Agnes Offers Free Unlimited Access While Codex Faces Bans and Claude Demands ID Verification

Agnes has launched a new generation of large‑language models for coding, introducing Agnes‑2.5‑Flash, a high‑performance text model optimized for coding, agent, and everyday development scenarios. Compared with the previous generation, Flash shows clear improvements in code understanding, engineering repair, multi‑step task execution, and complex reasoning, positioning it as the main model for daily developer use. Notably, the model is offered with an unlimited‑free‑access policy.

The company also announced the upcoming flagship model Agnes‑2.5‑Pro, aimed at complex agents, professional development, and high‑difficulty engineering tasks. Pro is described as more capable in long‑chain reasoning, large‑code‑base manipulation, multi‑file modifications, and system‑level analysis, and is expected to be a paid offering that will soon be released.

Alongside the models, Agnes released the AgnesCode desktop client, an AI‑powered coding workspace that integrates the model, skills, applications, local projects, and generated artifacts. Users interact via a single input box, issuing natural‑language commands that the system translates into code, file operations, tool invocations, and result retrieval, effectively turning the AI from a chat interface into a full‑featured development assistant.

To evaluate real‑world performance, the authors conducted hands‑on tests. First, they asked the model to generate classic games such as Snake and Space Invaders, which produced playable results. They then increased difficulty by having the system recreate a simplified Super Mario, demonstrating strong multi‑step execution and external‑tool integration via the "Coding Harness" feature.

Using the integrated data service from Machine Heart, the authors queried recent Loop Engineering articles. AgnesCode performed a six‑minute run, invoking tools multiple times and returning a detailed report that not only matched keywords but also understood the underlying intent, showing superior semantic retrieval.

The model also generated a knowledge‑graph of Ming‑dynasty literary figures, illustrating its capability in knowledge‑graph construction.

For office‑task evaluation, the authors supplied a link to ACL best‑paper listings and asked AgnesCode to produce a colorful PPT with at least three slides per paper. The resulting presentation covered content selection, layout, and visual design, requiring only a single prompt.

Overall, Agnes‑2.5‑Flash and AgnesCode produced high‑quality outputs across coding, data retrieval, knowledge organization, and document generation, with minor coding imperfections that can be corrected easily. The authors note that these results are achieved with the free Flash model, while the yet‑unreleased Pro model already demonstrates impressive demos such as a complete particle game, a 3D room, and a sophisticated typhoon tracker.

In the broader industry context, recent OpenAI incidents saw Codex accounts abruptly banned, affecting 5‑10% of subscribers, and Anthropic introduced mandatory KYC verification for Claude, which excludes mainland Chinese IDs. These developments raise the barrier for Chinese developers seeking top‑tier AI coding tools, highlighting the significance of Agnes’s free, stable, and domestically accessible offering.

Agnes concludes by summarizing its three‑part offering: Agnes‑2.5‑Flash for low‑cost strong coding ability, the upcoming Agnes‑2.5‑Pro for flagship experiences, and the AgnesCode desktop that links model, harness agent, and local projects into a seamless workflow.