Verdict: DeepSeek V4 Pro is the strongest open-weight coding model available in July 2026 — and at $0.87 per million output tokens, it costs roughly 57x less than Claude Fable 5's $50/M. But the leak-driven narrative that it "destroyed" Claude Fable 5 is misleading: the 0.2-point SWE-bench gap everyone quotes compares V4 Pro to the previous-generation Claude Opus 4.6, not to Fable 5. On independent benchmarks, Fable 5 leads V4 Pro by double digits on coding and agentic tasks. The real story is that the open-vs-closed gap has narrowed dramatically on price-performance, while the capability gap — especially for long-horizon agentic work — still favors closed models.
Last verified: 2026-07-21
- V4 Pro: 1.6T params / 49B active, MIT license, $0.435/M input + $0.87/M output
- V4 Flash: 284B params / 13B active, $0.14/M input + $0.28/M output
- Claude Fable 5: $10/M input + $50/M output, 1M context, released June 9, 2026
- SWE-bench Verified: V4 Pro 80.6% (vendor-reported) vs Opus 4.6 80.8% — not vs Fable 5
- V4 GA rollout began ~July 19, 2026; legacy API endpoints retire July 24, 2026
- Pricing/limits change often — last checked July 21, 2026.
What Is DeepSeek V4 Pro and Why Is Everyone Talking About It?
DeepSeek V4 Pro is the flagship tier of DeepSeek's V4 model family, a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters and 49 billion active per token. It shipped on April 24, 2026 as a preview, open-weight under the MIT license, with a 1-million-token context window and weights published on Hugging Face. The technical report is available on arXiv (2606.19348).
The reason it's generating headlines in July 2026 is the GA (general availability) rollout. On July 4, 2026, leaked model strings — deepseek-v4-pro-202606 and deepseek-v4-flash-202605 — appeared on X (Reported, via kie.ai). The naming pattern matches DeepSeek's convention of model-name followed by year-month, and both dates land after the April preview. By July 14, grayscale testing reportedly began for select users, with GA rollout starting around July 19 (wan27.org, Reported). DeepSeek has not published an official GA announcement as of July 21, 2026 — so all GA-specific claims carry a Reported confidence label, not Confirmed.
How Does DeepSeek V4 Pro Compare to Claude Fable 5 on Benchmarks?
Claude Fable 5 leads DeepSeek V4 Pro on coding and agentic benchmarks by a meaningful margin — but the often-quoted "0.2-point tie" is comparing V4 Pro to the wrong Claude model. Here's what the data actually shows:
| Benchmark | DeepSeek V4 Pro (Max) | Claude Opus 4.6 | Claude Fable 5 | Source |
|---|---|---|---|---|
| SWE-bench Verified | 80.6% | 80.8% | ~96% (estimated from Sonnet 5's 85.2%) | DeepSeek tech report, CodingFleet |
| SWE-bench Pro | 54.4% | — | 80% | BenchLM.ai |
| Arena Elo (Code) | 1,457 | — | 1,649 | AI Release Tracker |
| LiveCodeBench | 93.5% | 88.8% | — | DeepSeek tech report |
| Codeforces Elo | 3,206 | — | — | DeepSeek tech report |
The critical context: the 80.6% SWE-bench Verified score for V4 Pro comes from DeepSeek's own technical report — it's the vendor's number about the vendor's model, not an independent evaluation. On independent leaderboards, V4 Pro has been outscored by other open models. Artificial Analysis ranks V4 Pro second to Kimi K2.6 on several public coding tests (kie.ai, Reported). Scale's SEAL SWE-bench Pro leaderboard lists no DeepSeek V4 entry at all; the top open-weights entry is qwen3-coder-480b at 38.7% (MorphLLM).
Meanwhile, Claude Fable 5 — Anthropic's Mythos-class model released June 9, 2026 — sits above the Opus line and is built for "days-long, complex, and asynchronous tasks" including multi-day autonomous coding sessions. On BenchLM's aggregate BenchAlign score, Fable 5 leads V4 Pro (High) 83.68 to 55.47, with a 19.4-point margin in coding and a 14-point margin in agentic tasks (BenchLM.ai, Verified July 20, 2026).
