AnyModel vs OpenCode — local coding agent benchmark

Historical evidence · interpretation clarified October 2026. This is one local Qwen3-Coder 30B model, six small tasks, three repetitions per arm, measured in June 2026. The fixed arm has 17/18 passing artifacts but 16/18 clean completed runs: one passing artifact was produced by a run that timed out. Artifact success is not successful agent completion. This is not a frontier-model comparison, a current hosted-model ranking, or a general speed guarantee. The raw results and original measurements below are retained unchanged. Current product scope · Baseline raw JSON · Fixed raw JSON.

Model: qwen3-coder:30b on Ollama (M1 Max 32GB, 100% GPU) · Wed, 10 Jun 2026 21:17:50 GMT · 54 runs · identical seeded tasks, artifact-verified, fresh workspace per run

AnyModel v1.16.2

Claude Code → anymodel proxy → Ollama

task success (0/18 runs)

median turn-around: 3.7s
tool calls executed: 9 (bash 3 · edit 0 · read 4)
timeouts: 0 · zero-tool turns: 9
skill invocations observed: 2

OpenCode v1.16.2

OpenCode → Ollama direct

78%

task success (14/18 runs)

median turn-around: 118.4s
tool calls executed: 17 (bash 0 · edit 17 · read 0)
timeouts: 0 · zero-tool turns: 3
skill invocations observed: 0

AnyModel (fixed) v1.17.0

Claude Code → anymodel proxy → Ollama

94%

task success (17/18 runs)

median turn-around: 10.2s
tool calls executed: 55 (bash 11 · edit 18 · read 10)
timeouts: 1 · zero-tool turns: 0
skill invocations observed: 6

Median wall time per task (seconds — lower is better)

AnyModel

3.7

OpenCode

118.4

AnyModel (fixed)

10.2

Per-task results (pass / runs · median time · tool calls)

task	AnyModel	OpenCode	AnyModel (fixed)
write-file	0/33.0s · 0 calls	2/3105.4s · 2 calls	3/34.3s · 3 calls
bash-count	0/33.3s · 1 calls	2/3112.0s · 2 calls	3/38.6s · 5 calls
fix-bug	0/33.8s · 3 calls	3/3130.6s · 3 calls	3/3283.7s · 29 calls
multi-file	0/33.7s · 0 calls	3/3119.6s · 6 calls	3/311.3s · 6 calls
edit-json	0/33.6s · 3 calls	3/3120.4s · 3 calls	3/310.3s · 6 calls
skill-stamp	0/34.8s · 2 calls	1/3101.1s · 1 calls	2/37.2s · 6 calls

Failures

arm	run	class	time	workspace (logs)
AnyModel	write-file #1	zero-tool-execution	4.4s	`/tmp/bench-runs/baseline-1.16.2/anymodel/write-file/rep1`
AnyModel	write-file #2	zero-tool-execution	3.0s	`/tmp/bench-runs/baseline-1.16.2/anymodel/write-file/rep2`
AnyModel	write-file #3	zero-tool-execution	2.9s	`/tmp/bench-runs/baseline-1.16.2/anymodel/write-file/rep3`
OpenCode	write-file #3	zero-tool-execution	103.5s	`/tmp/bench-runs/baseline-1.16.2/opencode/write-file/rep3`
AnyModel	bash-count #1	zero-tool-execution	8.0s	`/tmp/bench-runs/baseline-1.16.2/anymodel/bash-count/rep1`
AnyModel	bash-count #2	wrong-artifact	3.3s	`/tmp/bench-runs/baseline-1.16.2/anymodel/bash-count/rep2`
AnyModel	bash-count #3	zero-tool-execution	2.9s	`/tmp/bench-runs/baseline-1.16.2/anymodel/bash-count/rep3`
OpenCode	bash-count #1	wrong-artifact	111.5s	`/tmp/bench-runs/baseline-1.16.2/opencode/bash-count/rep1`
AnyModel	fix-bug #1	wrong-artifact	5.6s	`/tmp/bench-runs/baseline-1.16.2/anymodel/fix-bug/rep1`
AnyModel	fix-bug #2	wrong-artifact	3.8s	`/tmp/bench-runs/baseline-1.16.2/anymodel/fix-bug/rep2`
AnyModel	fix-bug #3	wrong-artifact	3.2s	`/tmp/bench-runs/baseline-1.16.2/anymodel/fix-bug/rep3`
AnyModel	multi-file #1	zero-tool-execution	6.8s	`/tmp/bench-runs/baseline-1.16.2/anymodel/multi-file/rep1`
AnyModel	multi-file #2	zero-tool-execution	3.7s	`/tmp/bench-runs/baseline-1.16.2/anymodel/multi-file/rep2`
AnyModel	multi-file #3	zero-tool-execution	3.7s	`/tmp/bench-runs/baseline-1.16.2/anymodel/multi-file/rep3`
AnyModel	edit-json #1	wrong-artifact	4.9s	`/tmp/bench-runs/baseline-1.16.2/anymodel/edit-json/rep1`
AnyModel	edit-json #2	wrong-artifact	3.6s	`/tmp/bench-runs/baseline-1.16.2/anymodel/edit-json/rep2`
AnyModel	edit-json #3	wrong-artifact	3.5s	`/tmp/bench-runs/baseline-1.16.2/anymodel/edit-json/rep3`
AnyModel	skill-stamp #1	zero-tool-execution	4.8s	`/tmp/bench-runs/baseline-1.16.2/anymodel/skill-stamp/rep1`
AnyModel	skill-stamp #2	wrong-artifact	5.5s	`/tmp/bench-runs/baseline-1.16.2/anymodel/skill-stamp/rep2`
AnyModel	skill-stamp #3	wrong-artifact	3.7s	`/tmp/bench-runs/baseline-1.16.2/anymodel/skill-stamp/rep3`
OpenCode	skill-stamp #1	zero-tool-execution	101.0s	`/tmp/bench-runs/baseline-1.16.2/opencode/skill-stamp/rep1`
OpenCode	skill-stamp #3	zero-tool-execution	101.1s	`/tmp/bench-runs/baseline-1.16.2/opencode/skill-stamp/rep3`
AnyModel (fixed)	skill-stamp #1	wrong-artifact	9.8s	`/tmp/bench-runs/fixed-1.17.0/anymodel/skill-stamp/rep1`

Methodology. Each run: fresh temp workspace, seeded fixtures, one prompt, 300s timeout. Success = artifact verifier only (file contents / passing test), never model self-report. AnyModel arm: claude -p → anymodel proxy → Ollama native API; tool calls counted from Claude Code stream-json events; proxy log sliced per run. OpenCode arm: opencode run → Ollama /v1; tool calls parsed from its execution trace. Same Ollama server, same context length (32768), flash attention + q8_0 KV cache. Skill task seeds the same SKILL.md into .claude/skills, .opencode/skill, and .opencode/skills.