Home / Guides / DeepSeek V4 Flash

Published July 31, 2026 · applies to pgup-ai/jbot-review-action v0

The official DeepSeek V4 Flash 0731 can review pull requests free through OpenCode—or directly through DeepSeek at roughly 11× lower uncached input and 27× lower output list prices than Gemini 3.6 Flash. Vendor-published agent scores put the two models in the same performance band, with an important harness caveat explained below.

Free OpenCode setup in three steps

1. Create an OpenCode API key. Sign in to OpenCode Zen, then save the key as the repository secret OPENCODE_API_KEY (Settings → Secrets and variables → Actions).
2. Commit the workflow. Add the file below as .github/workflows/jbot-review.yml. OpenCode's deepseek-v4-flash-free alias is updated to the official model; the explicit pin makes that choice visible.
3. Open a pull request. J-Bot reads the base…head diff and posts diff-anchored findings with a verdict. Blocking findings are checked by a second session before they post.

name: J-Bot Code Review
on:
  pull_request: { types: [opened, reopened, ready_for_review, synchronize] }

concurrency:
  group: jbot-review-${{ github.event.pull_request.number }}
  cancel-in-progress: true

permissions:
  contents: read
  pull-requests: write
  issues: write
  checks: read

jobs:
  review:
    if: github.event.pull_request.draft == false
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v7
        with: { fetch-depth: 0 }
      - uses: pgup-ai/jbot-review-action@v0
        with:
          provider: opencode
          model: deepseek-v4-flash-free
          opencode-api-key: ${{ secrets.OPENCODE_API_KEY }}
          github-token: ${{ secrets.GITHUB_TOKEN }}

Free gateway or first-party API

J-Bot already defaults each provider to the model shown above. Pinning it makes the workflow self-explanatory. The OpenCode route is free during its current feedback window; DeepSeek's direct route uses the published token prices below.

The official 0731 agent benchmarks

DeepSeek published nine agent and coding-agent results with the official release. These are DeepSeek's reported scores, not J-Bot evaluations:

DeepSeek identifies DSBench-FullStack and DSBench-Hard as internal test sets. For the public code-agent benchmarks, it used the not-yet-released DeepSeek Harness in minimal mode, max effort, temperature=1.0, and top_p=0.95. Read the release notes →

Same agent-performance band, very different price

The overlapping vendor tables put official V4 Flash in Gemini 3.6 Flash's agentic coding range. DeepSeek's reported scores are slightly higher on the two named overlaps—but this is directional, not a controlled head-to-head.

Why the qualification matters: DeepSeek used its own forthcoming harness at max effort; Google's Terminal-Bench result uses the Terminus-2 harness, and Google labels its software-engineering result DeepSWE v1.1. Prices are directly comparable public API list rates checked July 31, 2026. DeepSeek says a future peak window will charge 2× its regular rates, with the effective date still to be announced. Sources: DeepSeek pricing and Google's Gemini 3.6 Flash model card.

Normal reviews versus the published benchmark setup

J-Bot's native-provider default is medium reasoning effort. DeepSeek ran the released code-agent benchmarks at max effort. To ask J-Bot for the closest available effort setting, add this input:

model-options: '{"reasoningEffort":"max"}'

This still does not reproduce DeepSeek's benchmark: its harness has not been released, and J-Bot uses its own review prompts, read-only tools, finding schema, and verification pass. Use max when depth matters more than latency; keep the default medium for everyday PR throughput.

Where the diff goes

• Your runner stays in control. The review runs inside GitHub Actions with read-only repository access; write permissions are used only to post review output.
• The provider you select receives the diff. With provider: deepseek, that is DeepSeek's first-party API. With provider: opencode, that is OpenCode Zen and its upstream.
• The free OpenCode route is a feedback route. OpenCode states that data collected through DeepSeek V4 Flash Free may be used to improve the model. Do not send sensitive code through that route unless its terms fit your policy.
• Fork PRs cannot read the key. GitHub does not pass repository secrets to pull_request workflows from forks.

FAQ

Related