Skip to content
Four runsRecorded on screenBilled per API key

Empryo against OpenCode and Claude Code

EmpryoagainstOpenCodeandClaude Code

Same prompt, same repository and the same model unless a chart says otherwise. Each tool ran on its own Anthropic API key, so the console shows exactly what each one cost.

See the chartsRead the fine print
than OpenCode on a failing test
3.2× faster
2m 37s against 8m 29s
than OpenCode, 2.2× than Claude Code
5.6× cheaper
Same run, per the Anthropic console
Empryo's cost counter matched the bill
To the cent
OpenCode's left out 65% of its bill
context than Claude Code on newer Opus 4.7
35% less
42.9k against 66.3k tokens

A failing test: three tools, one prompt

3 April 2026. Prompt: “fix the failing test”. All three on Claude Opus 4.6, each on its own API key.

Time to fix

Time to fix, in minutes
ToolMinutesAgainst the baseline
Empryo, release 1.9.02m 37sBaseline
Claude Code, release 2.1.917m 41s2.9× the baseline
OpenCode, release 1.3.108m 29s3.2× the baseline
Screen recording, 3 April 2026. Post on X

Billed cost

Billed cost, in us dollars
ToolUS dollarsAgainst the baseline
Empryo, Opus 4.6 only$0.57Baseline
Claude Code, Opus 4.6 and Haiku$1.272.2× the baseline
OpenCode, Opus 4.6 only$3.175.6× the baseline
Anthropic console, total per API key, 3 April 2026. Post on X

Empryo ran the test, went straight to the code it was built with and read only the lines it needed. Its own counter said $0.57, and the console billed exactly $0.57. Claude Code also used Haiku on its key. When the recording ended six minutes in, OpenCode's counter read $1.21; its key was billed $3.17 by the end of the run.

Tokens in the whole run
427.7k
Read from the cache
384.0k, 91%
Written by the model
5.8k
Largest context
34.5k of 1M
The evidence, 6 images
  • Empryo 1.9.0 after the run: its context panel shows 427.7k tokens, 91% of them read from the cache, and a cost of $0.57, all on anthropic/claude-opus-4-6. Below it: completed in 2m 37s.
    Empryo's own counter when it finished: $0.57.
  • Anthropic console cost page for Empryo's API key, Claude Opus 4.6, 3 April 2026. Total token cost: USD 0.57.
    The Anthropic console for Empryo's key: USD 0.57, the same as its counter.
  • Anthropic console cost page for Claude Code's API key, all models, 3 April 2026. Total token cost: USD 1.27.
    The console for Claude Code's key: USD 1.27, Opus and Haiku.
  • Anthropic console cost page for OpenCode's API key, all models, 3 April 2026. Total token cost: USD 3.17.
    The console for OpenCode's key: USD 3.17.
  • Last frame of the recording, about six minutes in. Left: Empryo finished, all 45 tests pass, completed in 2m 37s. Top right: OpenCode still working, its counter at $1.21. Bottom right: Claude Code still working after 6m 20s.
    Six minutes in: Empryo done, the other two still working.
  • Table from the post. OpenCode: 8m 29s, $3.17, full Opus usage. Claude Code: 7m 41s, $1.27, mixed Opus and Haiku. Empryo: 2m 37s, $0.57, full Opus usage, the baseline.
    The table published with the recording.

Recording and images: x.com/BniWael/status/2040172009666015641

One bug, against OpenCode

9 April 2026. A restored session undid compaction. Both on Claude Opus 4.6, both fixed it with the same solution.

Time to a correct fix

Time to a correct fix, in minutes
ToolMinutesAgainst the baseline
Empryo, release 2.9.56m 22sBaseline
OpenCode, release 1.3.1011m 18s1.8× the baseline
Screen recording, 9 April 2026. Post on X

Billed cost

Billed cost, in us dollars
ToolUS dollarsAgainst the baseline
Empryo, Opus 4.6$1.70Baseline
OpenCode, Opus 4.6$3.522.1× the baseline
Anthropic console, keys session-bug-forge and session-bug-opencode. Post on X

  • Empryo

    Billed$1.70
    It said$1.70

    Matched the bill to the cent

  • OpenCode

    Billed$3.52
    It said$1.24

    Left out 65% of its bill

What each tool said it cost, against what the console billed its key. The dashed outline is money spent but not shown. OpenCode's counter did not count its subagents.

