DeepSeek V4 as a Codex CLI Provider: Frontier-Class Coding at a Fraction of the Cost

codex-clideepseekv4custom-providercost-optimisationbenchmarksconfigurationmodel-selection

DeepSeek V4 as a Codex CLI Provider: Frontier-Class Coding at a Fraction of the Cost

DeepSeek V4 landed today — 24 April 2026 — and the numbers deserve attention. V4-Pro scores 80.6% on SWE-bench Verified while charging $3.48 per million output tokens1. For comparison, Claude Opus 4.6 scores 80.8% on the same benchmark at $25 per million output tokens2. That is a 7× price gap at near-identical coding performance. This article walks through configuring DeepSeek V4 as a Codex CLI provider, examines where the benchmarks genuinely hold up, and identifies the workflows where GPT-5.5 or Claude remain the better choice.

What Shipped

DeepSeek released two models under the MIT licence1:

  V4-Pro V4-Flash
Total parameters 1.6 trillion 284 billion
Active parameters 49 billion 13 billion
Context window 1 million tokens 1 million tokens
Max output 384K tokens 384K tokens
Training tokens 33 trillion 32 trillion
Weights size 865 GB 160 GB

Both use a Mixture-of-Experts (MoE) architecture with three key innovations: Hybrid Attention (CSA + HCA) that cuts KV cache to 10% of V3.2’s at 1M context, Manifold-Constrained Hyper-Connections that tame residual amplification from 3,000× to 1.6×, and the Muon optimiser for stable training at scale13.

Benchmark Reality Check

Headlines focus on SWE-bench, but the full picture is more nuanced:

Benchmark V4-Pro V4-Flash GPT-5.5 Claude Opus 4.7
SWE-bench Verified 80.6% 79.0% ⚠️ not reported 80.8%
SWE-bench Pro 55.4% — 58.6% 64.3%
Terminal-Bench 2.0 67.9% 56.9% 82.7% 65.4%
LiveCodeBench 93.5% 91.6% — 88.8%
Codeforces 3206 — — —

Sources: 1245

Three observations for practitioners:

  1. SWE-bench Verified is V4-Pro’s strong suit — it matches the frontier models on repository-level bug fixing1.
  2. Terminal-Bench 2.0 reveals the gap — GPT-5.5 leads V4-Pro by nearly 15 percentage points on complex multi-step terminal workflows4. If your work involves heavy iterative debugging and tool orchestration, GPT-5.5 remains substantially ahead.
  3. LiveCodeBench and Codeforces favour V4-Pro — pure algorithmic coding is a genuine strength, with V4-Pro’s 3206 Codeforces rating exceeding GPT-5.4’s 31681.

Pricing Comparison

The cost differential is the headline:

Model Input (cache miss) Input (cache hit) Output
DeepSeek V4-Flash $0.14/M $0.028/M $0.28/M
DeepSeek V4-Pro $1.74/M $0.145/M $3.48/M
GPT-5.4 — — —
GPT-5.5 ⚠️ credit-based ⚠️ credit-based ⚠️ credit-based
Claude Opus 4.6 $5/M — $25/M

Sources: 12

A production agentic pipeline processing 50 million output tokens per month costs approximately $174 on V4-Pro versus $1,250 on Claude Opus 4.63. V4-Flash is even more dramatic — at $0.28/M output, it undercuts GPT-5.4 Nano1.

Configuring Codex CLI for DeepSeek V4

DeepSeek’s API is OpenAI-compatible6, which means a straightforward custom provider definition in config.toml:

# ~/.codex/config.toml

[model_providers.deepseek]
name = "DeepSeek V4"
base_url = "https://api.deepseek.com"
env_key = "DS_KEY"
wire_api = "chat"

Set the environment variable referenced by env_key to your DeepSeek API key:

export DS_KEY="sk-your-key-here"

Model Profiles

Define profiles for different workflows:

[profiles.ds-pro]
model = "deepseek-v4-pro"
model_provider = "deepseek"

[profiles.ds-flash]
model = "deepseek-v4-flash"
model_provider = "deepseek"

Launch with a profile:

codex --profile ds-pro "Refactor the payment module to use the strategy pattern"
codex --profile ds-flash "Add unit tests for the UserService class"

Reasoning Modes

V4-Pro and V4-Flash support three reasoning modes: non-thinking, thinking, and thinking_max6. The default is non-thinking. To enable extended reasoning, you currently need to pass the appropriate parameters through the API — Codex CLI’s reasoning_effort config maps to OpenAI’s reasoning effort parameter, so you may need to experiment with how DeepSeek interprets these values through the compatibility layer.

flowchart LR
    A[Task arrives] --> B{Complexity?}
    B -->|Simple fix / lint| C[V4-Flash\nnon-thinking\n$0.28/M]
    B -->|Standard feature| D[V4-Pro\nnon-thinking\n$3.48/M]
    B -->|Complex refactor| E[V4-Pro\nthinking\nhigher token use]
    B -->|Critical / high-stakes| F[GPT-5.5 or\nClaude Opus 4.7]

Non-Interactive Pipelines

DeepSeek V4 works with codex exec for CI/CD automation:

codex exec \
  --profile ds-flash \
  --approval-mode full-auto \
  "Run the test suite and fix any failing tests" \
  -o results.json \
  --output-schema ./ci-schema.json

For cost-sensitive CI pipelines, V4-Flash at $0.28/M output makes per-commit agent runs economically viable in a way that frontier models do not.