The bottom line on benchmarks: V4 Pro is the best open-weight coding model. It is not at parity with Fable 5. The "destroyed" narrative confuses a comparison against the previous generation (Opus 4.6) with a comparison against the current top tier (Fable 5).
How Much Does Each Model Cost?
DeepSeek V4 Pro is dramatically cheaper than Claude Fable 5 — roughly 57x less on output tokens. Here's the pricing comparison at list rates as of July 2026:
| Model | Input (\(/M tokens) | Output (\)/M tokens) | Context Window | License | |
|---|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 1M tokens | MIT (open-weight) |
| DeepSeek V4 Pro | $0.435 | $0.87 | 1M tokens | MIT (open-weight) |
| Claude Opus 4.8 | $5.00 | $25.00 | 1M tokens | Proprietary |
| Claude Fable 5 | $10.00 | $50.00 | 1M+ tokens | Proprietary |
Sources: DeepSeek API docs, Anthropic pricing, MorphLLM
A practical cost example: for a workload of 20 requests per day (50K input + 10K output each, ~1M input + 200K output daily), V4 Pro costs about $18/month. Claude Fable 5 costs about $600/month for the same volume — a 33x difference in monthly spend (MorphLLM).
DeepSeek V4 Pro also offers prompt caching at $0.003625/M input tokens on cache hits — 138x cheaper than Fable 5's $1/M cache-hit rate. This makes V4 Pro especially cost-effective for workloads with repetitive system prompts or large context that stays constant across requests.
What Does the V4 GA Leak Actually Claim?
The GA (general availability) leak claims center on four improvements over the April preview, all currently Reported — not confirmed by DeepSeek:
- Stronger agentic behavior — the model can run longer chains of tool-calling tasks without degrading, which matters for autonomous coding agents.
- Better coding performance — testers report tighter answers with less rambling across multi-turn sessions.
- 3D and game generation — testers generated full HTML games (including Minecraft-style and No Man's Sky-style hybrids) from single prompts, though this is a demo capability, not a production workflow.
- Performance approaching Opus 4.8 — early testers report overall performance near Opus 4.8 levels, though this predates comparison with Fable 5 (wan27.org, Reported).
The architecture itself — parameter counts, MoE structure, context window — has not changed between Preview and GA. What has changed is reasoning quality, agentic capability, and potentially pricing strategy. The core V4 hybrid attention design (Compressed Sparse Attention + Heavily Compressed Attention) that enables the 1M-token context at 27% of V3.2's inference FLOPs and 10% of its KV cache remains the same (arXiv 2606.19348, Confirmed).
Which Model Should You Use for Your Workload?
Choose based on your actual workload, not leaderboard headlines. Here's a decision framework:
Choose DeepSeek V4 Pro if:
- You write a lot of code and need a strong, cost-effective coding model for bounded, single-file or multi-file tasks
- You handle huge documents (the 1M context window means you can feed entire codebases or long documents in one call)
- You want to run models on your own hardware (MIT license means you can download weights from Hugging Face and self-host)
- Your API spend is a material cost and a 30-57x reduction matters
- You need fine-tuning, air-gapped deployment, or sovereign-cloud compliance
Choose Claude Fable 5 if:
- You run long-horizon agentic sessions (multi-day autonomous coding, complex migrations, multi-stage knowledge work)
- Your agent operates across many tools with retries, partial failures, and recovery — Fable 5's "relentlessly proactive" behavior handles messier real-world conditions (Simon Willison, Confirmed)
- You need vision/multimodal capabilities (V4 Pro is text-only in the April 2026 preview)
- You need SOC 2 / HIPAA compliance on the API endpoint (use AWS Bedrock or Azure AI Foundry for Fable 5)
- The cost difference is outweighed by fewer botched runs that require human cleanup
Use both (routing strategy): If you have sustained volume above ~10,000 tasks/month, route the easy 90% of traffic to V4 Pro/Flash and reserve Fable 5 for the hard 10%. This is the cost-optimal approach that many production teams are adopting in 2026 (Totalum).
