Arcee, a U.S. AI startup with fewer than 100 employees, has open-sourced Trinity-Large-Thinking, a reasoning model that challenges Anthropic's flagship Opus 4.6 on agent benchmarks — at a fraction of the cost.
Agent Benchmarks: Close to Opus, First in One
On PinchBench, which measures real-world agent workflow performance, Trinity-Large-Thinking scored 91.9, just 1.4 points behind Opus 4.6's 93.3. In the Tau2-Airline benchmark simulating customer service interactions, it achieved 88.0, the highest among all compared models. The results underscore the model's strength in multi-turn dialogue and tool use.
Pricing is the knockout punch: API output costs $0.90 per million tokens, roughly 96% cheaper than Opus 4.6. For applications running long agent loops that burn tokens relentlessly, the cost gap may matter more than the score gap.
400B MoE: 13B Compute Per Token
The secret lies in architecture. Trinity-Large-Thinking uses a sparse mixture-of-experts design with 256 experts, activating only 4 per token. The massive 400B-parameter model needs just 13B compute per inference, 2–3 times more efficient than dense models of similar scale.
The key upgrade from the Preview version (released in late January) is the added reasoning chain. The Thinking version “thinks” before answering, boosting stability in multi-step tool calls and long-context coherence. Arcee says the model is built not to break in lengthy agent loops.
Base model training took 33 days and cost $20 million; post-training for the Thinking version required another 9 months of refinement.
CEO Lucas Atkins wrote in the announcement: “Getting here took difficult technical work, hard calls…Nobody did that. They kept pushing.”
General Reasoning Lags; Targets Chinese Open-Source Rivals
The focus on agents comes with trade-offs. On general reasoning benchmarks, Trinity-Large-Thinking scores 76.3 on GPQA-D (vs Kimi K2.5's 86.9 and Opus 4.6's 89.2) and 83.4 on MMLU-Pro, the lowest among comparables.
Arcee isn't chasing that race. The company positions Trinity-Large-Thinking as “the strongest open-source model outside China in many dimensions,” aiming squarely at Chinese open-source contenders like DeepSeek and Kimi, not at Opus or GPT.
The model is live on OpenRouter with a 5-day free trial on OpenClaw. Its predecessor, Preview, has processed over 3.37 trillion tokens on OpenRouter since late January, ranking first in the U.S. and fourth globally among open-source models on OpenClaw. For a small startup, that adoption proves the market's appetite for cheap, capable models.
Weights are available on Hugging Face under the Apache 2.0 license, free for download, modification, and commercial deployment.