Agnes 2.5 Pro Beta Brings 1M-Token Reasoning at Lower API Prices

Agnes AI releases a new paid reasoning model

Singapore-based Agnes AI has made Agnes 2.5 Pro Beta available through its API as a new proprietary reasoning model for advanced coding, long-context analysis, agent workflows and multimodal understanding.

The model is released and callable now, according to Agnes AI's official documentation. It should not be confused with Agnes 2.5 Pro Alpha, an earlier open-weight Apache 2.0 research model. The new Beta is a separate proprietary hosted model.

One model, three common API protocols

Agnes documents three supported interfaces for agnes-2.5-pro-beta:

• OpenAI-compatible Chat Completions;
• OpenAI-compatible Responses API; and
• an Anthropic Messages-compatible endpoint.

That multi-protocol support is aimed at reducing integration friction for developers moving existing applications or agents between providers. Applications still need testing for behavioral differences; API-shape compatibility does not imply identical model semantics.

The Beta accepts text and image input and produces text output. It supports tool definitions in agent workflows and streaming responses.

1M-token context and 65,536-token maximum output

Agnes lists a 1 million token context window and a maximum output length of 65,536 tokens.

Those limits make the model relevant to repository-scale coding, long document analysis and agent workflows that retain substantial working context. Actual useful context can be lower than the nominal maximum depending on retrieval strategy, latency, prompt structure and model performance over very long inputs.

Pricing is the main competitive signal

Agnes lists the current hosted price as:

• $0.10 per million uncached input tokens;
• $0.01 per million cached input tokens; and
• $0.30 per million output tokens.

The company notes that pricing and availability can depend on account, region, billing configuration and later updates, so developers should verify current console pricing before committing production workloads.

This pricing is materially lower than many frontier reasoning APIs, but cost alone does not determine production value. Developers should compare accuracy, latency, tool reliability, safety behavior, rate limits and total token consumption on their own workloads.

Beta is proprietary, not an open-weight update

The release-status distinction is important because the Agnes 2.5 family contains both open and closed variants.

Agnes 2.5 Pro Alpha is documented as an open-source multimodal reasoning model whose official weights are available on Hugging Face under Apache 2.0.

Agnes 2.5 Pro Beta, by contrast, is documented as a paid proprietary model with no open weights.

That makes the Beta primarily a hosted-service release rather than a self-hosting release.

Benchmark references need independent interpretation

Agnes says the Beta has participated in Artificial Analysis intelligence and agent-capability evaluation. Third-party coverage reports a higher Artificial Analysis Intelligence Index result than earlier Agnes 2.5 Pro variants, but benchmark scores can change as evaluation suites and provider implementations evolve.

The safest way to evaluate the release is to treat external leaderboard results as one signal and run application-specific tests for coding, multimodal reasoning, long-context retrieval and tool execution before migration.

Why this release matters

The main significance is the combination of long context, multimodal input, multiple API compatibility layers and aggressive hosted pricing from a smaller non-US, non-Chinese model provider.

That broadens provider choice for developers who want to reduce inference cost or avoid tightly coupling an application to one API ecosystem.

Released now: Agnes 2.5 Pro Beta through Agnes AI's hosted API.

Not released as open weights: the Beta remains proprietary. Developers seeking a self-hostable Agnes model should look separately at the Alpha release and verify its capabilities against their use case.