Agnes 2.5 Pro Beta: Singapore’s AI at a Fraction of the Price

Singapore’s Sapiens AI released Agnes 2.5 Pro Beta on August 26, 2026, a reasoning model that scores 49 on the Artificial Analysis Intelligence Index and is priced at $0.10 per million input tokens. That puts the Singapore-built model in the same conversation as frontier releases from Meta, DeepSeek, and Anthropic, while undercutting most of them on cost.

The launch is the latest signal that Singapore is producing competitive homegrown AI rather than just consuming it. Sapiens AI has raised $20 million, is approaching $20 million in annual recurring revenue, and is shipping a model family that includes the first paid tier cheaper than its own predecessor.

What Is Agnes 2.5 Pro Beta?

Agnes 2.5 Pro Beta is a paid reasoning model from Agnes AI, the commercial product line of Singapore-based startup Sapiens AI. It is designed for production workloads that need stronger reasoning depth than flash-class models, including complex code generation, scientific reasoning, long-context document analysis, and agentic workflows.

On the surface, the model looks similar to most modern frontier releases. It supports text and image input, generates text output, and uses OpenAI-compatible API endpoints, so existing code can usually be ported by swapping the base URL. What stands out is the combination of price, context, and benchmark score.

Key Specifications at a Glance

• Release date: August 26, 2026
• Developer: Sapiens AI (Singapore), operating as Agnes AI
• Model type: Proprietary, paid reasoning model
• Context window: 1 million tokens (roughly 1,500 A4 pages)
• Maximum output: 65,536 tokens per request
• Input modalities: Text and image URL
• Output modality: Text
• Intelligence Index: 49 (Artificial Analysis v4.1.1)
• Output speed: 159.5 tokens per second
• Cache discount: 90% on cached input
• Endpoints: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages

The detailed benchmark breakdown published by Artificial Analysis shows strong performance in agentic coding and scientific reasoning, with 70% on Terminal-Bench v2.1 and 91% on GPQA Diamond. Weaker areas include 16% on CritPt (physics reasoning) and 17% on AA-Omniscience Accuracy, which the firm notes is well above the class median but still indicates room to grow on long-tail factual knowledge.

One practical caveat: Artificial Analysis flags Agnes 2.5 Pro Beta as “very verbose.” It generates around 150 million tokens to run the full Intelligence Index, more than double the median of 66 million. For applications where output cost matters, that verbosity is worth budgeting for.

The Agnes 2.5 Pro Family: Alpha, Stable, and Beta Compared

Agnes 2.5 Pro is not a single model but a family of three related releases. Each one targets a slightly different stage in the development pipeline, and the pricing differences between them are sharper than the model names suggest.

Model Lineup

• Agnes 2.5 Pro Alpha (July 24, 2026): The open-weights research release. Weights are published on Hugging Face under Apache 2.0, making it usable for self-hosting and commercial deployment. Priced at $0.45 input / $0.90 output per 1M tokens through the hosted API.
• Agnes 2.5 Pro (August 1, 2026): The commercial stable release. Same benchmark numbers as Alpha, but weights are proprietary. Identical pricing to Alpha at $0.45 / $0.90 per 1M tokens.
• Agnes 2.5 Pro Beta (August 26, 2026): The newest release and the most interesting on price. It jumps to an Intelligence Index of 49 (up from 39) while cutting input price by roughly 78% and output price by 67%.

Side-by-Side Specifications

The Beta release does not replace Alpha or the stable commercial model. All three remain available through the same API hub, and the documentation pages for Agnes 2.5 Pro Beta, Agnes 2.5 Pro, and Agnes 2.5 Pro Alpha explicitly cross-link to one another.

Benchmark Performance vs. the Global Frontier

Agnes 2.5 Pro Beta sits in an unusual bracket. It is priced like an entry-level model but benchmarks like a mid-tier frontier system. The Intelligence Index of 49 places it at #6 out of 174 models in its class, and roughly 2.7 times higher than the median of 18 for reasoning models in its price tier.

To put that in context, here is how it compares to the other models mentioned in the question: Muse Spark 1.2, DeepSeek V4 Flash, GPT-5.6 Luna, MiniMax-M3, MiMo V2.5 Pro, Claude Sonnet 5, and Claude Opus 4.6, all scored under the same benchmark methodology.

Intelligence Index Comparison

The numbers tell a sharper story once you line up the direct competitors. Among reasoning models in the sub-$0.30 per 1M output band, Agnes 2.5 Pro Beta sits at the top of the index at 49, ahead of the open-weights MiMo V2.5 Pro (43) and MiniMax-M3 (45) at lower input prices than either. Against the mid-tier reasoning models that score 50 to 55, it trades 3 to 6 points of intelligence for a 2x to 10x reduction in cost.

