DeepSeek V4 Pro and V4 Flash: 1.6T MoE, 1M Context, Serving Benchmarks, Pricing (2026 Guide)

TL;DR

Last updated September 1, 2026 with the InferenceX V4 Pro serving runs, the self-host cost math, weekday-only peak billing, V4-Flash-Vision-Exp, and current OpenRouter rates.

DeepSeek V4 Pro is a 1.6T-parameter mixture-of-experts model with 49B active per token. DeepSeek V4 Flash is its 284B sibling with 13B active. Both are open weight under MIT, both default to a 1M-token context with 384K max output, and both are GA: V4-Flash-0731 since July 31, 2026 and V4-Pro-0813 since August 13. Off-peak API rates are $0.66/M input and $1.98/M output for Pro, $0.22/M and $0.66/M for Flash, double at weekday peak. Artificial Analysis scores Pro-0813 at 53 on its Intelligence Index, 1 point above Flash-0731. Per output token off-peak, V4-Pro is 12.6x cheaper than Claude Opus 4.8 and 15.2x cheaper than GPT-5.5.

Serving is where Pro gets expensive. InferenceX measures 216 tok/s for one user on a B300 node and 40 tok/s per user at concurrency 64 on B200. An 8x B200 node rents for $53.52 an hour on Lambda, about $0.12 per million tokens at full utilization and $0.40 at 30 percent. Morph serves V4 Flash at $0.1234375/M input and $0.3475/M output with no node to keep busy.

What Is DeepSeek V4?

DeepSeek V4 is an open-weight mixture-of-experts (MoE) model family from DeepSeek under the MIT license, previewed April 24, 2026 and now GA. It ships in two variants: V4-Pro (1.6T total parameters, 49B active per token; current checkpoint V4-Pro-0813) and V4-Flash (284B total, 13B active; current checkpoint V4-Flash-0731). Both default to a 1M-token context window with 384K max output, and the weights are on Hugging Face. The API speaks both the OpenAI ChatCompletions and Anthropic formats, so it drops into tools like Claude Code and OpenCode without a proxy.

Most serverless hosts quantize V4 activations to fp8 to cut cost, which moves output away from the reference weights. Morph serves morph-dsv4flash (DeepSeek V4 Flash) at 16-bit bf16 with no fp8 activation quantization, so output matches the published weights, at $0.1234375/M input and $0.3475/M output. See Morph Open Source Models and pricing.

Release Date and Models

DeepSeek released V4 on April 24, 2026 as a preview, announced on its API docs news page. Two models shipped the same day, and each has since rolled to a GA checkpoint:

• deepseek-v4-pro: 1.6T total parameters, 49B active per token. Serves DeepSeek-V4-Pro-0813 since August 13, 2026.
• deepseek-v4-flash: 284B total parameters, 13B active per token. Serves DeepSeek-V4-Flash-0731 since July 31, 2026.

Both run with a 1M-token context window by default across all official DeepSeek services, support JSON output, tool calls, and thinking plus non-thinking modes with three reasoning effort levels (low, high, max), and expose 384K max output tokens. The legacy deepseek-chat and deepseek-reasoner endpoints retired on July 24, 2026 15:59 UTC. The API supports both the OpenAI ChatCompletions format and the Anthropic API format, which is what makes the Claude Code setup below a three-line config.

Weights are on Hugging Face (deepseek-ai/DeepSeek-V4-Pro-0813, deepseek-ai/DeepSeek-V4-Flash-0731), released under the MIT license per the model cards.

What Changed August 15 to September 1, 2026

Everything below is verified against DeepSeek's API docs, changelog, and the cited third-party pages on September 1, 2026.

V4 Pro vs V4 Flash: Specs and Pricing

Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. The other 17 weekday hours and all of Saturday and Sunday bill off-peak at half the peak rate. The cache-hit input prices remain the standout line items. A cache hit on V4-Pro costs $0.022/M off-peak, 30x less than its cache-miss rate. Agentic coding loops that resend the same system prompt and file context on every turn see most of their input tokens land in cache, so effective per-session cost runs far below the list cache-miss rate.

Which to pick: Pro activates 3.8x more parameters per token. Flash costs 3x less on every line item, and Artificial Analysis scores Pro-0813 only 1 point above Flash-0731 (53 vs 52 on its Intelligence Index). Flash is the default for batch pipelines, evals, coding, and high-volume extraction; reach for Pro when a hard reasoning task measurably needs it.

