Apodex 1.1

Apodex 1.1

Quick answer: Apodex 1.1 is Apodex AI's closed, 397B-parameter flagship agent model, described in the technical report "Apodex 1.1: Scaling Agentic Intelligence for Complex Work." Evaluated under Apodex's Agent Team multi-agent coordination harness, it scores 92.4% (F1) on DeepSearchQA, 77.7% on SWE-bench Verified, 70.8% on Terminal-Bench 2.1, 63.3% on FrontierScience-Research, and 56.1% on Humanity's Last Exam. It's available only via Apodex's API (platform.apodex.ai) — there are no open weights for the flagship, unlike the smaller Apodex 1.1 Mini.

At a Glance

Where Apodex 1.1 leads

• 92.4% F1 on DeepSearchQA and 56.1% on Humanity's Last Exam — strong deep-research and broad-knowledge performance under its Agent Team harness
• 77.7% SWE-bench Verified and 70.8% Terminal-Bench 2.1 — competitive agentic coding/terminal-operation results
• 63.3% FrontierScience-Research and 54.3% FrontierFinance — among the first published frontier-model results on both of these newer, harder professional/science agent benchmarks
• Built around a distinct architectural framing: "Environment Scaling" (executable file/search/code worlds) plus "Agentic Coordination Scaling" (the Agent Team protocol — explicit task board, asynchronous human intervention, asymmetric verification)

Where it lags

• Closed weights, API-only — no self-hosting or fine-tuning option, unlike Apodex 1.1 Mini
• No published pricing as of this writing; access is via Apodex's own platform (platform.apodex.ai) rather than a documented public price list
• Reported scores mix two very different harness settings (minimal ReAct vs. full Agent Team) across benchmarks — the higher Agent Team numbers are not directly comparable to other labs' single-pass or lighter-scaffold results
• Modality, context window, and detailed architecture are not disclosed beyond the technical report's high-level description

Best for: Teams evaluating the ceiling of Apodex's multi-agent "Agent Team" coordination approach on professional-work, research, and coding agent tasks, where API-only access is acceptable.

What Apodex 1.1 Is

Apodex 1.1 is the flagship, closed model in Apodex AI's 1.1 family, described alongside its open-weight Mini sibling in the paper "Apodex 1.1: Scaling Agentic Intelligence for Complex Work" (arXiv:2608.23283). Apodex frames its core contribution as "working capability" — sustained, verifiable progress toward a real-world objective through tool, file, and code interaction, as distinct from static reasoning or knowledge recall. The paper develops this along two axes: Environment Scaling, expanding the diversity and verifiability of executable file/search/code environments used in training, and Agentic Coordination Scaling, training the model to decompose long-horizon tasks, delegate parallel work to sub-agents, integrate asynchronous results, and replan — formalized as the Agent Team protocol (an explicit task board, asymmetric verification between agents, adaptive effort allocation, and evidence-grounded synthesis of sub-agent outputs).

Both the flagship and Mini share a common execution substrate Apodex calls AgentOS — a persistent, three-namespace filesystem (/inputs read-only, /workspace scratch space, /outputs a single-writer-lease publish target) with tiered context compaction and soft/hard wall-clock task budgets, plus a unified training recipe combining supervised fine-tuning with PIVOT-RL (reinforcement learning localized to consequential decision points — "pivots" — in long trajectories, rather than optimizing every step uniformly). Apodex reports nearly every benchmark result under two settings: a minimal single-agent ReAct scaffold, and the full Agent Team multi-agent setup — the latter consistently scoring higher and representing the model's practical ceiling under Apodex's own harness.

At 397B parameters, Apodex 1.1 is presented as achieving its results with a substantially smaller model than many frontier systems it's compared against, a point the paper emphasizes as evidence that its training methodology (rather than raw scale) drives the gains.

Specifications

Pricing

No public price list has been found as of this writing — access is via Apodex's own API platform (platform.apodex.ai). Check Apodex's official documentation for current pricing before production use.

Public Benchmark Scores

Scores are self-reported by Apodex under its Agent Team multi-agent coordination harness (the higher of two settings Apodex reports throughout the paper — a minimal ReAct scaffold produces lower scores on every benchmark, e.g. 69.5% vs 78.8% on the paper's own plain-"GDPval" metric). Not Benchgen measurements — see the source paper for full methodology. SWE-bench Verified and Terminal-Bench 2.1 are reported as single values without a separate ReAct/Agent Team breakdown.

Additional benchmarks reported by Apodex, not added to Benchgen:

• GDPval (78.8% win rate, Agent Team) — Apodex reports this on the original win-rate methodology (Patwardhan et al. 2025); Benchgen's only GDPval-family page, GDPVal-AA v2, uses a structurally different Elo-rating scale (anchored to a human baseline of 1,000) from Artificial Analysis's own reproduction — the two are not the same metric, so this score is not mapped to that page.
• BioMysteryBench (35.3%, Agent Team) — an Anthropic-published bioinformatics evaluation; no independent Benchgen page exists yet.
• IMO 2025, IMO 2026, USAMO 2026, ProofBench (24.3–96.7% across settings) — scored under the MathArena protocol but as distinct competition-year configurations; Benchgen currently tracks other MathArena tracks (ArXivMath, BrokenArXiv, MathArena Apex 2025) but not these specific competition instances.
• FrontierSearchBench, FrontierResearchBench, HDS6 (Heavy-Duty Solver 6) — Apodex's own internal evaluation suites, not independently published or reproducible; treated as prose-only context, not benchmark pages.

Apodex 1.1 vs Alternatives

Apodex 1.1's DeepSearchQA F1 score sits close to but below Kimi K3's reported result, while its SWE-bench Verified and HLE numbers position it competitively among agentic systems evaluated under multi-agent coordination harnesses rather than single-pass scaffolds.

How Apodex 1.1 Performs on Real Agent Tasks

Apodex's central claim is that its Agent Team protocol — explicit task decomposition, parallel sub-agent delegation, asymmetric verification, and evidence-grounded synthesis — produces meaningfully better results than a minimal ReAct scaffold on the same underlying model, across nearly every benchmark in the paper. That's a useful signal for builders: it demonstrates the model's ceiling capability under heavy orchestration, but it also means Apodex 1.1's headline numbers depend as much on the coordination harness as on the base model itself. Teams evaluating Apodex 1.1 for production agentic workloads should weigh whether they can replicate a comparable multi-agent coordination setup, since a simpler single-agent deployment is likely to land closer to the paper's lower ReAct-setting scores.

Related Models & Benchmarks

• Compare: Apodex 1.1 Mini
• Benchmarks: DeepSearchQA · SWE-bench Verified · Terminal-Bench 2.1 · FrontierScience-Research · FrontierFinance · Humanity's Last Exam · APEX-Agents

Frequently Asked Questions

Specs from the Apodex 1.1 technical report (arXiv:2608.23283, Aug 2026). Last updated 2026-08-31.