Versus GPT-5.6 Luna (max) at 52, Agnes 2.5 Pro Beta is 3 points behind on intelligence but half the input price and a quarter of the output price. Against Claude Sonnet 5 (max) at 55, the gap widens to 6 points and the price gap stretches to 20x on input and 33x on output. Against the open-weight MiMo V2.5 Pro at 43, Agnes is 6 points ahead at less than a quarter of the input cost. And at 160 tokens per second, it is the fastest model in this comparison group, which matters more than the index number suggests for interactive applications.

Among the models that score in the 40s, Agnes 2.5 Pro Beta is the cheapest by a wide margin and the only one that pushes output above 150 tokens per second. Among the models that score in the low 50s, it is meaningfully cheaper on both input and output, and the only one that is faster than 130 tokens per second.

Agnes 2.5 Pro Beta lands 5 points below DeepSeek V4 Flash on intelligence but charges less than half the input price and a third of the output price. This is part of a broader price collapse reshaping the commercial AI market, with similar undercutting patterns coming from Chinese labs like DeepSeek and Zhipu. As covered in our analysis of why Chinese AI models are so much cheaper than OpenAI and Anthropic, Singapore-built Agnes AI is now joining that same price-pressure wave from a different angle. It is 5 points below Muse Spark 1.2 at roughly one-tenth the input cost. The performance gap is real, but so is the price gap.

Among open-weight competitors specifically, Agnes 2.5 Pro Beta lands 4 points above MiniMax-M3 at one-third the input cost and one-quarter the output cost, and 6 points above MiMo V2.5 Pro at less than a quarter of the input cost. Against the Anthropic and OpenAI mid-tier flagships, the trade is more intelligence for much more money: Claude Sonnet 5 (max) is 6 points ahead at 20x the input cost, and Claude Opus 4.6 (max) is 4 points below at 50x the input cost. At 159.5 tokens per second, Agnes 2.5 Pro Beta is unusually fast for a reasoning model that returns visible chain-of-thought to the caller.

Where Agnes 2.5 Pro Beta Stands Out

• GPQA Diamond (graduate-level science Q&A): 91%
• AA-LCR (long-context reasoning): 78%
• Terminal-Bench v2.1 (agentic coding and terminal use): 70%
• AA-Omniscience non-hallucination rate: 67%

Where It Trails the Frontier

• CritPt (physics reasoning): 16%
• AA-Omniscience accuracy: 17%
• τ3-Banking (agentic tool use in finance): 36%
• GDPval-AA v2 Elo: 1,456 (48%)

For most developer and analyst workloads, the strengths line up with where the model is being marketed. For specialized scientific work in physics or for finance-specific agentic tasks, the gap is more noticeable.

Sapiens AI: Singapore’s Homegrown Answer to DeepSeek

Agnes AI is the public brand of Sapiens AI, a Singapore-headquartered startup founded by Bruce Yang, a name that has appeared on Forbes Asia’s 30 Under 30 list. The company raised a $10 million Series A on February 13, 2026, led by LOOK FORWARD VCC, taking total funding to $20 million.

Sapiens AI has been explicit about its positioning. The Straits Times headline from the May 2026 launch describes Agnes as “Singapore’s homegrown answer to DeepSeek,” and the company’s strategic focus is consumer-facing AI for emerging markets including Indonesia, the Philippines, Brazil, and the Middle East.

Singapore itself has been pushing AI as a national priority. The country launched its National AI Strategy 2.0 in late 2024, and in 2025 released a Model AI Governance Framework for Generative AI through IMDA, the Infocomm Media Development Authority. AI Singapore, the government-backed research program, has also been a long-standing pillar of the country’s AI ecosystem. Agnes AI is one of the first Singapore-headquartered labs to produce a model that benchmarks competitively with frontier Chinese and US releases.

The financial picture is unusual for a sub-frontier model. Sapiens AI reported approaching $20 million in annual recurring revenue as of March 2026, just four months after the Series A. That is far faster monetization than most open-weight Chinese labs have achieved, and it reflects the company’s focus on paid API access plus a viral free tier.

Pricing Breakdown and the Free API Tier

Agnes 2.5 Pro Beta’s headline price is the part most coverage focuses on, but the full pricing structure has more layers worth understanding.

Paid Tier: Agnes 2.5 Pro Beta

• Input (cache miss): $0.10 per 1M tokens
• Output: $0.30 per 1M tokens
• Input (cache hit / cache read): $0.01 per 1M tokens

The 90% cache discount is the key number for production workloads. Most agent loops re-send the same system prompt and tool definitions on every call, and a cache hit price of $0.01 per million tokens means the marginal cost of repeated context is effectively zero.