DeepSeek V4 variant matrix: Pro vs Pro-Max vs Flash

Three names circulate, which causes confusion. V4-Pro is the served 1.6T MoE. V4-Pro-Max is the benchmark configuration of Pro (the entry that posts the 80.6% SWE-bench Verified score); DeepSeek does not list a separate Pro-Max API price, so the output rate below is Pro's. V4-Flash is the smaller, cheaper variant.

Read the DeepSWE column carefully. The 8% pass@1 figure came from the independent DeepSWE tracker running the April preview. The 62.7 (Pro-0813) and 54.4 (Flash-0731) figures are DeepSeek's own, from the GA model cards, with no independent reproduction yet. The reconciliation section below covers the preview-era spread.

DeepSeek V4 Pro Serving Benchmarks

Benchmark scores tell you what V4 Pro can do. Serving numbers tell you what it costs to get that out of a GPU. V4 Pro's FP4 weights alone are about 0.8 TB (1.6T parameters at 4 bits), so every public run uses an 8-GPU Blackwell node or a multi-node rack. The independent data comes from SemiAnalysis InferenceX, which publishes the exact scripts and GitHub Actions runs behind each number.

InferenceX runs of V4-Pro-0813 (August 25, 2026)

Read the two rows as the ends of one curve. One user on a B300 node gets 216 tok/s with a 350 ms first token because the whole node works on one request. At concurrency 64 on B200, each user gets 40 tok/s and waits 4.25 s for the first token because 64 prefills of 8K tokens each queue for the same GPUs. Aggregate throughput is what you pay for; per-user speed is what your agent feels.

The B200 vs B300 frontier at three interactivity targets

InferenceX's comparison page sweeps concurrency and reports the best measured throughput per chip at fixed per-user speeds for V4-Pro-0813 in FP4 on the 8K/1K lane. The cost column is InferenceX's own, from its hardware cost model, not a cloud list price.

The interactivity tax is steep: going from 59 to 145 tok/s per user cuts B200 throughput per chip by 4.6x. B300's extra HBM matters most at the fast end, where the decode batch is small and the run is memory-bound.

Rack scale and other silicon

• GB300 NVL72, disaggregated prefill/decode (Dynamo + SGLang): about 11,200 tok/s per GPU at roughly 50 tok/s per user with MTP speculative decoding in June 2026, up from about 2,200 tok/s per GPU on the day-0 April stack, per the SGLang team's PyTorch blog post on the public InferenceX GB300 lane.
• AMD MI355X, single 8-GPU node, TP8 (SGLang): 2,256 tok/s per GPU on the 8K/1K lane by May 21, 2026, up 110.5x from the FP8-only day-0 number, per InferenceX.
• Public API providers (Artificial Analysis medians): Pro-0813 streams 54.1 output tok/s with a 1.65 s first token; Flash-0731 streams 107.9 tok/s with a 1.50 s first token. Flash is 2x faster to a user for the same reason it is 3x cheaper: 13B active parameters versus 49B.

Filter the attributed runs below, or see every model on the dedicated inference benchmarks page. For how these numbers translate into GPU counts, see LLM inference explained.

Self-Host DeepSeek V4 Pro vs Call V4 Flash on Morph

The self-host question has a fixed cost and a utilization problem. Lambda lists an 8x B200 SXM6 node at $6.69 per GPU-hour on demand, so one node costs $53.52 an hour, or about $39070 a month if it never sleeps. Divide that by the tokens InferenceX measured on B200 and you get the cost per million tokens at each interactivity target:

The math: $6.69 per GPU-hour divided by (tok/s per chip times 3,600 seconds), per million tokens. InferenceX counts the tokens on its 8K/1K lane the same way for its own $/M column; its lower figures come from a hardware cost model rather than a cloud list price. Thirty percent utilization is generous for a single-team deployment. Coding agents are bursty and a node bought for weekday peaks idles overnight.

At full utilization a rented B200 node serves V4 Pro for about the same per-token price Morph charges for V4 Flash. At 30% utilization the node costs 3.4x more per token, and 15x more if your agents need 145 tok/s per user. You also still owe the engineering time that took SGLang from 2,200 to 11,200 tok/s per GPU between April and June. Morph's cache-read rate is $0.031/M, which is what agent loops that resend the same context actually pay on most input tokens. Pro's 1-point Intelligence Index lead over Flash is the whole quality delta you would be buying.

