Benchmark Almanac

Thinkingbox-bench

Claude Opus 5 66.5% leads since August 2026
Not in model cards

Agents complete 507 business workflows through tools and simulated user conversations. Executable checks require correct backend updates without forbidden side effects.

Some tasks also grade final responses against rubrics. The paper reports repeated-attempt reliability, but the linked dataset repository returned 404 during this audit.

Made by
Microsoft and collaborators
Released
August 2026
Kind
Customer service
Level
University
Scored
Checked automatically
Answers
Agent task
Measures
Capability
Language
English
In and out
Text → Text, Actions
Interaction
Agentic
Environment
APIs and tools
Items from
Generated
Refresh
One fixed set
Run by
A host
In views
Agents
Strongest model tested
GPT-6 Astra #4 on the capability index
Last model added
September 2026
Scores refreshed
November 20th, 2026

In the authors’ words

Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state. Built on this sandbox, Thinkingbox-bench contains 507 policy-conditioned workflows across business scenarios, including retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support. Each attempt is evaluated by task-specific executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects; designated tasks additionally check required properties of the final response. Our experiments reveal that even the strongest proprietary and open-weight models show steep reliability drops: Claude Opus 5 falls from 66.50% pass@1 to 47.53% pass^20, and Kimi-K3 from 57.37% pass@1 to 17.60% pass^20. Moreover, many failed trials terminate cleanly after valid state-changing actions, so response- or tool-call-level signals poorly proxy end-to-end completion. Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks. We release Thinkingbox, Thinkingbox-Bench, and Thinkingbox-integrated reinforcement learning framework.

Abstract of One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741, CC0)
18
Models scored in its paper

Claude Opus 5 leads with 66.5% among models released by September 2026

Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

Source: The Thinkingbox-bench paper

#ModelScore ±SettingDateSource
1 Claude Opus 5 66.5% ±3.7 507 tasks; 20 trials; medium, Thinkingbox Jul 2026 Official leaderboard
2 GPT-5.4 65.4% ±3.1 507 tasks; 20 trials; medium, Thinkingbox Mar 2026 Official leaderboard
3 GPT-5.6 Sol 61.9% ±3.3 507 tasks; 20 trials; medium, Thinkingbox Jul 2026 Official leaderboard
4 Claude Sonnet 4.6 59.2% 55.2 to 61.7 507 tasks; 20 trials; medium, Thinkingbox Feb 2026 Official leaderboard
5 GPT-6 Astra 58.3% ±4 507 tasks; 20 trials; medium, Thinkingbox Sep 2026 Official leaderboard
6 Kimi K3 57.4% ±3.1 507 tasks; 20 trials; medium, Thinkingbox Jul 2026 Official leaderboard
7 Qwen 3.8 27B 51.7% ±3.1 507 tasks; 20 trials; medium, Thinkingbox Aug 2026 Official leaderboard
8 GPT-5.2 46.3% ±3.2 507 tasks; 20 trials; medium, Thinkingbox Dec 2025 Official leaderboard
9 DeepSeek-V4-Pro 43.3% ±3.1 507 tasks; 20 trials; medium, Thinkingbox Apr 2026 Official leaderboard
10 Kimi K2.6 37.7% ±2.8 507 tasks; 20 trials; medium, Thinkingbox Apr 2026 Official leaderboard
11 GLM-5.1 33.2% ±3 507 tasks; 20 trials; medium, Thinkingbox Apr 2026 Official leaderboard
12 Qwen3.6 27B 32.9% ±2.8 507 tasks; 20 trials; medium, Thinkingbox Apr 2026 Official leaderboard
13 Claude Opus 4.6 32.1% ±3.5 507 tasks; 20 trials; medium, Thinkingbox Feb 2026 Official leaderboard
14 o3-pro 19.3% ±2.3 507 tasks; 20 trials; medium, Thinkingbox Jun 2025 Official leaderboard
15 Grok 4.3 14.4% ±2.2 507 tasks; 20 trials; medium, Thinkingbox Apr 2026 Official leaderboard
16 Qwen3.5-9B 5.65% ±1.2 507 tasks; 20 trials; medium, Thinkingbox Feb 2026 Official leaderboard
17 Mistral Large 3 4.66% 3.7 to 5.8 507 tasks; 20 trials; medium, Thinkingbox Dec 2025 Official leaderboard
18 MiniMax-M2.5 0.19% 0 to 0.5 507 tasks; 20 trials; medium, Thinkingbox Feb 2026 Official leaderboard

Rankings by part

We read 5 parts from the benchmark’s paper, each ranking models on one part of the test, such as Retail, Auto and Booking. Claude Opus 5 leads 3 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Retail

14 models we track have a score here.

