Frontier Models
Ox Alpha Matches Frontier Models on DeepSWE at 63 Percent Score
The anonymous model released August 20 delivers competitive long-horizon coding performance alongside DeepSeek V4 Pro while remaining free with a 1,048,576 token context window on OpenRouter.
Ox Alpha is an anonymous reasoning model released on August 20 2026 that matches mid-tier frontier AI performance on the DeepSWE coding benchmark with free access and a 1,048,576 token context window.
The emergence of Ox Alpha on OpenRouter introduces an anonymous provider capable of delivering frontier-adjacent results on a specialized software engineering benchmark. Developers have begun integrating the model into agentic workflows where extended context supports multi-step coding tasks. The free pricing structure differentiates it from commercial alternatives that typically charge per token for comparable capabilities.
What performance did Ox Alpha record on the DeepSWE benchmark?
Ben Davis evaluated Ox Alpha using the complete DeepSWE suite that includes 113 distinct tasks focused on long-horizon software engineering challenges. The final score reached approximately 63 percent which aligns closely with DeepSeek V4 Pro at the same level and sits near the 65 percent range reported for Grok 4.6 and Gemini 3.7 Flash. This outcome followed an earlier subset test on 10 tasks that produced an 80 percent pass@1 rate surpassing Fable at 65 percent and GPT-5.6 Sol at 52 percent.
The discrepancy between the subset and full benchmark results underscores the value of comprehensive testing for agentic models. Subset evaluations can overstate capabilities when tasks share similar patterns while the full suite reveals performance across varied software engineering scenarios. Davis observed that the 63 percent figure better reflected sustained usage patterns after extensive interaction with the model.
What technical specifications define the Ox Alpha release?
Ox Alpha launched on August 20 2026 through the OpenRouter and OpenCode platforms with a context window sized at 1,048,576 tokens. The listing specifies free access and routes requests to a single stealth provider that maintains high uptime without disclosing its identity. Input modalities include text images and video which enables multimodal agentic applications beyond pure text coding tasks.
The model targets coding agentic workflows and production workloads according to the OpenRouter description. Its design emphasizes sustained reasoning over extended sessions which aligns with the requirements of the DeepSWE benchmark. No additional pricing tiers appear in the initial release leaving users with unlimited free usage subject to provider capacity.
| Model | DeepSWE Score | Context Window | Access Model | Input Modalities |
|---|---|---|---|---|
| Ox Alpha | ~63% | 1,048,576 tokens | Free on OpenRouter | Text image video |
| DeepSeek V4 Pro | ~63% | Not specified in brief | Commercial | Not specified in brief |
| Grok 4.6 | ~65% | Not specified in brief | Commercial | Not specified in brief |
How does the DeepSWE benchmark evaluate model capabilities?
DeepSWE functions as a long-horizon software engineering benchmark that presents 113 tasks designed to test sustained agentic performance. All models undergo evaluation through the mini-swe-agent framework to ensure consistent execution conditions across submissions. The benchmark received an update on August 20 2026 coinciding with the Ox Alpha release date.
A 63 percent score indicates reliable handling of the majority of complex coding scenarios while leaving room for improvement on the remaining tasks. This level of performance supports practical deployment in production environments where partial automation of software engineering workflows delivers measurable productivity gains.
- Review initial subset results that indicated 80 percent pass rate on 10 tasks.
- Execute the complete 113-task DeepSWE evaluation to obtain the 63 percent score.
- Compare outcomes against DeepSeek V4 Pro Grok 4.6 and Gemini 3.7 Flash.
- Assess the 1,048,576 token context window for suitability in long agentic sessions.
- Monitor OpenRouter uptime and any future provider disclosures.
What market implications arise from the Ox Alpha release?
Free access to a model that matches established frontier performance on a demanding benchmark may compel commercial providers to reconsider pricing for coding-focused offerings. Independent developers and smaller teams gain an entry point to high-context agentic tools without incurring usage fees which could accelerate experimentation in software engineering automation.
The anonymous provider model raises questions about long-term reliability and feature updates. Stakeholders in enterprise settings may hesitate to adopt the model for mission-critical workloads until the source identity and support commitments become clearer. OpenRouter continues to serve as the primary distribution channel with zero listed price and consistent availability.
What expert reactions have surfaced regarding Ox Alpha?
Ben Davis documented the testing process in detail noting that the full benchmark result aligned more closely with practical experience than the initial subset. The developer highlighted extensive personal usage of the model and described it as a very good model suitable for ongoing agentic work.
Actual DeepSWE run on the ox alpha mystery model is done. Ended at ~63% NOT the 80% my first subset test got, which makes way more sense. I've been using this thing a ton and it is definitely a very good model.Ben Davis, Developer
What developments are anticipated for Ox Alpha and similar releases?
Continued community testing will likely produce additional performance data across specialized coding domains not covered in the initial DeepSWE run. The combination of free access and extended context may inspire competing anonymous or low-cost releases that target similar agentic use cases.
Future updates to the DeepSWE benchmark could incorporate new task categories that further differentiate model strengths in long-horizon software engineering. Observers will track whether the stealth provider maintains the current free tier or introduces changes to availability and modalities over time.
Frequently asked
What score did Ox Alpha achieve on the DeepSWE benchmark?
Ox Alpha recorded approximately 63 percent on the full 113-task DeepSWE benchmark according to evaluations conducted by Ben Davis.
When and where was Ox Alpha released?
Ox Alpha became available on August 20 2026 through OpenRouter and OpenCode as a free model with a 1,048,576 token context window.
How does Ox Alpha compare to other models on DeepSWE?
The 63 percent score matches DeepSeek V4 Pro and approaches the 65 percent range associated with Grok 4.6 and Gemini 3.7 Flash.
Sources
- OpenRouter — Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads. Released: Aug 20, 2026. Context: 1M. Price: Free. Modalities: Text, images and video as input.
- DataCurve — DeepSWE is a long-horizon software engineering benchmark with 113 tasks updated August 20, 2026. All models run on mini-swe-agent for consistency.
- X (Twitter) — Actual DeepSWE run on the ox alpha mystery model is done. Ended at ~63% NOT the 80% my first subset test got, which makes way more sense. I've been using this thing a ton and it is definitely a very good model.