---
type: analysis
slug: research-apodex-1-1-is-a-397b-agent-stack-with-an-open-35b-mini-not-a-verified-general
title: "Apodex 1.1 is a 397B agent stack with an open 35B Mini, not a verified general leader"
description: "The August 2026 technical report documents a real environment-and-coordination training recipe and ships Apache-2.0 Mini weights plus an open Agent Team runtime. Independent boards that do not use that runtime do not support a general leading-band claim."
published_at: 2026-10-25T10:00:00.000Z
topics: ["search","agents","research"]
source_urls: ["https://arxiv.org/abs/2608.23283","https://platform.apodex.ai/docs/models","https://platform.apodex.ai/docs/pricing","https://huggingface.co/apodex/Apodex-1.1-mini","https://raw.githubusercontent.com/ApodexAI/FrontierAgent/main/README.md","https://artificialanalysis.ai/models/apodex-1-1","https://artificialanalysis.ai/models/comparisons/gemini-3-8-flash-vs-apodex-1-1","https://www.prnewswire.com/news-releases/apodex-1-1-moves-ai-beyond-deep-research-to-verifiable-execution-302866271.html","https://samaya.ai/benchmarks/frontier-finance","https://www.mercor.com/apex/apex-agents-leaderboard/","https://arxiv.org/abs/2601.14242","https://openai.com/index/gdpval/","https://huggingface.co/collections/apodex/apodex-11","https://www.apodex.ai","https://github.com/ApodexAI/FrontierAgent"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
The newsroom flagged arXiv:2608.23283, Apodex 1.1: Scaling Agentic Intelligence for Complex Work, submitted 24 August 2026 and revised the next day. The report argues that answer quality is not working capability: sustained, verifiable progress through files, search, and code, with state, recovery, and a delivery contract. It trains that capability along two surfaces, Environment Scaling and Agentic Coordination Scaling, bound by a shared harness and AgentOS, then evaluates ReAct against Agent Team. The company also shipped a hosted workbench, an OpenAI-compatible API, Apache-2.0 Mini weights, and FrontierAgent.

This matters for builders because the scores that look frontier-band are Agent Team system scores, while the cheap core API is a no-tool single-pass model. Several comparison rows are Apodex self-reproductions under its own harness. Independent evaluators using different protocols place the 397B model well below current frontier composites, so a news item that repeats the abstract’s leading-band sentence would overclaim.

## What we checked

- **Does Apodex 1.1 reach the leading performance band on professional, finance, and scientific work despite a substantially smaller model than many frontier systems?** partly confirmed. The report’s Agent Team configuration records 54.3 on FrontierFinance and 63.3 on FrontierScience-Research, the highest rows in those vendor tables, and the hosted flagship is documented at 397B parameters. That is not a general lead: Samaya’s official FrontierFinance board updated 28 September 2026 does not list Apodex and tops at 56.0 percent, while Artificial Analysis, without Agent Team, scores Apodex 1.1 at Intelligence Index 26, GDPval-AA Elo 1191, Humanity’s Last Exam 34 percent, and Terminal-Bench 4.0 0 percent. [1][2][9][6][7]
- **Is the 35B Apodex 1.1 Mini locally deployable with public weights and strong working capability?** partly confirmed. Apache-2.0 weights for apodex/Apodex-1.1-mini are on Hugging Face, with SGLang and vLLM serve commands, a 262,144-token context recipe, and a Qwen3.5-style tool-call template; FrontierAgent is the matching open ReAct and Agent Team runtime. Vendor Mini Agent Team scores are 50.2 on FrontierFinance, 51.7 on FrontierScience-Research, and 27.7 on APEX-Agents. No independent public Mini leaderboard was found in this session. [4][13][5][1]

