows, Qwen3.8-Max leads with its PaperBench score of 93.0 and OSWorld-Verified score of 86.1. For teams that need real-time scientific information alongside trained reasoning, Grok 4’s built-in web search and X integration at $5 per 1,000 calls is a practical advantage that neither Qwen3.8-Max nor Claude Mythos Preview matches at the API level.