Choose V4 Flash (not Pro) if:
- You need speed and high-volume processing, not maximum reasoning depth
- You want to run on modest hardware (13B active parameters is tractable on A100 80GB or single H100 nodes)
- Your tasks are standard chat, retrieval, summarization, or mid-tier agent work
How Do You Migrate to DeepSeek V4 Before the July 24 Deadline?
DeepSeek is retiring the legacy deepseek-chat and deepseek-reasoner API model names on July 24, 2026 at 15:59 UTC. After that, any code calling those names returns errors — no grace period, no fallback (DeepSeek API docs, Confirmed). Here's the migration:
- Find every reference. Search your codebase, environment variables, and automation platforms (n8n, Make, Zapier) for the strings
deepseek-chatanddeepseek-reasoner. Also check vendor software that might route to DeepSeek behind the scenes. - Replace model names.
deepseek-chat→deepseek-v4-flash(non-thinking mode).deepseek-reasoner→deepseek-v4-flashwith thinking enabled, ordeepseek-v4-proif you need the higher tier. - Set the thinking parameter. V4 collapses what used to be separate models into a single model ID with a reasoning-effort setting. In OpenAI-compatible format:
extra_body={"thinking": {"type": "enabled"}}or"disabled". Tune depth withreasoning_effortset to"high"or"max"(DeepSeek API docs, Confirmed). - Run a behavior test, not just a connectivity check. Thinking mode on V4 can change latency and output length. Run a real task through it, not just a ping.
- Check your aggregator config. If you reach DeepSeek through OpenRouter or a similar gateway, verify any pinned model names aren't stale.
The trap to watch for: deepseek-reasoner maps to V4 Flash with thinking, not V4 Pro. If you relied on the reasoner alias for hard reasoning, a straight rename keeps you at Flash-tier quality. You need to explicitly switch to deepseek-v4-pro for the higher tier (byteiota.com, Confirmed).
What Makes V4's API Compatibility Special?
DeepSeek V4 supports both OpenAI ChatCompletions and Anthropic API formats natively. This means you can swap the model name in existing Claude Code, OpenClaw, or OpenCode configurations and often keep everything else the same — the base URL, API key structure, and request format carry over with minimal changes (MorphLLM, Confirmed).
The Anthropic-compatible endpoint lives at api.deepseek.com/anthropic. If your tooling speaks the Anthropic dialect but not OpenAI's, this is your path. This dual-format support is a strategic differentiator — it lowers the switching cost for teams already invested in the Claude ecosystem who want to cut API spend. For a practical guide on running Claude Code with alternative models, see our guide to running Claude Code for free.
What Are V4's Three Reasoning Modes?
V4 introduces a three-tier reasoning system controlled by parameters, not separate model IDs:
| Mode | Setting | When to Use | Token Cost |
|---|---|---|---|
| Non-think | thinking: disabled |
Simple queries, chat, summarization, speed-critical tasks | Lowest |
| Think High | reasoning_effort: "high" (default) |
Most tasks — best balance of quality and cost | Medium |
| Think Max | reasoning_effort: "max" |
Hard reasoning problems, complex coding, maximum accuracy | 3-5x more output tokens than Think High |
Source: DeepSeek API docs, aimadetools.com
Using Think Max on simple tasks just burns tokens and time. Using Non-think on hard reasoning gives you sloppy answers. The key is matching the mode to the task complexity — something that's easy to get wrong when you first migrate from the old separate-model approach.
How to Read AI Model Leak Headlines Responsibly
The DeepSeek V4 GA leak cycle illustrates a broader pattern worth understanding. Here's a framework for evaluating any AI model leak:
- Check the source type. Is it an official changelog, a vendor tech report, or a social media post? DeepSeek publishes changelogs when things actually ship — that's your source of truth, not screenshots.
- Separate vendor-reported from independently verified. The 80.6% SWE-bench score is DeepSeek's own number. Scale's SEAL leaderboard has no V4 entry. Until independent harnesses reproduce it, treat it as a vendor claim.