Two-Tier Routing: The Practical Architecture

The most effective pattern is not choosing one model — it is routing between them. Use DeepSeek for the bulk of your work and escalate to GPT-5.5 for the tasks where Terminal-Bench performance matters:

# Default: DeepSeek V4-Pro for everyday coding
model = "deepseek-v4-pro"
model_provider = "deepseek"

# Escalation profile for complex multi-step debugging
[profiles.frontier]
model = "gpt-5.5"

# CI profile: V4-Flash for cost-sensitive automation
[profiles.ci]
model = "deepseek-v4-flash"
model_provider = "deepseek"

flowchart TD
    subgraph "Daily Development"
        A[Interactive sessions] --> B[DeepSeek V4-Pro]
        C[Quick questions / tests] --> D[DeepSeek V4-Flash]
    end
    subgraph "Escalation"
        E[Complex debugging] --> F[GPT-5.5]
        G[Security review] --> F
    end
    subgraph "CI/CD"
        H[Per-commit checks] --> D
        I[Nightly full review] --> B
    end

Known Limitations and Caveats

Before committing to DeepSeek V4 as your primary provider, understand the trade-offs:

  1. wire_api compatibility — DeepSeek supports the chat wire API (OpenAI Chat Completions compatible), not the responses wire API6. This means some Codex CLI features that rely on Responses API primitives (such as native apply_patch tool routing) may behave differently. Test your critical workflows thoroughly.

  2. Terminal-Bench gap — V4-Pro’s 67.9% versus GPT-5.5’s 82.7% on Terminal-Bench 2.04 translates to meaningfully worse performance on multi-step agentic workflows involving iterative tool use, error recovery, and plan revision. For complex debugging sessions, this gap is felt.

  3. Reasoning mode support — Codex CLI’s Alt+, / Alt+. reasoning effort shortcuts target OpenAI’s reasoning effort parameter. Mapping these to DeepSeek’s thinking / non-thinking / thinking_max modes requires care — the semantics are not identical6.

  4. Rate limits and availability — DeepSeek’s API has historically experienced availability issues during peak demand7. For production-critical workflows, configure failover:

# Primary: DeepSeek
model = "deepseek-v4-pro"
model_provider = "deepseek"

# Failover: switch profiles manually or via gateway
[profiles.fallback]
model = "gpt-5.4"
  1. Data residency — DeepSeek’s API routes through infrastructure in China7. Organisations with data sovereignty requirements should evaluate whether this meets their compliance posture, or consider self-hosting the open weights behind a private endpoint.

When to Use What

Workflow Recommended Model Rationale
Feature implementation V4-Pro Strong SWE-bench, 7× cheaper than Claude
Unit test generation V4-Flash Adequate quality at $0.28/M
Complex multi-step debugging GPT-5.5 82.7% Terminal-Bench dominance
CI/CD per-commit checks V4-Flash Economic at scale
Security-sensitive code review GPT-5.5 / Claude Opus 4.7 Higher precision on nuanced analysis
Competitive programming V4-Pro 3206 Codeforces, best-in-class
Regulated industries (data residency) GPT-5.5 or self-hosted V4 weights API routing considerations

Migration from DeepSeek V3

If you are already using DeepSeek models, update your model identifiers. The deepseek-chat and deepseek-reasoner model IDs deprecate on 24 July 20266. The migration is a single-line change:

# Before
model = "deepseek-chat"

# After
model = "deepseek-v4-pro"   # or "deepseek-v4-flash"

Conclusion

DeepSeek V4 does not replace GPT-5.5 — it occupies a different point on the cost-performance curve. At $3.48/M output with 80.6% SWE-bench Verified, V4-Pro is the most cost-effective frontier-adjacent coding model available today. The practical architecture is two-tier: DeepSeek for the 80% of tasks where it matches the frontier, GPT-5.5 for the 20% where Terminal-Bench performance and deep agentic reasoning justify the premium. Configure both providers, define profiles, and let the task complexity guide your routing.

Citations

  1. DeepSeek, “DeepSeek-V4-Pro Technical Report”, Hugging Face, 24 April 2026. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  2. VentureBeat, “OpenAI’s GPT-5.5 is here: narrowly beats Claude Mythos Preview on Terminal-Bench 2.0”, 23 April 2026. https://venturebeat.com/ai/openais-gpt-5-5-is-here-and-its-no-potato-narrowly-beats-anthropics-claude-mythos-preview-on-terminal-bench-2-0 ↩ ↩2 ↩3

  3. BuildFastWithAI, “DeepSeek V4-Pro Review: Benchmarks, Pricing & Architecture”, April 2026. https://www.buildfastwithai.com/blogs/deepseek-v4-pro-review-2026 ↩ ↩2

  4. MarkTechPost, “OpenAI Releases GPT-5.5: Terminal-Bench 2.0 and GDPval Scores”, 23 April 2026. https://www.marktechpost.com/2026/04/23/openai-releases-gpt-5-5-a-fully-retrained-agentic-model-that-scores-82-7-on-terminal-bench-2-0-and-84-9-on-gdpval/ ↩ ↩2 ↩3

  5. Lushbinary, “DeepSeek V4 vs Claude Opus 4.7 vs GPT-5.5: Benchmarks & Pricing”, April 2026. https://lushbinary.com/blog/deepseek-v4-vs-claude-opus-4-7-vs-gpt-5-5-comparison/ ↩

  6. DeepSeek, “API Documentation — Integrate with Coding Agents”, April 2026. https://api-docs.deepseek.com/guides/coding_agents ↩ ↩2 ↩3 ↩4 ↩5

  7. Simon Willison, “DeepSeek V4 — almost on the frontier, a fraction of the price”, 24 April 2026. https://simonwillison.net/2026/Apr/24/deepseek-v4/ ↩ ↩2