  1. 1 Kimi K3 ↑5 82.2%
  2. 2 Claude Opus 5 ↓1 80.7%
  3. 3 GPT-5.4 ↓1 76.3%
  4. 4 Claude Sonnet 4.6 72.3%
  5. 5 GPT-6 Astra 71.7%
  6. 6 Claude Opus 4.6 ↑7 68.6%
  7. 7 DeepSeek-V4-Pro ↑2 68.2%
  8. 8 GPT-5.6 Sol ↓5 67.7%
  9. 9 Qwen 3.8 27B ↓2 64%
  10. 10 GLM-5.1 ↑1 58.7%

The best score, 66.5%, costs $0.32 a task

Runs on the cost frontier score higher than any run that's cheaper. GPT-5.6 Sol at 507 tasks; 20 trials; medium comes within 4.6 points of the best for $0.08 a task, 4 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

MiniMax OpenAI Moonshot AI Anthropic Cost frontier One model at each effort level
0% 20% 40% 60% 80% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 5, 507 tasks; 20 trials; medium GPT-5.6 Sol, 507 tasks; 20 trials; medium Kimi K2.6, 507 tasks; 20 trials; medium MiniMax-M2.5, 507 tasks; 20 trials; medium 0% 20% 40% 60% 80% $0.001 $0.01 $0.10 $1.00 $10 Cost per task, log scale Claude Opus 5, 507 tasks; 20 trials; medium GPT-5.6 Sol, 507 tasks; 20 trials; medium

Source: The Thinkingbox-bench paper

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5 507 tasks; 20 trials; medium Thinkingbox 66.5% $0.32
GPT-5.4 507 tasks; 20 trials; medium Thinkingbox 65.4% $0.09
GPT-5.6 Sol 507 tasks; 20 trials; medium Thinkingbox 61.9% $0.08
Claude Sonnet 4.6 507 tasks; 20 trials; medium Thinkingbox 59.2% $0.16
GPT-6 Astra 507 tasks; 20 trials; medium Thinkingbox 58.3% $0.17
Kimi K3 507 tasks; 20 trials; medium Thinkingbox 57.4% $0.14
Qwen 3.8 27B 507 tasks; 20 trials; medium Thinkingbox 51.7% $0.09
GPT-5.2 507 tasks; 20 trials; medium Thinkingbox 46.3% $0.09
DeepSeek-V4-Pro 507 tasks; 20 trials; medium Thinkingbox 43.3% $0.22
Kimi K2.6 507 tasks; 20 trials; medium Thinkingbox 37.7% $0.05
GLM-5.1 507 tasks; 20 trials; medium Thinkingbox 33.2% $0.11
Qwen3.6 27B 507 tasks; 20 trials; medium Thinkingbox 32.9% $0.11
Claude Opus 4.6 507 tasks; 20 trials; medium Thinkingbox 32.1% $0.25
o3-pro 507 tasks; 20 trials; medium Thinkingbox 19.3% $3.63
Grok 4.3 507 tasks; 20 trials; medium Thinkingbox 14.4% $0.06
Qwen3.5-9B 507 tasks; 20 trials; medium Thinkingbox 5.65% $0.06
Mistral Large 3 507 tasks; 20 trials; medium Thinkingbox 4.66% $0.06
MiniMax-M2.5 507 tasks; 20 trials; medium Thinkingbox 0.19% $0.006

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Benchmark paper Not tested since Sep 2026

Source: The Thinkingbox-bench paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 5 GPT-5.4 Claude Sonnet 4.6 GPT-5.2 o3-pro 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 5 Claude Sonnet 4.6

Source: Official Thinkingbox-bench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Tool use

All 265 in Tool use
  • AutomationBench In use Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state. Gemini 4 Neon 56.3%
  • MCP Atlas In use Multi-step workflows over real MCP servers. Scores tool-use competency on realistic tasks. Gemini 4 Argon 90.7%
  • BFCL V4 In use Models call functions and tools, including multi-turn agentic tasks with web search and memory. Checked against expected calls and outcomes. Qwen3.7-Max 75%
  • τ³-bench Banking In use Agents act as banking customer-service reps over many turns with a simulated user, following policy. Pass^1 on the final database state. Qwen 3.8 Max 55.1%
  • Toolathlon-Verified In use 108 long-horizon tasks across 32 apps through about 600 MCP tools. Execution-based checkers inspect final environment state. GLM-5.3-Flash 78.4%
  • DeepSearchQA In use Questions that need multi-step search and synthesis. Final answers graded against references. Kimi K3 95%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark

Catalog record 1 warning 1 pass
  • Warning Task access We don’t know whether the tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (18 entries).
  • Pass Models scored 18 models we track.
  • Pass Newest score The newest is from September 2026, for GPT-6 Astra.
  • Note Frontier models Only 1 of the top 10 has a score. The strongest tested is GPT-6 Astra (#4).
  • Pass Headroom The best is 66.5%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 18 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check