Carbon-X Benchmarks: How It Stacks Up Against GPT-6.1 and Opus

Carbon-X, a 14.8B-parameter open-weight DNA language model built on Qwen3-14B, averaged 66.7% accuracy across five DNA benchmarks with reasoning disabled, just below GPT-6.1 Sol's 67.5% and ahead of Opus 5.5's 60.9%, at an estimated 1,800 times lower inference cost per problem. Analyzing 2.4 million ClinVar variants of uncertain significance would cost about $300 with Carbon-X versus $32,000 to $4 million with the frontier models tested, and the 14.8B-parameter weights fit on a single 80GB GPU in bfloat16. Carbon-X led every tested model on coding-variant effect prediction (74.9%) and next-base prediction (39.4%), while Opus 5.5 scored 1.1% on NextBase.

Carbon-X Benchmarks: How It Stacks Up Against GPT-6.1 and Opus Carbon-X hits 66.7% average across five DNA benchmarks, nearly matching GPT-6.1 Sol at roughly 1/1,800 the inference cost. What is Carbon-X and how does it benchmark against GPT-6.1 and Opus? Carbon-X is a 14.8B-parameter open-weight language model built for DNA understanding, trained jointly on genetic sequences and natural-language text. Across five DNA benchmarks covering variant-effect prediction, sequence understanding, and next-base prediction, it averages 66.7% accuracy with reasoning disabled. That places it just below GPT-6.1 Sol’s 67.5% and ahead of Opus 5.5, Gemini 4 Argon, GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, while costing an estimated 1,800 times less per problem than the frontier models it was measured against. TL;DR - Carbon-X averages 66.7% across five DNA benchmarks BioReason Coding, BioReason Non-SNV, GenomeQA, NextBase, FrontierVEP , nearly matching GPT-6.1 Sol’s 67.5% and beating Opus 5.5’s 60.9%. - The model runs with no extended reasoning , answering directly and using far fewer output tokens than frontier models that rely on chain-of-thought, which is the main driver of its cost advantage. - Cost estimates for analyzing 2.4 million ClinVar variants of uncertain significance come out to about $300 with Carbon-X versus $32,000 to $4 million with the frontier models tested. - Carbon-X leads every tested model on coding-variant effect prediction 74.9% and next-base prediction 39.4% , two tasks where several frontier models struggled badly. - It starts from Qwen3-14B and adds a hybrid tokenizer that encodes DNA as non-overlapping 6-mers alongside a standard text BPE vocabulary, letting one decoder handle both modalities in a single token stream. - The 14.8B-parameter weights fit on a single 80GB GPU in bfloat16, so labs can run variant analysis on sensitive or unpublished sequences without sending data to an external API. - Opus 5.5 posted a near-zero score 1.1% on NextBase , next-base prediction, despite leading on complex variant effects and single-base variant effects, showing frontier generalist models aren’t uniformly strong https://goremy.ai/blog/articles/gpt-6-1-sol-vs-claude-opus-5-5-benchmark-reality-check across genomics tasks. - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. How does Carbon-X actually process DNA? Carbon-X doesn’t bolt a DNA encoder onto a language model. It uses one decoder-only transformer for both DNA and text, with a hybrid tokenizer that splits the difference: text gets the standard Qwen3 BPE vocabulary, while DNA gets encoded as non-overlapping 6-mers, meaning each DNA token represents six base pairs. The full vocabulary comes to 157,184 tokens, 151,669 from BPE and 5,515 dedicated to DNA 4,096 6-mers, 1,364 tail tokens for sequences whose length isn’t a clean multiple of six, plus special markers for tagging DNA spans . This matters practically: you can paste a raw DNA sequence directly into a chat message, and the tokenizer auto-detects runs of at least six A/C/G/T bases and wraps them in <dna tags automatically. Mixed prompts, reference and mutated sequences side by side, context about the gene or strand, and plain English questions all live in the same input. The architecture itself is a 40-layer transformer with a 5120 hidden size, grouped-query attention 40 query heads, 8 KV heads , SwiGLU activations, and RMSNorm with QK-norm. It supports a 40,960-token context window, long enough to hold reference sequences spanning thousands of base pairs alongside a detailed prompt. How was Carbon-X trained to specialize in genomics? Carbon-X starts from Qwen3-14B’s pretrained weights and bolts on DNA token embeddings, initialized randomly and trained for about 10 billion tokens with the rest of the model frozen. Only after the DNA embeddings stabilize does the full model unfreeze for broader training. From there, mid-training runs for 302 billion tokens with a data mixture that’s 55% raw DNA spanning eukaryotic genomes, mRNA, and prokaryotic genomes , 20% general English text, 15% biology-specific text, and 10% instruction data. The team notes that learning rate was the single most sensitive setting at this stage: too low and the model never picks up DNA patterns, too high and it starts forgetting English. Supervised fine-tuning follows, built on 20 dataset subsets covering variant-effect reasoning, sequence classification, and basic DNA manipulation tasks like counting and searching within sequences. The final stage applies reinforcement learning