Paid Tier Comparison: Alpha vs Stable vs Beta

The Beta tier is 4.5x cheaper on input, 3x cheaper on output, and 4.5x cheaper on cached input than its stable sibling. For a developer running a multi-turn agent loop with heavy prompt caching, that translates to a meaningful change in unit economics.

Free Tier: What Is Included and What Are the Limits?

Agnes AI runs one of the most generous free tiers in the commercial AI market. The core rules, as reported in industry coverage and confirmed through the company’s own platform communications:

• Free models: agnes-2.0-flash and agnes-2.5-flash for text and image input. Output is free for these models with no announced end date.
• No credit card required to sign up.
• No time limit on free access. The company has stated access to its own models is “indefinitely free.”
• Single omni-modal API covers text, image, and video through the same account.
• Image and video generation: agnes-image-2.0-flash and agnes-image-2.1-flash are also free across all output-resolution tiers and input reference images. Our hands-on guide to Agnes Image 2.5 Flash covers the newest image model in the line, which shares the same free pricing.

The catch is rate limits, not pricing. Free-tier users are subject to requests-per-minute limits, which can produce errors under burst load. The company recommends exponential backoff for retries.

There is also a data usage flag worth knowing about. Some industry observers have noted that the free tier trains on user data, so production traffic should be tested on the paid tier before being committed to a customer-facing feature. This is similar to how several open-weight providers handle their no-cost tiers, but it is a non-trivial consideration for enterprise use.

Agnes 2.5 Pro Beta itself is not free. The free tier is limited to the Flash models. Beta is a paid reasoning model with no public free quota.

How Much Does Agnes 2.5 Pro Beta Cost in Practice?

A quick back-of-the-envelope example: a 1,000-token system prompt sent 1 million times with cache hits, generating 500-token answers each time, would cost roughly:

• Input cost (with cache): 1,000 tokens x 1M calls x $0.01 / 1M = $10
• Output cost: 500 tokens x 1M calls x $0.30 / 1M = $150
• Total: $160 for one million completions

For the same workload on the stable Pro model at full price ($0.45 input, $0.90 output, $0.045 cache), the bill would be $900, more than five times higher.

How to Access Agnes 2.5 Pro Beta: Developer Quickstart

The API is OpenAI-compatible, which makes integration a single base-URL swap for most existing code. The same hub also supports the Anthropic Messages API and the OpenAI Responses API.

Base Configuration

• Base URL: https://apihub.agnes-ai.com/v1
• Chat Completions endpoint: POST /v1/chat/completions
• Responses endpoint: POST /v1/responses
• Messages endpoint: POST /v1/messages (Anthropic-compatible)
• Model name: agnes-2.5-pro-beta
• Authentication: Bearer token (or x-api-key for the Messages API)

Basic Chat Request

This curl example sends a technical question and asks for a 1,200-token answer.

curl https://apihub.agnes-ai.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agnes-2.5-pro-beta",
    "messages": [
      {"role": "system", "content": "You are a precise technical assistant."},
      {"role": "user", "content": "Explain the tradeoffs between optimistic locking and pessimistic locking in distributed systems."}
    ],
    "temperature": 0.3,
    "max_tokens": 1200
  }'

Streaming Request

Setting stream: true returns tokens as they are generated. For a reasoning model, the first chunk will be delayed while the model finishes its thinking pass.

curl https://apihub.agnes-ai.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agnes-2.5-pro-beta",
    "messages": [
      {"role": "user", "content": "Create a step-by-step migration plan for moving a monolith to services."}
    ],
    "stream": true
  }'

Image Input Request

Agnes 2.5 Pro Beta accepts image URLs alongside text, which is useful for visual reasoning and chart analysis.

{
  "model": "agnes-2.5-pro-beta",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "Summarize this chart and call out any anomalies."},
        {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}}
      ]
    }
  ]
}

Integration Tips From the Official Documentation

• Use a large max_tokens value. Reasoning models can consume part of the output budget before producing a visible answer. If you get finish_reason: “length” or stop_reason: “max_tokens”, raise the limit and retry.
• For long-context tasks, structure the prompt with section labels and filenames so the model can refer back to specific sources.
• For debugging, ask the model for root-cause analysis before requesting a patch. The Beta tier is strong on multi-step reasoning.
• Track cache read, input, and output token usage separately. The cache read is 10% of input cost, so prompt-cache strategies can drop effective spend dramatically.

