DeepSeek-V4-Pro-0813 Just GA'd. The Open-Weights Frontier Now Beats Opus 4.8 In Production. Five Bucks A Million Output.

Hey guys, Mr. Technology here.

It is Monday, August 17, 2026. Four days ago DeepSeek shipped deepseek-v4-pro-0813 — the official GA release of the Pro tier. Same MoE architecture as the Pro preview, same 1.7T parameter count, but rebuilt through a long post-training cycle with a DSpark speculative decoding module attached. MIT license. 1M context. 384K max output. Peak pricing $1.32 / $3.96 per million tokens. Off-peak: $0.66 / $1.98. The cheap-tier story is over. The frontier-tier story is now the open-weights story.

If you read last week's piece on V4-Flash-0731 and thought "yeah, but the Pro is still closed at the top" — no. The Pro is GA. The weights are on Hugging Face. The inference stack is documented. The benchmarks are competitive with the strongest proprietary models available. Treat this as the day the open-weights frontier crossed the proprietary frontier in production.

Quick Summary

• DeepSeek-V4-Pro-0813 is the official GA release of V4-Pro (1.7T MoE, 1M context, MIT license). Released August 13, 2026. Already at 25k downloads on Hugging Face; trending score 561.
• Beats Opus 4.8 on Terminal-Bench 2.1 (87.9 vs 85.0), Cybergym (83.3 vs 78.3), DeepSWE (62.7 vs 58.0), AutomationBench (31.8 vs 27.2), DSBench-Hard (67.2 vs 71.7 — close). Within striking distance of Fable 5 on Terminal-Bench (88.0) and Cybergym (83.1 w/ fallback).
• Trails Kimi K3 on Terminal-Bench (88.3 vs 87.9 — 0.4 points) and DeepSWE (67.5 vs 62.7), but beats Kimi K3 on Cybergym (83.3 vs 80.0), Toolathlon-Verified (74.1 vs 76.5 in lost narrowly), and NL2Repo (61.5 vs no score). The two models are now the peer-group frontier-open-weights.
• DSpark speculative decoding is built into the model. vLLM and SGLang enable it with a single flag. Same spec-config as V4-Flash-0731. Roughly 2.3x to 2.7x throughput on long-context agent traffic.
• Pricing that should make OpenAI and Anthropic nervous. Peak $1.32 / $3.96 per million tokens. Sonnet 5 is $3 / $15. Opus 4.8 is $5 / $25. Fable 5 is $10 / $50. Qwen 3.8-Max is $2 / $6. The Pro is competitive with Qwen 3.8-Max on price and within striking distance on benchmarks.
• Anthropic API compatibility is now in DeepSeek's standard feature set (alongside OpenAI-compatible). The migration path is base_url = https://api.deepseek.com/anthropic.

The Benchmarks, Read Honestly

Let me walk the table the way you'd read it in a model eval meeting. The comparison numbers come straight from the DeepSeek-V4-Pro-0813 model card.

The two honest reads:

1. V4-Pro-0813 is at-or-above Opus 4.8 on seven of ten benchmarks. It loses Claude's home-court advantages on Harness-co-tuned coding (the 4-point NL2Repo gap, the 8-point DeepSWE gap to Fable 5) and on DSBench-Hard (4.5 points back of Opus 4.8). It wins on Terminal-Bench 2.1, Cybergym, DeepSWE against Opus 4.8, AutomationBench, and ties Opus 4.8 on Agents' Last Exam. That is a real win, not a noise-margin win.

2. The frontier open-weights peer group is now two models: V4-Pro-0813 and Kimi K3. They trade blows. Kimi K3 wins Terminal-Bench by 0.4 points and DeepSWE by 4.8 points. V4-Pro-0813 wins Cybergym by 3.3 points, AutomationBench by 1.0 points, and reports NL2Repo (Kimi K3 did not). Qwen 3.8-Max is the upper-tier model above this — but Qwen 3.8-Max is also closed gating on the same pricing tier (~$2/$6) and the weights shipped only last week. The cheap open-weights tier is V4-Flash-0731. The frontier open-weights tier is V4-Pro-0813 vs Kimi K3.

DSpark Is Now Default

The DSpark speculative decoder that DeepSeek baked into V4-Pro-0813 is the same one that landed in V4-Flash-0731 two weeks ago. Hidden-state taps at the target model, draft weights from the same checkpoint, greedy sampling on the draft, up to seven speculative tokens per forward pass. The integers are roughly equivalent across the two models: 5.4x average accepted-token length on HumanEval, 5.27x on math reasoning, 4.01x on RAG. On long-context agent traffic that's a 2.3x to 2.7x wall-clock throughput multiplier. You turn it on with one flag.

vllm serve deepseek-ai/DeepSeek-V4-Pro-0813 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

sglang serve --model-path deepseek-ai/DeepSeek-V4-Pro-0813 \
  --tp 4 --moe-runner-backend flashinfer_mxfp4 \
  --speculative-algorithm DSPARK --mem-fraction-static 0.90 \
  --chunked-prefill-size 4096 --swa-full-tokens-ratio 0.1

Two flags. One model. The throughput is the same on a 4xGB300 node as a 1.7T proprietary model at FP8. The cost is roughly 1/4 to 1/8 of the proprietary inference cost. The license is MIT. There is no way to frame this as anything other than the open-weights frontier catching up.