with verifiable rewards, using GRPO across two rounds on roughly 150,000 selected prompts, with task-specific verifiers checking variant predictions, classification accuracy, and instruction-following. All training data went through three decontamination checks against the evaluation benchmarks: whole-span DNA overlap, text n-gram overlap, and sequence homology search, which matters for genomics benchmarks where near-duplicate sequences across species can otherwise inflate scores. Where does Carbon-X win and where does it lose? The benchmark table tells a mixed story, and that’s worth sitting with rather than smoothing over. Carbon-X takes the top score on two of five benchmarks: BioReason Coding 74.9%, predicting benign versus pathogenic effects of mutations in protein-coding regions and NextBase 39.4%, predicting the next bases in a DNA sequence . On the other three, it trails the best frontier model but stays competitive: 85.1% on BioReason Non-SNV versus Opus 5.5’s 94.8%, 63.6% on GenomeQA versus GPT-6.1 Sol’s 67.2%, and 70.4% on FrontierVEP versus Opus 5.5’s 76.2%. Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. The more striking result is how inconsistent the frontier models are. Opus 5.5 leads on complex variant effects and single-base variant effects but scores just 1.1% on NextBase, essentially failing at predicting the next bases of a sequence. Gemini 4 Argon scores 90.3% on BioReason Non-SNV but only 6.5% on BioReason Coding. These aren’t small gaps. They suggest that general-purpose reasoning models, even at their strongest settings, don’t transfer evenly across different genomics task types, while Carbon-X’s specialized training produces steadier performance across the board. Carbon-X also holds up as a general assistant, scoring 79.3% prompt-level strict accuracy on IFEval 84.0% averaged across the four IFEval metrics , which indicates it didn’t lose basic instruction-following ability in the process of specializing on DNA. Is Carbon-X worth using over a frontier model for genomics work? The deciding factor is almost entirely about cost and scale rather than peak accuracy. Carbon-X answers directly without extended chain-of-thought reasoning, so each query burns through comparatively few output tokens. Frontier models tested at their best configurations GPT-6.1 Sol and Opus 5.5 at “Max” reasoning effort generate much longer reasoning traces to hit their scores, and that reasoning is expensive at scale. The model card estimates that running all 2.4 million variants of uncertain significance in ClinVar, the public database of genetic variants, through Carbon-X would cost around $300. The same job through the frontier models evaluated would run from $32,000 up to $4 million, depending on which model and reasoning setting. For any lab or tool doing high-volume variant triage rather than one-off lookups, that gap changes what’s financially feasible. There’s also a data-locality argument. Carbon-X’s weights fit on a single 80GB GPU in bfloat16, which means researchers working with unpublished or sensitive genetic data can run the model entirely on their own infrastructure instead of routing sequences through a third-party API. For clinical or pre-publication research contexts, that’s often a hard requirement, not a nice-to-have. The tradeoff is accuracy ceiling. If a single prediction carries high stakes and cost is not the binding constraint, GPT-6.1 Sol’s 67.5% average and lead on FrontierVEP and GenomeQA may justify the premium. For bulk screening, research pipelines, or any workflow where a few hundred dollars versus tens of thousands matters, Carbon-X’s 66.7% average at a fraction of the cost is the more practical choice. Frequently Asked Questions What benchmarks was Carbon-X tested on? Five: BioReason Coding and BioReason Non-SNV variant-effect prediction from the research literature , GenomeQA multiple-choice sequence understanding across six sub-tasks , and two built for this evaluation, FrontierVEP single-nucleotide variant classification and NextBase next-base sequence prediction . How much cheaper is Carbon-X than GPT-6.1 Sol or Opus 5.5? The model card estimates roughly 1,800 times lower inference cost per problem, based on measured throughput on an H200 GPU at $5 per hour versus standard API pricing for the frontier models at their best reasoning settings. Can Carbon-X run on consumer or single-GPU hardware? The 14.8B-parameter weights fit on a single 80GB GPU in bfloat16 precision, which covers common datacenter GPUs but exceeds typical consumer GPU VRAM, so it’s not a laptop-friendly model despite its relatively small parameter count. Does Carbon-X use chain-of-thought reasoning like GPT-6.1 or Opus? No. Carbon-X was post-trained and evaluated with extended thinking disabled, answering directly. This is a deliberate design choice that keeps output token counts, and therefore cost, low. What base model does Carbon-X build on? Remy is new. The platform isn't. Remy is the latest expression of years of platform work. Not a hastily wrapped LLM. Carbon-X starts from Qwen3-14B, with a custom DNA tokenizer and vocabulary added on top, then trained through staged mid-training, supervised fine-tuning, and reinforcement learning with verifiable rewards on genomics tasks.