Ornith-1.0 Hits SWE-Bench Verified 82.4: What MIT-Licensed Agentic Coding at Frontier Level Costs You in 2026

A New Open-Source Frontier in Coding

DeepReinforce dropped Ornith-1.0 on Hugging Face on June 25-26, 2026. The release is a family of four MIT-licensed models post-trained for agentic coding: 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE, built on top of Gemma 4 and Qwen 3.5 base weights with a self-improving training loop the team calls "self-scaffolding."

The reported benchmark numbers for the family flagship (the 397B-MoE configuration): SWE-Bench Verified 82.4, SWE-Bench Pro 62.2, Terminal-Bench 2.1 77. The 35B-MoE variant — the one most likely to fit on commodity inference hardware — reports SWE-Bench Verified 75.6, SWE-Bench Pro 50.4, and Terminal-Bench 2.1 of 64.2 (Terminus-2 harness) / 62.8 (Claude Code harness).

These numbers put the open-source coding frontier within striking distance of Claude Opus 4.8, GPT-5.5, and the proprietary tier. The interesting question is no longer "can open weights do agentic coding well?" It's what does it cost to self-host vs. pay API rates? Let's run the math.

Self-Hosting Cost: 35B-MoE on Cloud GPU

The Ornith model card recommends an 8×80GB GPU node with tensor parallelism for the 35B-MoE variant — concretely, an 8× H100 80GB node or 8× A100 80GB node. Cloud pricing for an 8× H100 instance in mid-2026:

• AWS p5.48xlarge: ~$50-60/hour on-demand
• Lambda Labs 8× H100: ~$22-28/hour
• Together / RunPod / CoreWeave: ~$24-32/hour

At a midpoint of $28/hour, the monthly cost of running one inference node continuously is ~$20,000. At a vLLM-typical throughput of ~3,000 output tokens/second on the MoE variant, that's ~7.8B output tokens per month if the node runs at 100% utilization.

Practical utilization is rarely 100%. A team running real workloads averages 25-50% utilization (peak hours dense, off-hours sparse). Call it 40% utilization: ~3.1B output tokens per month for $20K = ~$6.45 per million output tokens, all-in (excluding inference engineer time).

API Rate Comparison

For the same SWE-Bench Verified ~75 tier, the relevant proprietary alternatives in mid-2026:

• Claude Sonnet 4.6: $3 input / $15 output per million tokens
• GPT-5.5: $2.50 input / $20 output per million tokens (roughly)
• DeepSeek V4 Pro: $0.20 input / $2 output per million tokens
• Qwen 3.7 Plus: $0.50 input / $4 output per million tokens

Self-hosted Ornith 35B-MoE at ~$6.45 per million output tokens is roughly half the price of Claude Sonnet output, but more than DeepSeek V4 Pro and Qwen 3.7 Plus. The breakeven against Sonnet is around 1.3B output tokens per month per inference node. Below that volume, paying API rates is cheaper. Above it, self-hosting wins.

When Self-Hosting Ornith Makes Sense

The cost math says self-hosting wins above ~1.3B output tokens/month. There are at least three other reasons that change the calculus:

Data residency. If your code must stay within a specific jurisdiction or off third-party APIs entirely, self-hosting is the only path. A self-hosted Ornith node is structurally compliant with the most strict data residency requirements.

Latency control. A dedicated inference node has no queueing, no rate-limit pressure, no "model overloaded" errors. For interactive coding workflows where seconds matter, this is a real productivity gain on top of cost considerations.

Fine-tuning. Self-hosted models can be fine-tuned on your codebase, internal style guides, or domain-specific patterns. API models can't. For teams with strong internal conventions, a fine-tuned Ornith may produce more on-style code than any frontier API model.

The Smaller Variants: 9B and 31B Dense

The 9B-Dense and 31B-Dense Ornith variants are more accessible. The 9B fits on a single 24GB consumer card (RTX 4090, RTX 5090) with quantization. The 31B-Dense needs roughly 2× 40GB GPUs or a single 80GB A100/H100. SWE-Bench Verified scores drop accordingly (the 9B is in the mid-50s tier, the 31B Dense is in the upper-60s tier) but the cost-per-token also drops substantially.

For individual developers or small teams who don't need frontier benchmark performance, the 9B variant on a single RTX 5090 ($2,000-2,500 hardware investment plus electricity) can plausibly handle a large fraction of routine coding tasks at near-zero marginal cost per token. The hardware pays for itself against API rates somewhere in the 200-500M output tokens range, which a heavy individual user crosses in 12-24 months.

The 397B-MoE: Frontier-Tier, Specialized Use Case

The 397B-MoE variant — the one that posts the headline 82.4 / 62.2 / 77 numbers — needs a much larger inference footprint. Practical deployment requires 16-32 H100 GPUs depending on quantization and batch size. Per-node cost: $50K-$100K monthly. Volume break-even against Claude Opus 4.8 ($15 input / $75 output): roughly 800M-1.5B output tokens per month per node, which is a lot but achievable for organizations running hundreds of agents in parallel.

For most teams, the 35B-MoE variant is the right entry point. The 397B-MoE matters mostly because its existence demonstrates that open weights have caught up to the proprietary frontier on coding tasks — which is a structural shift in the market regardless of whether you personally deploy it.

What This Means for Coding Cost Economics

Three structural implications of the Ornith release:

1. Pricing pressure on the proprietary tier. When open weights deliver Sonnet-tier coding performance at self-host cost parity, the pricing floor for "good enough" coding APIs drops. Expect Claude Sonnet, GPT-5.5, and other mid-tier APIs to face increased competitive pressure on their output token pricing over the next two quarters.

2. Hybrid routing becomes more attractive. Teams running OpenRouter or LiteLLM routing can now plausibly route easy tasks to Ornith (cheap), reserve Claude/GPT for hard tasks (expensive), and significantly cut their average cost per task. This is the model orchestration pattern Anthropic and OpenAI hate, but it works.

3. Self-hosted as a real option for mid-sized teams. Previously, self-hosting frontier-tier coding models was an enterprise-only conversation. With Ornith 35B-MoE working on a single 8×H100 node, a 50-developer engineering org can plausibly run a private inference deployment that covers most coding workloads.

Bottom Line

Ornith-1.0 is the strongest open-source coding family released to date. The 35B-MoE variant is the practical choice for teams with non-trivial usage volume and infrastructure tolerance, breaking even against Claude Sonnet at roughly 1.3B output tokens per month. The 9B variant is the practical choice for individual developers willing to invest in a consumer GPU. The 397B-MoE variant is the demonstration that open weights have reached the proprietary frontier on agentic coding — which is the bigger structural news than any single deployment.