The official Agnes 2.5 Pro Beta documentation includes a full integration checklist and additional request examples for coding, streaming, and image understanding tasks.

Frequently Asked Questions

Is Agnes 2.5 Pro Beta free to use?

No. Agnes 2.5 Pro Beta is a paid model. Pricing is $0.10 per 1M input tokens, $0.30 per 1M output tokens, and $0.01 per 1M cached-input tokens. The free tier covers the Flash-class models (Agnes 2.0 Flash and Agnes 2.5 Flash), not the Pro Beta tier.

How does Agnes 2.5 Pro Beta compare to MiniMax-M3 and MiMo V2.5 Pro?

Both are competitive open-weight models at similar price points. MiniMax-M3 scores 45 on the Artificial Analysis Intelligence Index at $0.30 input and $1.20 output per 1M tokens, while MiMo V2.5 Pro scores 43 at $0.435 and $0.87. Agnes 2.5 Pro Beta is 4 to 6 points ahead of either, at one-third to one-quarter the input cost. It is also the only one of the three that supports image input natively through the hosted API.

How does Agnes 2.5 Pro Beta compare to GPT-5.6 Luna?

GPT-5.6 Luna (max) scores 52 on the Intelligence Index, three points above Agnes 2.5 Pro Beta, and is priced at $0.20 input and $1.20 output per 1M tokens. The price gap favors Agnes 2.5 Pro Beta on both axes, with 50% lower input cost and 75% lower output cost. GPT-5.6 Luna is only available through OpenAI, while Agnes 2.5 Pro Beta is available through the OpenAI-compatible, Anthropic-compatible, and OpenAI Responses APIs.

How does Agnes 2.5 Pro Beta compare to DeepSeek V4 Flash?

On the Artificial Analysis Intelligence Index, DeepSeek V4 Flash scores 52 versus Agnes 2.5 Pro Beta at 49, a gap of 3 points. On price, Agnes 2.5 Pro Beta is roughly half the input cost and less than half the output cost. For workloads where the 3-point intelligence gap matters more than cost, DeepSeek V4 Flash is the better fit. For cost-sensitive applications, Agnes 2.5 Pro Beta delivers comparable reasoning depth at significantly lower spend.

Can I self-host Agnes 2.5 Pro Beta?

Not directly. The Beta model is proprietary and only available through the hosted Agnes AI API. If self-hosting is required, the open-weights Agnes 2.5 Pro Alpha is available on Hugging Face under Apache 2.0, though with a lower Intelligence Index of 39.

What is the difference between Agnes 2.5 Pro Beta and the Alpha version?

Alpha (July 24, 2026) is the open-weights research model scoring 39 on the Intelligence Index and priced at $0.45 / $0.90 per 1M tokens. Beta (August 26, 2026) is the newer, proprietary release that scores 49 on the Intelligence Index and is priced at $0.10 / $0.30 per 1M tokens. Beta is 26% smarter by index and 4.5x cheaper on input.

What API endpoints does Agnes 2.5 Pro Beta support?

Three: the OpenAI-compatible Chat Completions API, the OpenAI Responses API, and the Anthropic Messages API. All three are served from the same base URL, so switching endpoints does not require re-authentication. Existing OpenAI or Anthropic client code generally works after changing the base URL and model name.

Is the Agnes AI free tier really unlimited?

The company states access to its own Flash-class models is free with no time limit and no credit card required. The practical limits are requests-per-minute rate limits and burst-error behavior under heavy load, which the company says should be handled with exponential backoff. The free tier also trains on user data, which is worth verifying before routing production traffic.

Who built Agnes AI?

Agnes AI is the commercial product line of Sapiens AI, a Singapore-based startup founded by Bruce Yang. The company raised a $10M Series A in February 2026 led by LOOK FORWARD VCC and was approaching $20M ARR as of March 2026.

The Bottom Line

Agnes 2.5 Pro Beta is the kind of release that quietly shifts the floor of what “cheap reasoning” means. At $0.10 per 1M input tokens, with 90% cache discounts and an Intelligence Index of 49, it undercuts most of the OpenAI-and-Anthropic-priced tier without falling off the benchmark curve.

It is not a frontier model. Claude Opus 5, GPT-5.6 Sol, and Kimi K3 still lead on raw intelligence, and the Beta tier’s verbosity is a real cost consideration. But for teams that need solid agentic and scientific reasoning at sustained production scale, and that want to support a non-US, non-Chinese provider, Agnes 2.5 Pro Beta is a serious option.

For Singapore, the launch is a proof point that the country’s National AI Strategy is producing exportable, competitive models. For developers, it is one more reminder that the price of intelligence in 2026 is no longer set in San Francisco alone.