If you need Pro specifically, or a private endpoint, the honest path is a dedicated deployment sized from measured throughput rather than list prices. Use the dedicated inference calculator to size it, or read how Morph dedicated inference prices GPU-hours against the tokens they actually produce.

DeepSeek V4 Architecture

Per DeepSeek's release notes, V4's attention combines token-wise compression with DSA (DeepSeek Sparse Attention). DeepSeek's own documentation is sparse on internals; the detail below comes from the model card and third-party technical analyses of the weights, and is marked as such.

Attention: token-wise compression + DSA

The official description: each layer compresses the KV cache token-wise, then applies DeepSeek Sparse Attention over the compressed representation. This is what makes a 1M-token default context economical to serve at $0.66/M off-peak input.

Third-party analyses (Lambda's launch breakdown and model-card readers) describe the mechanism as Compressed Sparse Attention: KV caches compressed 4x along the sequence dimension, with a lightning indexer selecting the top 1,024 compressed KV entries per query. Treat the 4x and top-1,024 figures as secondary-source; DeepSeek's news page does not publish them.

CSA and HCA: the hybrid that makes 1M context cheap

The Hugging Face DeepSeek V4 blog names the attention stack a hybrid of two mechanisms interleaved across layers, not a single "DSA" block (DSA is the V3.2-era term). The two are Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).

• CSA compresses KV entries 4x along the sequence dimension using softmax-gated pooling with a learned positional bias. A lightning indexer (run in FP4, a ReLU-scored multi-head dot product) selects the top-k compressed blocks per query, and a sliding-window branch handles the most recent uncompressed tokens.
• HCA compresses KV entries 128x and drops sparse selection entirely: every query attends densely to every compressed block.

Per the HF blog, V4-Pro is a 61-layer stack where layers 0 and 1 are HCA and layers 2 through 60 alternate CSA and HCA. Both paths store most KV entries in FP8 and keep BF16 only for the RoPE dimensions. The compound effect: the KV cache lands at roughly 2% the size of 8-head grouped-query attention in bfloat16 (V4-Pro at about 10% of V3.2's KV memory, V4-Flash at about 7%), and single-token inference runs at 27% of V3.2's FLOPs for Pro and 10% for Flash. That is what makes a 1M-token default context economical even at $0.66/M off-peak input.

DeepSeek V4 Pro architecture

V4-Pro is the 1.6T-total-parameter MoE with 49B active per token. Versus V3 (671B total, 37B active), that is 2.4x the total parameter count and 1.3x the active compute per token, with the context window expanded 8x from 128K to 1M. The sparse-attention stack is what keeps the larger model servable: DeepSeek prices Pro at $1.98/M output off-peak, 3x the output price of V4-Flash but a fraction of closed-model rates.

DeepSeek V4 Flash architecture

V4-Flash shares the attention design but at 284B total parameters with 13B active per token. The smaller expert pool is why DeepSeek can price Flash at $0.22/M off-peak input, a third of Pro on every line item. It also inherits the same 1M context and 384K max output, so the variant choice is about quality per token, not context capability.

Benchmarks: SWE-bench Verified

The independently tracked number is DeepSeek-V4-Pro-Max at 80.6% on SWE-bench Verified (llm-stats tracker, June 2026). That is the highest open-weights entry, tied with Gemini 3.1 Pro, 0.1 points ahead of MiniMax M3 and 0.2 ahead of Qwen3.7 Max. Closed frontier models score higher: Claude Fable 5 leads at 95.0% (currently suspended, see note).

Read the price column against the score column. The 14.4-point gap between Fable 5 (95.0%) and V4-Pro-Max (80.6%) costs 25x more per output token at V4-Pro's off-peak rate. The 8-point gap to Opus 4.8 costs 12.6x more. Whether that trade is worth it depends entirely on whether your tasks live in the band those extra points unlock.

DeepSeek-reported launch numbers

At launch, DeepSeek reported V4-Pro-Max scoring 93.5 on LiveCodeBench Pass@1 and a 3206 Codeforces rating. These are vendor-run numbers from the April 2026 release coverage, not independent leaderboard entries; vendor scaffolds routinely score above standardized harnesses (see our SWE-bench Pro breakdown for how large that gap runs).