## What we found

- The mechanism in the report is concrete, not a slogan. Environment Scaling builds file, search, and code worlds with explicit verifiers; the file registry is stated as 33 domains, 318 occupations, and 1,208 deliverable clusters. Agentic Coordination Scaling trains decomposition, delegation, staged return, and replanning, then runs them as Agent Team on a persistent task board with asynchronous intervention, asymmetric verification, and adaptive extra effort. Training is unified SFT plus PIVOT-RL over environment trajectories and coordination traces. AgentOS keeps workspace state in named namespaces (/inputs, /workspace, /outputs) outside the transcript. The paper is explicit that evidence-grounded synthesis (section 3.2.5) is used in online products and is excluded from offline evaluations. [1]
- There are two hosted SKUs and two billing contracts. Official model docs list apodex-1.1 at 397B and apodex-1.1-mini at 35B, both with a 262,144-token window. Core LLM API list prices are $0.30 input, $0.03 cached, $3.00 output per million tokens for 1.1, and $0.10 / $0.01 / $1.00 for Mini; requests over 200K input tokens are billed at 2×. Core models are text-only, reasoning-on-by-default, and run no server-side tools. Deep Research 1.1 is $5 / $20, Deep Solve 1.1 $5 / $25, and Deep Discover 1.1 Preview $10 / $100, plus hosted tool fees (web search $5, web fetch $6, code sandbox $90 per 1,000 calls). Paper tables that include search, files, and subagents describe the agent product, not the core completion endpoint. [2][3]
- Vendor Agent Team scores on the 397B stack are 38.5 APEX-Agents, 78.8 GDPval win rate, 54.3 FrontierFinance, 63.3 FrontierScience-Research, 35.3 BioMysteryBench Human-difficult, 56.1 Humanity’s Last Exam, and 92.4 DeepSearchQA F1. ReAct is lower on every overlapping column, and Apodex 1.0 is lower still (16.5 / 59.3 / 40.3 / 28.3). The same report does not claim a coding lead: 70.8 on Terminal-Bench 2.1 and 77.7 on SWE-bench Verified, versus much higher selected comparators. On the internal FrontierResearchBench of 97 executable science workflows, Agent Team pass rate is 12.4 percent against 20.6 percent for GPT-5.6-Sol with Codex. Captions state that all external GDPval rows, the Claude Opus 5 APEX-Agents row, and DeepSeek V4 FrontierFinance rows are Apodex reproductions. [1][4][5]
- Independent measurement of the 397B API does not match the abstract. Artificial Analysis lists Apodex 1.1 as proprietary, released August 2026, Intelligence Index 26 (v4.3.2), GDPval-AA v2.1 Elo 1191, AA-Briefcase 1013, AutomationBench-AA 28 percent, Humanity’s Last Exam 34 percent, SciCode 45 percent, AA-Omniscience −22, and Terminal-Bench 4.0 0 percent. Gemini 3.8 Flash (High) is 41 / 1436 / 48 percent / 20 percent on Index, GDPval-AA, HLE, and Terminal-Bench 4.0 in the same comparison. The 1 September 2026 Apodex press release instead cited an Artificial Analysis Index of 44 on 31 August; that figure is not on the current model page. Artificial Analysis also states image input and undisclosed size, which conflicts with Apodex’s text-only core-API contract and 397B disclosure. [6][7][8][2]
- Official finance and professional-work boards further bound the claim. Samaya’s FrontierFinance page (220 queries, 11,543 rubric items, updated 28 September 2026) ranks Samaya System (high effort) at 56.0 percent and Claude Fable 5.1 with Finance Agent v2 at 55.9 percent; Apodex is absent. Several paper comparators match this board (Claude Fable 5 at 49.2, GPT-5.6-Sol at 46.8), but DeepSeek V4 Pro is 45.5 in the paper’s internal reproduction and 40.5 on the official board. Mercor’s live APEX-Agents leaderboard now reports Mean Score, with Gemini 4 Argon High at 82.4 percent ±4.3 percent on a page that describes 240 tasks and 31 worlds. The original APEX-Agents paper used Pass@1 on 480 tasks across 33 worlds, with Gemini 3 Flash then at 24.0 percent. Apodex’s 38.5 cannot be read onto the current Mercor ranking without an unstated metric and subset mapping. OpenAI’s GDPval gold set is 220  [9][10][11][12][5][1]
- The reproducible builder path is Mini plus FrontierAgent, not the 397B weights. The Hugging Face collection dated through 10 September 2026 lists Apodex-1.1-mini and quantized variants only; the card reports Apache 2.0, 36B parameters, 2,556 downloads in the last month, and qwen3_5_moe. FrontierAgent is Apache 2.0, supports ReAct and Agent Team, maps /inputs, /workspace, and /outputs, and bundles evaluation keys including GDPval, APEX, FrontierScience-Research, and FrontierSearchBench. The 1 September press release (Redwood City) confirms Mini open weights, FrontierAgent, and a hosted workbench, and quotes the team that high-stakes work still needs domain validation and human review. [13][4][5][8][14]

