For years, the frontier AI story has been told as a two-country rivalry — the United States versus China. OpenAI versus DeepSeek. Anthropic versus Qwen. That framing just got more complicated.
Singapore's Agnes AI, the consumer and developer platform built by Sapiens AI, has launched Agnes 2.5 Pro — a model posting benchmark numbers that sit comfortably in frontier territory. And it's available via a free API, today.
The numbers matter. But the geography matters more.
What Agnes 2.5 Pro Is Claiming
Agnes AI's internal evaluations put Agnes 2.5 Pro at 82.7 on SWE-bench Verified — the industry's most widely cited benchmark for real-world software engineering ability. To put that in context, the open-source model ecosystem has been scrambling to cross 60% on the same benchmark. A score above 75% is generally considered strong autonomous coding ability; 82.7 puts Agnes 2.5 Pro in the same bracket as models from Anthropic, OpenAI, and Google.
The full benchmark profile from Agnes AI's announcement includes:
- 82.7% on SWE-bench Verified (real GitHub issue resolution)
- 78.7% multilingual coding performance
- Strong gains on SWE Atlas, with Agnes 2.5 Flash (the smaller sibling released alongside Pro) showing clear improvements over Agnes 2.0 Flash across all seven internal tests
- Benchmark leads over GLM 5.2 and DeepSeek V4 Pro on multiple evaluation cuts
A note on the numbers: Agnes AI describes these as internal evaluations. Independent third-party verification is still pending, and benchmark scores — particularly vendor-reported ones — should always be treated as directional rather than definitive. That said, the company has a credible track record: Agnes AI was confirmed as the first Singapore-founded AI lab to appear in the global top 10 on both Claw-Eval and Artificial Analysis's image-editing leaderboard in May 2026.
Who Is Agnes AI?
Agnes AI is the developer-facing platform of Sapiens AI (formally Singapore Sapiens Technology PTE. LTD.), founded in July 2025 by Bruce Yang — a Raffles Institution alumnus and doctoral candidate in artificial intelligence at the National University of Singapore.
The company's trajectory has been remarkable for its age. By February 2026, it had surpassed 6 million users. It closed a $10 million Series A in February 2026, led by LOOK FORWARD VCC, and was approaching $20 million in annual recurring revenue by March 2026. A potential Singapore Exchange listing is reportedly being targeted by the end of 2026.
The platform operates as a unified AI gateway across text, image, and video — all available under a free API with no usage cap. In the past seven days alone, Agnes AI's models processed 5.41 trillion tokens combined (3.25 trillion text, 2.16 trillion multimodal) under its existing free-access programme. That is not a small operation.
Agnes explicitly positions itself as independent from the technology blocs of the US and China — a deliberate identity choice that is increasingly meaningful as AI becomes entangled with geopolitics. As the global AI investment landscape continues to shift, a credible third-party player outside both blocs carries real strategic weight.
What Was Just Released
The July 2026 release consists of two models:
- Agnes 2.5 Flash — Available now. A fast, lightweight model purpose-built for everyday coding and agentic tasks. It builds directly on Agnes 2.0 Flash (already live with 585,000+ active API users) with clear benchmark improvements across the board. Uncapped free API access at launch.
- Agnes 2.5 Pro — The flagship, targeting production-grade systems, complex agent pipelines, and tasks requiring sustained multi-step reasoning. Free API announced for launch.
Agnes also announced Agnes Code, a desktop coding workspace that integrates the new models directly into real development projects — a direct play in the agentic coding tooling space alongside products like Cursor and Windsurf.
Both models run on the same account, credits, and subscription as the rest of Agnes AI's platform — across chat, API, and Agnes Code. A developer pays once and deploys everywhere.
How It Compares to the Chinese Models It's Beating
The claim that Agnes 2.5 Pro beats GLM 5.2 and DeepSeek V4 Pro on multiple benchmark cuts deserves some context on what it's competing against.
GLM 5.2, from Beijing-based Zhipu AI (Z.AI), is currently one of the strongest open-weight coding models available. It leads SWE-bench Pro at 62.1% and posts an 81.0 on Terminal-Bench 2.1, making it a genuine frontier-class open-source coding model. It is MIT-licensed and Anthropic API-compatible.
DeepSeek V4 Pro is the cost leader — $0.87/M output tokens, 1.6 trillion total parameters, and 93.5% on LiveCodeBench (globally first across all models). Where DeepSeek dominates is in competitive programming and raw algorithmic reasoning. On SWE-bench Verified it scores 80.6%, and its 1M context window and 384K max output are hard to match.
If Agnes 2.5 Pro's 82.7 on SWE-bench Verified holds up under independent evaluation, it would sit above both on that specific benchmark — the most production-relevant measure of real-world coding agent capability. That would represent a genuine result, not just a marginal improvement.
The Broader Shift: Frontier Is No Longer a Two-Country Story
The AI lab landscape through 2025 was organized around a simple mental model: American labs set the frontier (OpenAI, Anthropic, Google DeepMind), Chinese labs challenged on price and open-weight access (DeepSeek, Qwen, GLM, Kimi). Every other country was a consumer, not a producer.
Agnes AI is a real challenge to that model. It is not a wrapper around a US or Chinese base model. It is building in the same environment as the major labs — developing proprietary models, running its own infrastructure, and competing on benchmark leaderboards directly. The fact that it reached the global top 10 on Artificial Analysis as a company less than one year old is not something the existing narrative accounts for easily.
Southeast Asia has long been a significant consumer and deployment market for AI. Singapore's position as a financial, regulatory, and technical hub makes it a natural candidate for hosting a credible independent AI lab. Agnes AI is making that concrete.
The question worth watching is whether independent benchmark validation confirms what Agnes AI's internal numbers show. If it does, the frontier AI story will need a third chapter — and it will start in Singapore.
How to Access Agnes 2.5 Pro
Agnes 2.5 Flash is available now through the Agnes AI developer platform at agnes-ai.com, with API access via an OpenAI-compatible endpoint. Agnes 2.5 Pro follows later this month under the same free-access model. Agnes Code (the desktop agent workspace) is available alongside the Flash release.
The API is free and currently uncapped — the same model Agnes AI has run at scale since its Agnes 2.0 series, which processed trillions of tokens per week under free access.