- Check what's being compared. The "0.2 gap" headline compares V4 Pro to Opus 4.6 (previous gen), not Fable 5 (current gen). Always verify which exact models are in the comparison.
- Benchmark scores measure narrow tasks. A 0.2-point gap on a public test can become a 5-point gap on your own messy codebase with weird edge cases. Test on your actual work before switching.
- Watch the open model space, not just the big labs. The pace of improvement in open-weight models (DeepSeek, Kimi, GLM) is currently faster than in closed models. The first open-weight model to top a coding leaderboard was a signal event, and the trend is accelerating.
For a deeper look at how open-weight models are changing the economics of AI, see our analysis of the compute war between free and paid AI models.
What This Means for You
If you're a developer or small team building with AI in 2026, the practical takeaway is this: the open-weight option is now good enough and cheap enough to be your default for most coding work, with a closed model on standby for the hard cases. DeepSeek V4 Pro at $0.87/M output is the rational pick for bounded coding tasks, document processing, and any workload where you can verify outputs locally before merging. Claude Fable 5 at $50/M output is the pick for long-horizon agentic sessions where a botched run costs more than the token savings.
If you're already using DeepSeek, migrate your API model names before July 24, 2026 — that deadline is real and the fix takes twenty minutes. If you're evaluating whether to switch, run 10 real tasks from your actual work through both models and compare. That tells you more than any leaderboard. And if you want to understand how Claude Fable 5 fits into an agentic workflow, see our guide on building a Claude Fable 5 agent OS for your business.
The broader trend: open models with MIT licenses are closing the gap fast enough that the decision is no longer "open vs closed" but "which workloads route to which model." That's a healthier market for everyone.
FAQ
Q: Is DeepSeek V4 Pro better than Claude Fable 5? A: No — on independent benchmarks, Claude Fable 5 leads V4 Pro by double digits on coding (89.2 vs 69.8 BenchAlign coding average) and agentic tasks (84.6 vs 70.6). However, V4 Pro is the strongest open-weight model available and costs roughly 57x less per output token. The right question is which model fits your workload and budget, not which is universally "better."
Q: What is the SWE-bench Verified score for DeepSeek V4 Pro? A: 80.6% for V4 Pro-Max, as reported in DeepSeek's own technical report (arXiv 2606.19348). This is a vendor-reported number, not an independent evaluation. It places V4 Pro within 0.2 points of Claude Opus 4.6 (80.8%) — the previous-generation model, not Fable 5. Scale's SEAL leaderboard has no V4 entry as of July 2026.
Q: Can I run DeepSeek V4 Pro locally? A: Yes. V4 Pro is MIT-licensed with weights on Hugging Face. At 865GB in full precision, it requires serious hardware (multi-GPU server). V4 Flash (160GB, 13B active) is more tractable on A100 80GB or H100 nodes. Quantized versions work with vLLM, SGLang, and llama.cpp.
Q: When do the old DeepSeek API endpoints stop working?
A: deepseek-chat and deepseek-reasoner are retired on July 24, 2026 at 15:59 UTC. Replace them with deepseek-v4-flash (non-thinking or thinking mode) or deepseek-v4-pro. Only the model string changes — base URL, API key, and request structure stay the same.
Q: Does DeepSeek V4 support the Anthropic API format?
A: Yes. DeepSeek V4's API speaks both OpenAI ChatCompletions and Anthropic formats. The Anthropic-compatible endpoint is at api.deepseek.com/anthropic. This means you can swap the model name in Claude Code or other Anthropic-dialect tools with minimal configuration changes.
Q: Is DeepSeek V4 GA officially announced?
A: As of July 21, 2026, DeepSeek has not published an official GA announcement. Leaked model IDs (deepseek-v4-pro-202606, deepseek-v4-flash-202605) appeared on July 4, and community reporting points to a GA rollout beginning around July 19. All GA-specific capability claims should be treated as Reported, not Confirmed, until DeepSeek publishes official release notes.
Discussion
0 comments