## What it means

Treat Apodex 1.1 as a model-plus-harness system. If you need the paper’s Agent Team numbers, you need Agent Team, the file/search/code tools, and the delivery contract, not a raw /v1/chat/completions call to apodex-1.1. The core API is the cheap, text-only, no-tool path; the Deep Research / Solve / Discover tiers are the expensive agent path, billed on intermediate tokens and hosted tools. Do not paste vendor tables into a competitive bake-off without tagging self-reproduced rows and without matching harness, metric, and date.

For local or auditable work, start from Mini weights and FrontierAgent, log trajectories, and re-run public subsets rather than trusting 78.8 GDPval or 54.3 FrontierFinance as portable facts. Keep a human verifier on load-bearing numbers: the report’s own operational boundary is that AgentOS does not make a source, method, or conclusion correct. Budget for verbosity; Artificial Analysis recorded 232 million output tokens to run its Intelligence Index. If a task is terminal-heavy or one-shot knowledge work under a standard harness, current independent numbers say this stack is not the frontier pick.

## Still open

- No independent public reproduction of Apodex Agent Team scores on FrontierFinance, FrontierScience-Research, APEX-Agents, or GDPval was found.
- Why Artificial Analysis Terminal-Bench 4.0 is 0 percent while the report lists 70.8 on Terminal-Bench 2.1 is unexplained; versions and harnesses differ.
- The press release’s Artificial Analysis Index of 44 (31 August 2026) versus the current Index 26 on v4.3.2 is unresolved; a methodology change is plausible but not documented on the pages opened.
- Whether the 397B flagship shares the Mini’s Qwen3.5 MoE family, and whether image input exists anywhere except the Hugging Face tag, is not settled by API docs versus Artificial Analysis.

## Sources

1. [Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283)](https://arxiv.org/abs/2608.23283) (primary)
2. [Apodex API: core models](https://platform.apodex.ai/docs/models) (primary)
3. [Apodex API pricing](https://platform.apodex.ai/docs/pricing) (primary)
4. [Apodex-1.1-mini model card](https://huggingface.co/apodex/Apodex-1.1-mini) (primary)
5. [FrontierAgent README](https://raw.githubusercontent.com/ApodexAI/FrontierAgent/main/README.md) (primary)
6. [Artificial Analysis: Apodex 1.1](https://artificialanalysis.ai/models/apodex-1-1)
7. [Artificial Analysis: Gemini 3.8 Flash vs Apodex 1.1](https://artificialanalysis.ai/models/comparisons/gemini-3-8-flash-vs-apodex-1-1)
8. [Apodex 1.1 press release (PR Newswire, 1 Sept 2026)](https://www.prnewswire.com/news-releases/apodex-1-1-moves-ai-beyond-deep-research-to-verifiable-execution-302866271.html) (primary)
9. [Samaya FrontierFinance benchmark and leaderboard](https://samaya.ai/benchmarks/frontier-finance)
10. [Mercor APEX-Agents leaderboard](https://www.mercor.com/apex/apex-agents-leaderboard/)
11. [APEX-Agents (Vidgen et al., arXiv:2601.14242)](https://arxiv.org/abs/2601.14242)
12. [OpenAI: Measuring performance on real-world tasks (GDPval)](https://openai.com/index/gdpval/)
13. [Hugging Face collection: Apodex-1.1](https://huggingface.co/collections/apodex/apodex-11) (primary)
14. [Apodex official website](https://www.apodex.ai) (primary)
15. [ApodexAI/FrontierAgent GitHub repository](https://github.com/ApodexAI/FrontierAgent) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Search](https://newruntime.com/topics/search.md): Explore the search topic hub.
- topic hub: [Agents](https://newruntime.com/topics/agents.md): Explore the agents topic hub.
- topic hub: [Research](https://newruntime.com/topics/research.md): Explore the research topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)