Drex Max: $0.06/M tokens, 67.59 on Decision Index

Nace AI has released Drex Max, a multimodal decision model designed for computer-use agents. The model accepts text and images to return probability distributions rather than generated text. Nace reports a Decision Index score of 67.59 on version 0.3.2, positioning it 6.9 points above TypeSafe’s Jev 1.14. At $0.06 per million input tokens with no output charges, Drex Max offers a pricing structure aimed at high-volume decision tasks in agent automation.

What changed

Drex Max is the first image-capable member of the Drex decision model family. Previous Drex variants accept only text input; Drex Max processes screenshots, scanned documents, charts, and video frames alongside text to output typed decisions. The model returns calibrated probability distributions for provided options rather than generating new tokens.

On Decision Index 0.3.2, Drex Max achieves 67.59 compared to Jev 1.14’s 60.7. On JevBench, Drex Max reaches 92.2 versus 86.7 for Jev 1.14. Performance varies across specific benchmarks. Drex Max leads on BFCL tool calling at 98.2 versus 96.4 for Jev and on OSWorld-Verified automation at 0.780 versus 0.734. On InfographicVQA, which tests chart and infographic comprehension, Drex Max scores 98.8. The model scores 92.3 on FinEntity financial tasks versus 81.9 for Jev 1.14.

Nace reports median latency of 174 milliseconds per screenshot decision against 253 milliseconds for Jev 1.14 on the same test set of 243 questions. On that same set, Drex Max processed 100 percent of screenshots while Jev 1.14 rejected 15.9 percent for exceeding size constraints.

Pricing stands at $0.06 per 1 million input tokens. Output tokens incur no charges. New accounts receive $25 in signup credit, and failed requests incur no cost. The model runs exclusively on Nace’s hosted API under the identifier “drex-max”; no open-weight or quantized variants have been announced. Model parameters remain undisclosed.

Drex Max handles up to 8 images per request, each up to 16 MiB in size, with a 32 MiB total budget per request. Each image is resized to a maximum of 1,638,400 pixels. The API supports up to 262,144 tokens of context per question, including image tokens. Each image costs up to 1,602 tokens per question.

Why it matters

Drex Max addresses a practical constraint in computer-use agent design. Agents automating browser and desktop workflows typically make dozens of decisions per task, each triggered by a screenshot. Previously, builders faced a choice between cheaper text-only models and expensive general-purpose frontier models for handling visual input.

At $0.06 per million tokens, Drex Max costs substantially less than frontier alternatives. This pricing structure compounds savings for agents making hundreds or thousands of decisions daily.

The 174-millisecond latency advantage matters for user-facing automation. A 79-millisecond improvement over Jev 1.14 appears marginal but compounds when agents run sequential tasks. Decision latency bottlenecks can add 1-2 seconds per workflow. Full screenshot acceptance also eliminates error-handling overhead; rejection-retry loops add latency and complexity.

Drex Max’s vision performance shows uneven coverage. On Decision Index Vision suite components, it scores 74.5 versus 63.9 for GPT-6 Luna Decisions. On charts and infographics it reaches 98.8. On finance benchmarks like FinEntity it scores 92.3. On banking request routing (BANKING77) it reaches 90.0 versus 79.7 for Jev 1.14. On knowledge tests like MMLU-Pro, however, it scores 72.5 versus 82.1 for Jev 1.14. On Hateful Memes moderation it trails 70.6 to 71.2 versus GPT-6 Luna Decisions.

The fundamental tradeoff is that Drex Max targets decision-making, not reasoning or knowledge recall. If your agent needs to classify customer inquiries, identify document types, or route messages by intent, Drex Max fits the requirement. If it needs open-ended reasoning or problem-solving, frontier models remain stronger despite higher costs.

What to test

Before deploying Drex Max to production, evaluate the following dimensions:

• Benchmark methodology: All scores are self-reported by Nace. Cross-vendor comparisons use different evaluation harnesses and may not control for prompt formatting or data distribution differences. Run benchmarks on decision tasks specific to your workflow before committing to production.

