Benchmark Almanac

Agent-Diff

DeepSeek-V3.2 88.1% leads since February 2026
Not in model cards

Agents use code to complete tasks across replicas of Slack, Linear, Box, and Calendar. The evaluator compares the resulting application state with the expected changes.

In February 2026, DeepSeek V3.2 led nine models with an overall score of 88.1. The table also reports Gemini 3 Flash, GPT-OSS 120B, and Claude Haiku 4.5.

Released
February 2026
Kind
Apps and MCP servers
Level
University
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Code, Actions
Interaction
Agentic
Environment
APIs and tools
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
Gemini 3 Flash #60 on the capability index
Last model added
December 2025
Scores refreshed
November 20th, 2026
9
Entries in its paper, 8 models we track

DeepSeek-V3.2 leads with 88.1% among models released by December 2025 DeepSeek-V3.2 leads with 88.1% among models released by December 2025

Best score per entry 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. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in December 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 DeepSeek-V3.2 No documentation 88.1%
  2. 2 Devstral 2 No documentation 86%
  3. 3 qwen3-vl-235b (model ID disputed) No documentation 79.2%
  4. 4 Kimi K2 Sep 2025, No documentation 75.4%
  5. 5 Grok 4.1 Fast No documentation 74.9%
  6. 6 Gemini 3 Flash No documentation 73.8%
  7. 7 gpt-oss-120b No documentation 68.5%
  8. 8 Claude Haiku 4.5 No documentation 49.3%
  9. 9 Llama 4 Scout No documentation 38%

Source: The Agent-Diff paper

#EntryScore ±SettingDateSource
1 DeepSeek-V3.2 88.1% ±2.4 No documentation Dec 2025 Official leaderboard
2 Devstral 2 86% ±3.1 No documentation Dec 2025 Official leaderboard
3 qwen3-vl-235b (model ID disputed) 79.2% ±3.3 No documentation Official leaderboard
4 Kimi K2 Sep 2025 75.4% ±3.5 No documentation Sep 2025 Official leaderboard
5 Grok 4.1 Fast 74.9% ±2.7 No documentation Nov 2025 Official leaderboard
6 Gemini 3 Flash 73.8% ±3.7 No documentation Dec 2025 Official leaderboard
7 gpt-oss-120b 68.5% ±4.1 No documentation Aug 2025 Official leaderboard
8 Claude Haiku 4.5 49.3% ±5.5 No documentation Oct 2025 Official leaderboard
9 Llama 4 Scout 38% ±4.1 No documentation Apr 2025 Official leaderboard
#ModelScore ±SettingDateSource
1 DeepSeek-V3.2 88.1% ±2.4 No documentation Dec 2025 Official leaderboard
2 Devstral 2 86% ±3.1 No documentation Dec 2025 Official leaderboard
3 Kimi K2 Sep 2025 75.4% ±3.5 No documentation Sep 2025 Official leaderboard
4 Grok 4.1 Fast 74.9% ±2.7 No documentation Nov 2025 Official leaderboard
5 Gemini 3 Flash 73.8% ±3.7 No documentation Dec 2025 Official leaderboard
6 gpt-oss-120b 68.5% ±4.1 No documentation Aug 2025 Official leaderboard
7 Claude Haiku 4.5 49.3% ±5.5 No documentation Oct 2025 Official leaderboard
8 Llama 4 Scout 38% ±4.1 No documentation Apr 2025 Official leaderboard

Rankings by part

We read 4 parts from the benchmark’s paper, each ranking models on one part of the test, such as Box, Calendar and Linear. DeepSeek-V3.2 leads 3 of them, and Gemini 3 Flash leads the other one. Arrows show how far a model moved from its overall rank among the models we track.

Box

8 models we track have a score here.

  1. 1 Gemini 3 Flash ↑4 80.3%
  2. 2 Devstral 2 79%
  3. 3 DeepSeek-V3.2 ↓2 76.6%
  4. 4 gpt-oss-120b ↑2 70.1%
  5. 5 Kimi K2 Sep 2025 ↓2 66.5%
  6. 6 Grok 4.1 Fast ↓2 58.5%
  7. 7 Claude Haiku 4.5 45.1%
  8. 8 Llama 4 Scout 33.7%

The best score, 88.1%, costs $0.03 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

xAI DeepSeek Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.02 $0.05 $0.10 $0.20 $0.50 $1.00 Cost per task, log scale DeepSeek-V3.2, No documentation Grok 4.1 Fast, No documentation 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 Cost per task, log scale DeepSeek-V3.2, No documentation Grok 4.1 Fast, No documentation

Source: The Agent-Diff paper

ModelEffortHarnessScoreCost per taskOutput tokens
DeepSeek-V3.2 No documentation 88.1% $0.03
Devstral 2 No documentation 86% $0.08
Kimi K2 Sep 2025 No documentation 75.4% $0.04
Grok 4.1 Fast No documentation 74.9% $0.01
Gemini 3 Flash No documentation 73.8% $0.05
gpt-oss-120b No documentation 68.5% $0.02
Claude Haiku 4.5 No documentation 49.3% $0.22
Llama 4 Scout No documentation 38% $0.02

The best score, 88.1%, costs $0.03 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

xAI DeepSeek Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.02 $0.05 $0.10 $0.20 $0.50 $1.00 Cost per task, log scale DeepSeek-V3.2, No documentation Grok 4.1 Fast, No documentation 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 Cost per task, log scale DeepSeek-V3.2, No documentation Grok 4.1 Fast, No documentation

Source: The Agent-Diff paper

ModelEffortHarnessScoreCost per taskOutput tokens
DeepSeek-V3.2 No documentation 88.1% $0.03
Devstral 2 No documentation 86% $0.08
Kimi K2 Sep 2025 No documentation 75.4% $0.04
Grok 4.1 Fast No documentation 74.9% $0.01
Gemini 3 Flash No documentation 73.8% $0.05
gpt-oss-120b No documentation 68.5% $0.02
Claude Haiku 4.5 No documentation 49.3% $0.22
Llama 4 Scout No documentation 38% $0.02

Number one over time

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

DeepSeek
Benchmark paper Passed 85% in Feb 2026, by DeepSeek-V3.2

Source: The Agent-Diff paper

DeepSeek-V3.2 on the official leaderboard passed 85% in December 2025

Models released before it came out already scored that high. 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%) DeepSeek-V3.2 Kimi K2 gpt-oss-120b Llama 4 Scout 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) DeepSeek-V3.2 Kimi K2

Source: Official Agent-Diff 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

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the benchmark’s paper (9 entries).
  • Pass Models scored 8 models we track, plus 1 other entry on its leaderboard.
  • Note Newest score The newest came 11 months ago, in December 2025, for Gemini 3 Flash. That’s expected for a saturated test.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Gemini 3 Flash (#60).
  • Note Headroom Saturated. The best is 88.1%, past 85% since December 2025.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 8 rows for models we track and 1 other entry 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