What Is DeepSeek V4?
DeepSeek V4 is DeepSeek’s open-weight (MIT-licensed) mixture-of-experts model family, first released as preview on April 24, 2026, and graduating to official stable release in mid-July 2026. It ships in two variants — V4-Pro and V4-Flash — and remains one of the most aggressively priced frontier-adjacent models on the market, with a 1-million-token context window by default across both tiers.
Key Features
- Two-tier lineup — V4-Pro (1.6T total parameters, 49B active) for maximum capability; V4-Flash (284B total, 13B active) for speed and cost efficiency.
- 1M-token context window — standard across both models, with up to 384K max output tokens.
- Hybrid reasoning modes — thinking and non-thinking modes selectable per request, plus JSON output and tool calling support.
- MIT-licensed weights — both models are on Hugging Face for self-hosting, with no restrictions on commercial use or modification.
- Dual API compatibility — supports both the OpenAI ChatCompletions format and the Anthropic API format, making it a near drop-in swap for existing integrations.
Pricing
V4-Pro: $0.435 input / $0.87 output per million tokens (cache-hit input $0.003625/M). V4-Flash: $0.14 input / $0.28 output per million tokens (cache-hit $0.0028/M). Per output token, V4-Pro runs roughly 28.7x cheaper than Claude Opus 4.8 and 34.5x cheaper than GPT-5.5. The mid-July official release adds peak-hour pricing — 2x baseline during Beijing business hours (9–12 and 14–18) — which wasn’t part of the original preview pricing, so factor that into cost modeling if you’re running high-volume workloads during those windows.
Important: the July 24 migration deadline
If you’ve been calling the legacy deepseek-chat or deepseek-reasoner aliases, they’ve been routing to V4-Flash since April 24 — but those aliases retire entirely on July 24, 2026 at 15:59 UTC, with no announced grace period. Full migration guide here. If your system has been stable on the legacy aliases since April, your effective migration to V4-Flash has arguably already happened behaviorally — you just need to update the model ID in your API calls before the cutoff.
Benchmarks
DeepSeek-V4-Pro-Max scores 80.6% on SWE-bench Verified — the highest score among open-weight models, tied with Gemini 3.1 Pro. Function calling reliability improved sharply over V3, with error rates dropping from roughly 15% to under 2%. Chinese-language capability is reported as notably strong versus GPT-5 and Claude. The trade-off: V4 remains text-only, with no multimodal input support publicly available as of this update.
How It Compares
Against GLM-5.2, DeepSeek V4 competes directly as an open-weight, MIT-licensed option, with V4’s 1M context window and stronger SWE-bench score being the main differentiators. Against proprietary frontier models like GPT-5.6 or Claude Opus 4.8, DeepSeek V4 doesn’t lead on raw reasoning quality, but the price gap is enormous — V4-Flash at $0.14/$0.28 per million tokens makes it viable for high-volume production use where per-token cost dominates the calculus. MiniMax M3 is the other open-weight model worth cross-shopping in this bracket.
Is DeepSeek V4 Worth It?
For cost-sensitive production workloads — especially agentic loops with stable system prompts that benefit from the deep cache discounts — V4-Flash is close to unbeatable on price. V4-Pro is worth the premium when you need meaningfully better reasoning and can still tolerate open-weight-tier quality versus the closed frontier leaders. The one hard action item: if you’re on the legacy API names, migrate before July 24 — it’s a one-line change, but the deadline has no fallback.
Last updated: July 17, 2026.
Related reading
- DeepSeek’s July 24 API deadline — what to migrate and when
- GLM-5.2 review — the other major open-weight contender