The evidence, 7 images
  • Anthropic console before the run: the keys session-bug-forge and session-bug-opencode both at USD 0.00, no data.
    Before the run: two fresh keys, both at $0.00.
  • Empryo's answer: after compaction the saved messages kept the whole history, so a restored session rebuilt it and the context meter went from 1% back to 15%. The fix replaces the saved messages with the compacted ones. Completed in 6m 22s.
    Empryo's fix, completed in 6m 22s.
  • Empryo's header after the run: 7% of the context window used and a cost of $1.70.
    Empryo's own counter: $1.70.
  • Anthropic console for the key session-bug-forge: total token cost USD 1.70.
    The console for Empryo's key: USD 1.70.
  • OpenCode after the run: Build, claude-opus-4-6, 11m 18s. Its footer shows 55.0K tokens and a cost of $1.24.
    OpenCode's own counter: $1.24 after 11m 18s.
  • Anthropic console for the key session-bug-opencode: total token cost USD 3.52.
    The console for OpenCode's key: USD 3.52.
  • Table from the post. Time: 6m 22s for Empryo, 11m 18s for OpenCode. Cost: $1.70 against $3.52. Both on Claude Opus 4.6, both a correct fix with the same solution.
    The table published with the recording.

Recording and images: x.com/BniWael/status/2042364421373121018

A layout change, against Claude Code on a newer model

16 April 2026. Empryo on Opus 4.6, Claude Code on Opus 4.7. Prompt: “The tabs bar should be inline with the legend of checkpoints, and legend of checkpoints should go to next line only on smaller terminal sizes.”

Time to finish

Time to finish, in minutes
ToolMinutesAgainst the baseline
Empryo, release 2.12.1, Opus 4.63m 29sBaseline
Claude Code, release 2.1.111, Opus 4.73m 21s0.96× the baseline
Screen recording, 16 April 2026. Post on X

Context used

Context used, in thousand tokens
ToolThousand tokensAgainst the baseline
Empryo, Opus 4.642.9kBaseline
Claude Code, Opus 4.766.3k1.5× the baseline
Screen recording, 16 April 2026. Post on X

Claude Code finished 8 seconds sooner and used 54.5% more context to do it. It ran on a Claude Max plan, so it has no per-run bill; Empryo's counter read $0.92.

What each change did, from the published table
AspectEmpryoClaude Code
Files changed23, one of them new
Responsive layoutflexWrap="wrap"hardcoded termWidth >= 110
Store subscriptionScoped to the legendApp level, whole tree
Props threaded throughNonesuppressCheckpointLegend
Legend removed from TabInstanceFullyPartly, a conditional remains
The evidence, 3 images
  • Opening frame: Empryo 2.12.1 on Claude Opus 4.6 on the left, Claude Code 2.1.111 on Opus 4.7 with medium effort on a Claude Max plan on the right, both given the same prompt.
    Before the run: the same prompt in both.
  • After the run. Left: Empryo's context panel, 42.9k of a 1M window, cost $0.92. Right: Claude Code cogitated for 3m 21s, context 66.3k.
    At the end: 42.9k of context for Empryo, 66.3k for Claude Code.
  • Table from the post. Model: Opus 4.6 against Opus 4.7. Time: 3m 29s against 3m 21s. Context: 42.9k against 66.3k, 54.5% more. Files changed: 2 against 3, one new. The responsive strategy, store subscription, prop threading and legend removal are compared row by row.
    The table published with the recording.

Recording and images: x.com/BniWael/status/2044826445382373759

An audit, against OpenCode

Prompt: “verify cost reporting is wired correctly”. Both on Claude Opus 4.6, same repository.

Time to report

Time to report, in minutes
ToolMinutesAgainst the baseline
Empryo, Opus 4.62m 00sBaseline
OpenCode, Opus 4.65m 56s3.0× the baseline
Empryo README, May 2026

Cost

Cost, in us dollars
ToolUS dollarsAgainst the baseline
Empryo, Opus 4.6$0.84Baseline
OpenCode, Opus 4.6$2.613.1× the baseline
Empryo README, May 2026

Correct findings

Correct findings, in findings out of 7
ToolFindings out of 7Against the baseline
Empryo, 100%7 of 7
OpenCode, 57%4 of 7
Empryo README, May 2026
Mistakes in each report
AspectEmpryoOpenCode
False alarms03
Wrong claims01

The fine print

All benchmarks
Share these runs