GA checkpoint benchmarks (self-reported, August 2026)

The GA checkpoints post-train heavily for agentic work, and the model-card numbers move accordingly. Flash-0731 vs the April preview: Terminal-Bench 2.1 61.8 to 82.7, DeepSWE 7.3 to 54.4, Toolathlon-Verified 49.7 to 70.3. Pro-0813 vs the April Pro preview: Terminal-Bench 2.1 72.1 to 87.9, DeepSWE 12.8 to 62.7, NL2Repo 38.5 to 61.5, CyberGym 52.7 to 83.3. Every cell is DeepSeek-run; no independent reproduction of a GA-specific figure exists yet. The independent signal so far is Artificial Analysis: Pro-0813 scores 53 on its Intelligence Index, 8 points above the April Pro but only 1 point above Flash-0731's 52.

Does DeepSeek V4 Really Score 80.6%? Self-Reported vs Independent

The 80.6% SWE-bench Verified number is real, but it sits at the top of a harness with a loose verifier. On DeepSWE, a written-from-scratch, contamination-free benchmark, the April V4-Pro preview scored 8% pass@1 versus GPT-5.5 at 70% and Opus 4.7 at 54%. The same ordering held on SWE-bench Pro. Read all three together before trusting any one. Note the checkpoint scoping: everything in this section measured the April preview. DeepSeek self-reports 62.7 on DeepSWE for Pro-0813, but on its own harness; the independent trackers have not re-run the GA build.

How to reconcile the spread. The yage.ai DeepSWE audit sampled tasks across both benchmarks and an external LLM judge, and found SWE-bench Pro's verifier rejects about 24% of functionally correct solutions and accepts about 8.5% of incorrect ones, while DeepSWE's verifier runs at 1.1% false negatives and 0.3% false positives. A tighter verifier and contamination-free tasks make DeepSWE harder to inflate.

That said, V4-Pro's 8% DeepSWE result is anomalously low relative to peers and to its own 76.2% on SWE-bench Pro. The audit's read is that the gap reflects a real long-horizon-agent capability difference rather than a verifier artifact, because V4-Pro ranks lowest on both independent harnesses. The SWE-bench Pro figures above are the ones reported in the yage.ai audit, not Scale's SEAL board (which still has no V4 entry); treat the Opus 4.7 and GPT-5.5 numbers as reported by the DeepSWE and audit trackers, not independently re-run by us.

DeepSeek V4 Flash and SWE-bench: What Exists

As of July 2026, Flash has aggregated and self-reported SWE-bench numbers but still no contamination-free independent re-run: llm-stats lists DeepSeek-V4-Flash-Max at 79.0% SWE-bench Verified, DeepSeek's own technical report says 73.7%, and our DeepSeek V4 Flash deep-dive covers the gap between those two figures. The state of the trackers:

• Scale's SEAL SWE-bench Pro leaderboards (public and private sets) list no DeepSeek V4 entry of any variant. The top open-weights entry there is qwen3-coder-480b-a35b at 38.7% on the public set.
• The llm-stats SWE-bench Verified tracker lists DeepSeek-V4-Pro-Max (80.6%) but no Flash entry.

If you see a Flash SWE-bench number quoted, check whether it is a vendor-run scaffold result or a community harness; neither currently appears on the two trackers above. For agentic coding where benchmark evidence exists, the published data points at Pro, not Flash.

Cost Math vs Opus 4.8, GPT-5.5, Gemini 3.1 Pro

Reddit and X discussion of V4 settled on a "17x cheaper" shorthand at launch. The August 16, 2026 repricing shrank the ratios: increases ran 50% to 1,100% across line items. The exact ratios now depend on which token type you compare and on the hour. At off-peak list prices:

The launch-era "17x cheaper" shorthand no longer holds. Against GPT-5.4 and Sonnet 4.6 output, V4-Pro off-peak is now 7.6x cheaper; at peak, 3.8x. Against Opus 4.8 it is 12.6x off-peak on output. Against Fable 5 it is 25.3x. Note MiniMax M3 output ($1.20/M) now undercuts V4-Pro at every hour. Cached input still widens the frontier gaps: a V4-Pro cache hit costs $0.022/M off-peak vs $0.50/M for an Opus 4.8 cache hit, 23x apart.

