Apodex 1.1 Moves AI Beyond Deep Research to Verifiable Execution

News provided by

Sep 01, 2026, 09:00 ET

Apodex 1.1 brings reasoning, multi-agent collaboration and built-in verification into a single platform, with a new open-weight model and open-source framework for local deployment

REDWOOD CITY, Calif., Sept. 1, 2026 /PRNewswire/ -- Apodex announced the release of Apodex 1.1, a reasoning model and online workbench that can carry complex, multi-step professional, scientific, and financial work from raw inputs through to verifiable deliverables.

Within a persistent workspace, the model analyzes papers, datasets, spreadsheets, images and code, adapts as evidence changes, recovers from setbacks, and traces every consequential conclusion back to the sources, calculations and files behind it, while also maintaining strong general reasoning, deep search, mathematics and coding capabilities.

On August 31, Artificial Analysis, a leading independent benchmarking company for AI, scored Apodex 1.1 at 44 on its Intelligence Index, roughly equivalent with DeepSeek V4 Pro 0424 and Kimi K2.6 — models built by foundation-model labs at roughly a trillion parameters. Artificial Analysis applies the same standardized harness to every model it scores, without the Agent Team coordination and verification layer Apodex 1.1 uses in the workbench. Apodex 1.1 reaches that level through post-training rather than scale — the same approach that produces Apodex 1.1 Mini, released today with open weights at 35 billion parameters.

Apodex 1.1 is available in the Apodex online workbench. Apodex has also released open weights for the 35-billion-parameter Apodex 1.1 Mini and open-sourced FrontierAgent, its execution framework for ReAct and Agent Team workflows.

Model Performance

Apodex 1.1 has been evaluated across professional work, financial analysis, scientific research, general reasoning, and deep search. Results are from Apodex's published comparison, with comparator scores shown alongside.

Complex professional work: 38.5 on APEX-Agents (GPT-5.5 38.5, GPT-5.6 Terra 38.9, GPT-5.6 Sol 39.9). 78.8 on GDPVal, which covers real deliverables from 44 occupations (GPT-5.6 Sol 79.3, Kimi K3 max 80.0).
Financial analysis: 54.3 on FrontierFinance, the highest result in the comparison (Claude Fable 5 49.2, Kimi K3 max 48.8, GPT-5.6 Sol 46.8).
Scientific research: 63.3 on FrontierScience-Research, the highest result in the comparison (DeepSeek V4 Flash 0731 55.0). 35.3 on the harder BioMysteryBench Human-difficult Set, between Claude 4.x and Claude Opus 5.
General reasoning and deep search: 92.4 on DeepSearchQA (Kimi K2.6 92.5, Claude Opus 4.7 91.7).

Technical Approach

Apodex 1.1's core advance goes beyond adding more tools around a model. The model's working capability incorporates tool use, task execution, failure recovery, and multi-agent coordination. Environment Scaling and Agentic Coordination Scaling are connected by a unified execution framework, AgentOS and the training system.

Environment Scaling treats executable environments as a scaling dimension. File, Search and Code Worlds expose the model to real files, source discovery, code execution, changing states, failures and verification rather than additional prompts. Apodex's file-task library spans 33 professional domains, 318 occupations and 1,208 deliverable types. Agentic Coordination Scaling trains the model to break down objectives, delegate parallel work, integrate findings, replan as evidence changes and stop branches that no longer add value. AgentOS holds files, evidence, execution logs, dependencies and task status independently of the conversation, so long-running work survives compression, redirection and recovery after failure.

Verification runs alongside execution. Statement Review and asymmetric verification test important claims against evidence, citations, calculations and delivery requirements, and an integrity gate zeroes any trajectory in which the system fabricates a tool result or claims an action that did not occur.

"Most evaluation still asks whether the final answer was right. On a task that runs for hours, that tells you very little because a successful run can contain badly reasoned steps, and a failed run can appear to be working until it doesn't," said Xinyu Wang, AI Research Scientist, Apodex. "Our TRACES methodology scores the process before publishing results against it so we can track the quality of the answers."

Open Weights and Local Deployment

Apodex 1.1 Mini scores 50.2 on FrontierFinance and 51.7 on FrontierScience-Research. When paired with FrontierAgent, which runs from a single command on macOS and Linux with no pre-built environment and no Docker requirement, Apodex 1.1 Mini makes professional workflows that previously depended on cloud frontier models. Because the weights are public, these results can be reproduced independently.

"An Agent Team is a behavior the model has, not infrastructure sitting on top of it — which is why the whole thing runs from one command on a laptop," said Simon Du, Lead Scientist at Apodex. "The coordination in FrontierAgent is the same coordination we run ourselves — the best version there is."

Apodex notes that verifiable execution does not make every source, method or conclusion correct. High-stakes work still requires domain-specific validation, reproducible analysis and human review. Pretraining for Apodex 2.0 is underway, moving persistent reasoning, native tool execution, self-correction and end-to-end verifiability into the base model rather than adding them through post-training and orchestration.

About Apodex

Apodex builds heavy-duty solvers: AI systems designed to carry complex, long-running professional and scientific work from raw input to finished deliverable, through a process that can be inspected and verified at every step. The company releases both hosted systems and open-weight models with open execution frameworks. Apodex is hiring across research, engineering and open-source contributions. Learn more at www.apodex.ai.

For More Information