Coding Horizon

Claude Code Has A New Problem Nobody Is Measuring

Every figure this video puts on screen, chased to a primary source.

1. The resource benchmark

Antigma’s own measurement page for Ante.

Source: Antigma Labs, “Resource Footprint”, Ante documentation. https://docs.antigma.ai/benchmarks/compare_table

Methodology, in Antigma’s words: “we ran the same 20 tasks in parallel through Ante, Claude Code, and Opencode, each inside Docker with identical constraints, and recorded CPU, memory, and disk usage throughout the run.”

Wall time (seconds)

Agent Wall time
Ante 940
Claude Code 627
Opencode 1076

CPU usage (%)

Agent Peak Avg P95 P99
Ante 94.4 1.3 6.2 12.3
Claude Code 89.5 12.1 31.0 43.4
Opencode 90.8 3.8 27.1 62.3

Average CPU ratio, Claude Code against Ante: 12.1 / 1.3 = 9.3x. The video says “roughly nine times higher”.

Memory usage (MiB)

Agent Peak Avg P95 P99
Ante 1968 683 1489 1550
Claude Code 13877 3685 8927 9535
Opencode 12944 2077 11266 12852

Peak memory ratio, Claude Code against Ante: 13877 / 1968 = 7.05x. The video says “roughly seven times higher”.

Disk usage (MiB)

Agent Peak Avg P95 P99
Ante 7041 3121 6975 6976
Claude Code 22467 4304 10128 10193
Opencode 59689 6046 29108 34744

Disk read rate (MB/s)

Agent Peak P95 P99
Ante 3.5 0.0 0.1
Claude Code 263.9 10.4 101.9
Opencode 284.1 0.1 10.6

Disk write rate (MB/s)

Agent Peak P95 P99
Ante 186.3 3.6 61.7
Claude Code 302.3 26.6 113.0
Opencode 302.9 14.5 296.6

Total disk I/O (MB)

Agent Total read Total write
Ante 24 2785
Claude Code 17444 15116
Opencode 2224 31427

Total write ratio, Claude Code against Ante: 15116 / 2785 = 5.4x. That is the row the “five times less disk I/O” headline rests on. Total read ratio is 17444 / 24 = 727x, and 17444 MB is 17.4 GB, which is what the video compares against Ante’s 24 MB.

2. What Ante is

Source: Antigma Labs, Ante repository README. https://github.com/AntigmaLabs/ante

Source: Antigma Labs, “Philosophy”, Ante documentation. https://docs.antigma.ai/start/philosophy

The stated thesis is “cellular-native” agents: “agents should be lightweight enough to run by the thousands”. “We’re running hundreds and thousands of agent replicas. Each one can’t cost a couple GB of memory.”

The sentence the video calls the most important one in the docs is on the resource page: “Mass parallelism only pays off if a single agent is cheap enough to multiply.”

3. Terminal sessions

Source: Antigma Labs, “Interactive TUI”, Ante documentation. https://docs.antigma.ai/usage/tui

“The agent drives interactive programs — dev servers, REPLs, ssh, even other CLI agents like Claude Code or Codex — in durable named tmux sessions … Sessions live in the ante- namespace and survive Ante restarts: a later conversation can pick up a running server where it was left.”

/term <name> attaches or detaches a named session (https://docs.antigma.ai/usage/slash-commands).

4. Local models

Source: Antigma Labs, “Offline Mode”, Ante documentation. https://docs.antigma.ai/local/offline

Offline mode runs the full agent loop against a local GGUF model through a managed llama.cpp build. “Once a model is on disk, inference needs no network connection at all.”

Source: Antigma Labs, “Verified Models”, Ante documentation. https://docs.antigma.ai/local/verified-models

Model Quant Download Max context Status
Qwen3.8 27B Q4_K_M 17 GB 256K Verified
MiniMax M2.7 Q3_K_XL 102 GB 192K Verified
Qwen3.5 122B A10B Q2_K 45 GB 256K Verified
Qwen3.6 35B A3B Q4_K_M 22 GB 256K Verified
Qwen3.6 27B Q4_K_M 17 GB 256K Evaled, 56.2% on Terminal Bench 2.1
GLM 4.7 Flash Q4_K_M 18 GB 198K Verified
Gemma 4 26B A4B Q4_K_M 16 GB 256K Verified
Devstral Small 2 24B Q4_K_M 14 GB 384K Verified
Qwen3.5 9B Q4_K_M 5.7 GB 256K Verified
Gemma 4 E4B Q4_K_M 4.7 GB 128K Verified

Antigma defines the two statuses on the same page. Verified means “the model loads, the chat template resolves, and tool calling works in Ante’s agent loop. Not yet through the full eval.” Evaled means it “ran the full public benchmark: same Terminal-Bench 2.1 harness, pinned builds, and auditable Harbor runs”.

The one Evaled row: “Qwen3.6 27B reached 56.2% accuracy across 89 Terminal-Bench 2.1 tasks with 445 trials — a 17 GB download holding its own on the same benchmark we use for frontier models.”

5. Ante’s own benchmark constraints

Source: Antigma Labs, “Terminal-Bench”, Ante documentation. https://docs.antigma.ai/benchmarks/eval

“All trials use the same parameters as the official Terminal-Bench leaderboard: 89 tasks, 5 trials per task, strict timeouts and hardware limits.” 89 x 5 = 445 trials.

Ante’s own headline figure at the time of writing is 82.7% on Terminal Bench 2.1 with DeepSeek V4 Flash 0731, across 368 of 445 trials passing.

6. Qwen3.8 27B

Source: Qwen, Qwen3.8-27B model card, Hugging Face. https://huggingface.co/Qwen/Qwen3.8-27B

Benchmark rows quoted in the video:

Benchmark Qwen3.8 27B Qwen3.6 27B Opus 4.6 Max
Terminal Bench 2.1 (Terminus) 73.0 63.4 78.2
SWE Bench Pro 61.7 53.5 53.4

The model card states that SWE Bench Pro figures other than Opus 4.6 Max were produced with the Claude Code harness at temperature 1.0, top_p 0.95, and a 256K context window. That is a different harness from the one Ante’s own 56.2% figure was produced under, which is exactly why the video says the two sets of numbers should not be combined.

Caveats