A concrete daily workload, 20 requests of 50K input + 10K output (1M input, 200K output per day), assuming zero cache hits and off-peak hours:

• V4-Flash: $0.35/day, about $11/month
• V4-Pro: $1.06/day, about $32/month
• Opus 4.8: $10.00/day, about $300/month
• GPT-5.5: $11.00/day, about $330/month
• Claude Fable 5: $20.00/day, about $600/month

Full Anthropic-side rates, including cache and batch multipliers, are in our Anthropic API pricing guide.

Where to Run It: First-Party API vs OpenRouter

The first-party API (api.deepseek.com) is the reference deployment: $0.66/$1.98 off-peak for Pro and $0.22/$0.66 for Flash (double at peak), with the cache-hit discounts above. On September 1, 2026 OpenRouter listed deepseek/deepseek-v4-pro from $0.87/M input and $1.74/M output with 1M context across 17 providers (DigitalOcean cheapest, CoreWeave fastest at 50 tok/s), and deepseek/deepseek-v4-flash from $0.0679/M input and $0.168/M output. Per-host rates vary by routing tier, so check the live OpenRouter provider list before committing volume. The image-input variant deepseek-v4-flash-vision-exp (released August 21, 2026) is first-party only and bills at Flash rates.

Reasons to pick first-party: documented cache-hit pricing at $0.022/M off-peak (Pro) and the Anthropic-format endpoint. Reasons to pick OpenRouter: one key across models, provider failover, no peak-hour surcharge, and easy A/B against MiniMax M3 or Qwen3.5 at the prices in the table above. Because the weights are open, you can also self-host; Flash at 284B total is the realistic target, Pro at 1.6T is multi-node territory.

Output fidelity is where serverless hosts diverge. Most serverless providers quantize activations to fp8 to cut serving cost, which moves output away from the reference weights. Morph Open Source Models serves DeepSeek with 16-bit (bf16) activations and does not quantize activations to fp8, so output matches the published weights. For coding agents specifically, Morph adds codegen-tuned speculative decoding plus custom low-level inference kernels built for code generation, which makes it the fastest and highest-quality option for codegen. morph-dsv4flash (DeepSeek V4 Flash) runs at $0.1234375/M input and $0.3475/M output; see pricing for the full list.

Use DeepSeek V4 in Claude Code and OpenCode

Claude Code

DeepSeek's API speaks the Anthropic format natively, so Claude Code needs only environment variables, no proxy:

export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-your-deepseek-key"
export ANTHROPIC_MODEL="deepseek-v4-pro"
export ANTHROPIC_SMALL_FAST_MODEL="deepseek-v4-flash"
claude

Pointing the small-fast model at deepseek-v4-flash keeps background tasks on the $0.66/M-off-peak-output tier. For routing multiple providers behind one endpoint instead, see Claude Code with LiteLLM.

OpenCode

OpenCode ships a DeepSeek provider. Run opencode auth login, select DeepSeek, paste your API key, then set the model in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "deepseek/deepseek-v4-pro"
}

Codex CLI

Codex's custom model_providers config only supports wire_api = "responses", and DeepSeek's API exposes ChatCompletions and Anthropic formats, not the Responses API. To drive V4 from Codex, route through a Responses-compatible proxy or use OpenRouter-backed tooling; for direct use, Claude Code and OpenCode are the paths that work without translation.

What Changed from DeepSeek V3

The structural shift is the attention stack: replacing MLA with token-wise compression plus DSA is what moves the default context from 128K to 1M without 1M-context pricing. The legacy V3-era endpoints survived only as aliases onto deepseek-v4-flash and retired on July 24, 2026.

Where DeepSeek R1 sits: R1 (671B total, 37B active, 128K context, MIT) launched January 20, 2025 as the model behind deepseek-reasoner. Per DeepSeek's changelog that endpoint moved to R1-0528 on May 28, 2025, then to the thinking modes of V3.1 (August 21, 2025), V3.2-Exp, and V3.2, then aliased to V4-Flash thinking mode on April 24, 2026 before retiring. V4 has no separate reasoning model. Both V4-Pro and V4-Flash expose thinking mode with low, high, and max reasoning effort, and R1's 49.2% on SWE-bench Verified sits 31 points under V4-Pro-Max's 80.6%. R1 remains on Hugging Face for self-hosting and its license explicitly permits distillation.