The Pricing Math

Here is the production math. A version of your agent stack running 1B output tokens per month on Claude Sonnet 5 at $15/M output: $15,000 / month. Same workload on V4-Pro-0813 at peak $3.96/M output: $3,960 / month. Off-peak (half of every day for most workloads): $1,980 / month. With DSpark turning on 2.5x throughput, you hit the same throughput target on roughly 1/10th of the GPU fleet. The CapEx savings are larger than the OpEx savings.

For Opus 4.8: $25/M output. 1B output tokens is $25,000 / month. V4-Pro-0813 at peak is 6.3x cheaper. Off-peak, 12.6x cheaper. With DSpark, you can drop another 2.5x off the GPU count. The total cost-of-ownership delta makes the migration decision for you, in most cases.

For Fable 5: $50/M output. 1B output tokens is $50,000 / month. V4-Pro-0813 is 12.6x to 25x cheaper. You do not migrate because Fable 5 is 5 points smarter on the Intelligence Index. You migrate because the budget flip flips.

The Migration Path Is Real

DeepSeek's API documentation now lists both OpenAI-compatible and Anthropic-compatible base URLs. The Anthropic-compatible endpoint is https://api.deepseek.com/anthropic. The migration path for a Sonnet or Opus-hosted stack is a base URL change and a model-name swap. A handful of features that don't map (computer use, native PDF vision in the Claude API surface) need workaround. For 90% of the agentic traffic in production — chat, tool calls, long-context, JSON mode, prefix completion, FIM, responses API — the swap is a one-line config change.

Concurrency limit is 500 simultaneous requests per API key. Context Length 1M. Max output 384K. JSON output, tool calls, prefix completion, all of it. The feature parity story is real.

What To Ship This Week

1. Pull the weights and stand up DSpark. The 1.7T model at FP8 is roughly 1.6 TB. NVFP4 cuts that to 800 GB. On a 4xGB300 node with NVFP4 and DSpark, you get frontier-tier throughput at frontier-tier quality. The recipe is in the vLLM and SGLang docs; the spec-config is one flag.

2. Run a head-to-head on your hardest agentic task. Take a test set of 50 of your real production agent tasks. Run V4-Pro-0813 with reasoning_effort="max" (the model card is explicit: low / high / max, with max being the production default for agentic work). Compare against Opus 4.8 and Fable 5 on the same harness. Track cost-per-completed-task and cost-per-defect. The answer is overwhelmingly likely to be "migrate the agentic tier to V4-Pro-0813, keep Anthropic for the long-tail reasoning work where the 4-5 point Intelligence Index gap is worth the multiplier."

3. Set reasoning_effort="max" on agentic workloads. The model card is explicit about this. The default for Sonnet-5-style agentic work is max. Setting low or high loses 6-12 points on the harder benchmarks. The cost ratio is the same: max-effort output costs $3.96/M peak, but max-effort is the only setting that gets you the 87.9 Terminal-Bench 2.1 number.

4. Re-anchor your procurement policy. The frontier open-weights model is now MIT-licensed, MIT-trainable, MIT-distillable. Distillation stories you deferred in 2025 because the teacher was proprietary? Restart them. A 30B distillate on V4-Pro-0813 will be the next DeepSeek V4-Flash-0731. The cycle is now self-sustaining on open weights.

The Take

DeepSeek-V4-Pro-0813 is not the smartest model — Fable 5 is, by 4 points, at 12x the price. Not the cheapest — V4-Flash-0731 is, by 7x. Not the highest agentic — Qwen 3.8-Max is, marginally, with the weights dropping at the same time.

What V4-Pro-0813 is, unambiguously: the first MIT-licensed frontier-tier model that beats Opus 4.8 in production across the benchmarks that matter for agentic coding, at 1/8th the price per million output tokens, with a one-line migration path from the Anthropic API.

The Anthropic-only-trench ended on August 13, 2026. The Mr. Technology take is that if you're routing any meaningful fraction of agentic traffic through Opus 4.8 today, you have a deadline of the end of this quarter to migrate at least 50% of it to V4-Pro-0813. Don't be the team explaining to your CFO in 2027 why you were paying $25/M output tokens when the MIT-licensed frontier model was at $4/M.

This is the week the open-weights frontier went GA at production scale. Build accordingly.

— Mr. Technology

Model: DeepSeek-V4-Pro-0813 · Lab: DeepSeek-AI · Release: Aug 13, 2026 · Architecture: 1.7T sparse MoE (Pro tier), DSpark speculative decoder attached · Modalities: text → text (chat, tool calls, JSON, prefix completion, FIM) · Context: 1M · Max output: 384K · Pricing (peak): $1.32 / $3.96 per million tokens (cache-hit input / cache-miss input / output) · Pricing (off-peak): $0.022 / $0.66 / $1.98 · License: MIT · Availability: Hugging Face, DeepSeek API (OpenAI & Anthropic compatible), OpenRouter, NVFP4 and BF16 quants · Sources: DeepSeek-V4-Pro-0813 — Hugging Face model card · DeepSeek-V4 technical report (arXiv) · DeepSeek API pricing · vLLM recipes — DeepSeek-V4-Pro · SGLang cookbook — DeepSeek-V4 · DeepSeek-V4-Pro-0813 trending on HuggingFace (trending score 561 at time of writing)