• Vision performance on your images: Drex Max’s vision benchmarks cover charts, infographics, documents, and photographs. If your agent processes a different image category such as mobile interfaces, handwritten forms, or low-contrast scans, test performance locally with representative samples.

• Latency under production load: Nace reports 174 milliseconds median latency per screenshot on single requests. Concurrent requests, rate limiting, and peak traffic patterns may behave differently. Test with your expected request volume and concurrency levels.

• Cost extrapolation: An agent making 1,000 decisions daily with 500 average input tokens per decision costs approximately $30 monthly at the published rate. Calculate this for your projected volume and compare to frontier model pricing under your negotiated contract.

• Image handling constraints: Drex Max rejects images with any side exceeding 1,568 pixels or under 280 pixels. If your screenshots fall outside this range, test resizing logic in your preprocessing pipeline. Verify that the 8-image and 262K-token limits accommodate your workflows.

• Probability calibration: Drex Max returns probability distributions rather than binary decisions. For example, if it returns 0.51 for option A and 0.49 for option B, confirm your downstream systems can safely handle such narrow margins. Verify confidence scores align with your risk tolerance.

• Compliance requirements: Nace reports SOC 2 Type II certification and mentions self-hosted or hybrid deployment options. If data residency or audit requirements are strict, verify these claims with Nace’s trust center before proceeding.

Common questions

What is the difference between Drex Max and other Drex models?

Drex Max is the only Drex model that processes images. Drex 1.5 is text-only and costs $0.04 per million input tokens. Drex 2.0 offers open weights for self-hosting but does not support images. Choose based on whether your agent needs visual input; if not, Drex 1.5 or 2.0 may offer different speed or cost tradeoffs.

Can I run Drex Max on my own infrastructure?

No. Drex Max is available exclusively on Nace’s hosted API. Drex 2.0 does offer open weights for self-hosting, though without image support. If on-device inference is critical, evaluate open-weight alternatives.

How does Drex Max compare to frontier models for agent automation?

Frontier models like GPT-6 Sol and Claude Opus 5.5 excel at reasoning and open-ended tasks but cost 50 to 170 times more per decision. Drex Max optimizes for closed-question scenarios where you supply options and the model returns a probability. If your agent’s decisions fit that pattern, Drex Max is cheaper and faster.

Is Drex Max better than Jev 1.14?

On the Decision Index, Drex Max scores 6.9 points higher. On MMLU-Pro knowledge tests, Jev 1.14 leads 82.1 to 72.5. On vision benchmarks, Drex Max wins on documents and charts but trails on some reasoning tasks. The best choice depends on your specific requirements and task distribution.

The conclusion

Drex Max provides a cost-effective point for multimodal decision models. At $0.06 per million tokens and 174 milliseconds per screenshot, it undercuts frontier models and matches or beats text-only competitors on speed. The Decision Index score of 67.59 represents a meaningful gap above Jev 1.14, though the margin narrows on knowledge and reasoning benchmarks outside the model’s optimization target.

The main limitation is that scores are self-reported and cross-vendor comparisons lack controlled methodology. Nace’s decision to keep weights closed and restrict to hosted API limits deployment flexibility. For teams needing multimodal agent decisions at scale with hosted infrastructure, Drex Max merits evaluation. For teams requiring open weights, on-device inference, or frontier-level reasoning capabilities, other options remain better fits.

Monitor for future updates on model size transparency, open-weight variants, and third-party benchmarks. Also test how Drex Max performs on vision tasks outside published benchmarks, as agent use cases such as mobile UI automation, form field recognition, and desktop interface navigation often diverge from academic evaluation datasets.

Related reading

• MiniMax M3.1 Flash: $0.15/M tokens cuts agent costs 70%
• d1-ultra decision model scores 65 on AA index
• Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model