Limitations

Frequently Asked Questions

When was DeepSeek V4 released?

April 24, 2026 as a preview, with V4-Pro and V4-Flash shipping the same day. GA came in two waves: V4-Flash-0731 on July 31, 2026 and V4-Pro-0813 on August 13, 2026. The legacy deepseek-chat and deepseek-reasoner endpoints retired on July 24, 2026 15:59 UTC.

What is the DeepSeek V4 architecture?

A mixture-of-experts transformer with token-wise compression plus DeepSeek Sparse Attention (DSA). Pro: 1.6T total / 49B active. Flash: 284B total / 13B active. Both: 1M context, 384K max output. Third-party analyses add 4x KV compression and a top-1,024 lightning indexer, which DeepSeek has not published first-party.

What does DeepSeek V4 cost?

Since August 16, 2026 the official API bills peak and off-peak. Off-peak: Flash $0.22/M input (miss), $0.007/M (cache hit), $0.66/M output; Pro $0.66/M, $0.022/M, $1.98/M. Peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday through Friday) bill double; weekends are off-peak all day. On September 1, 2026 OpenRouter listed Pro from $0.87/$1.74 and Flash from $0.0679/$0.168.

How fast does DeepSeek V4 Pro run on B200 and B300?

InferenceX runs of V4-Pro-0813 on SGLang in FP4 (8K input, 1K output) measure 216 tok/s for a single user on a B300 node with a 350 ms first token, and 40 tok/s per user at concurrency 64 on B200 with a 4.25 s first token. At 59 tok/s per user the InferenceX frontier is 15,410 tok/s per chip on B200 and 18,735 on B300. Public API providers stream Pro at a median 54.1 tok/s (Artificial Analysis).

What does it cost to self-host DeepSeek V4 Pro?

An 8x B200 node on Lambda is $53.52 an hour, about $39070 a month. At InferenceX's 15,410 tok/s per chip that is $0.12 per million tokens at 100% utilization and $0.40 at 30%. Calling V4 Flash on Morph is $0.148/M blended on the same 8:1 mix with no fixed cost. Full table in the self-host section above.

Where does DeepSeek R1 sit versus V4?

R1 was the January 2025 model behind deepseek-reasoner. That endpoint walked through R1-0528, V3.1, V3.2, and a V4-Flash alias before retiring on July 24, 2026. V4 folds reasoning into both Pro and Flash as an effort parameter (low, high, max). R1 scored 49.2% on SWE-bench Verified; V4-Pro-Max scores 80.6%.

What is deepseek-v4-flash's SWE-bench score?

llm-stats lists Flash-Max at 79.0% SWE-bench Verified; DeepSeek's technical report says 73.7%. Neither is an independent contamination-free re-run, and Scale SEAL still has no Flash entry. Full breakdown on the DeepSeek V4 Flash page.

Does DeepSeek V4 really score 80.6% on SWE-bench?

On SWE-bench Verified, yes: V4-Pro-Max posts 80.6% (llm-stats), but that harness has a loose verifier. On DeepSWE, a contamination-free written-from-scratch benchmark, the April V4-Pro preview scored 8% pass@1 versus GPT-5.5 at 70% and Opus 4.7 at 54%. DeepSeek self-reports 62.7 on DeepSWE for the Pro-0813 GA build, on its own harness, with no independent re-run yet.

How does V4-Pro-Max compare to Claude on SWE-bench Verified?

V4-Pro-Max 80.6% vs Opus 4.6 80.8%, Opus 4.8 88.6%, Fable 5 95.0% (llm-stats, June 2026). Per output token V4-Pro costs $1.98 off-peak ($3.96 peak) vs $25 for Opus 4.8 and $50 for Fable 5.

Is V4 open source?

Open weights on Hugging Face under MIT. The current GA checkpoints are deepseek-ai/DeepSeek-V4-Flash-0731 and deepseek-ai/DeepSeek-V4-Pro-0813. Download, run, fine-tune.

Can I use V4 in Claude Code?

Yes. Set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic, ANTHROPIC_AUTH_TOKEN to your DeepSeek key, and ANTHROPIC_MODEL=deepseek-v4-pro. The full snippet is in the setup section above.