DeepSeek in Claude Code (and the Engine They Gave Away)
DeepSeek open-sourced DeepInfer, the engine that runs its own API. Three days earlier it shipped V4.1-Flash-1017: MIT open weights, a 1M-token context, and $0.15 per million input tokens. A heavy day of AI coding costs about $2.10 with it, against $30 with Gemini 4 Argon. What you'll learn: → What DeepSeek released on 17 and 20 October, and why shipping the serving engine is rare → The model, honestly: DeepSeek's own benchmark gains, and where it sits against Claude Opus 5.5, Kimi K3.1 Pro and GPT-6.1 Terra → Inside DeepInfer: kernels, expert parallelism, DSpark speculative decoding and deterministic sampling → What it does NOT include: KV cache, prefill/decode disaggregation and load balancing → How to point Claude Code at deepseek-flash with three environment variables → The cost maths for 10M input + 1M output tokens, and DeepSeek's peak hours → Four things to check before you move your team over Honest trade-offs: it is not a frontier model on Terminal-Bench, the API runs on DeepSeek's servers, and DeepInfer needs data-centre GPUs. I cover all of it. ⏱️ Timestamps 0:00 – $30 or $2.10 for the same day of coding 0:30 – What happened 1:26 – The model, honestly 2:34 – Inside DeepInfer 3:57 – Inside Claude Code 5:28 – The cost maths 6:39 – Before you switch 7:40 – Verdict Use deepseek-flash in Claude Code export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic export ANTHROPIC_API_KEY=YOUR_DEEPSEEK_API_KEY export ANTHROPIC_MODEL=deepseek-flash claude The first two lines come from DeepSeek's Anthropic API guide. deepseek-flash is the model name on DeepSeek's pricing page. The terminal session in the video is an illustrative recreation, not a recorded run: your turns, results and cost will differ. Numbers used (checked 27 Oct 2026) • V4.1-Flash-1017: MIT, 763B parameters, 1M context. Terminal-Bench 4.0 is 42.3 (was 31.2), ProgramBench 29.1 (was 20.3), DeepSWE v1.1 77.0 (was 74.2). These are DeepSeek-reported. • Price per 1M tokens, off-peak: $0.15 input, $0.003 cache hit, $0.60 output. Prices double at peak, 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. • Search & Fetch APIs: $0.75 per 1,000 calls. They are free through server-side tools in DeepSeek Harness, Claude Code and Codex until 31 Oct 2026, 16:00 UTC. • Cost example (10M input + 1M output, list prices, no caching): DeepSeek $2.10 · GPT-6.1 Terra $15 · Gemini 4 Argon $30 (introductory price to 30 Nov). • Terminal-Bench 4.0 across labs: Claude Opus 5.5 66.4 · Kimi K3.1 Pro 65.7 · GPT-6.1 Terra 50.3 · DeepSeek 42.3. These are vendor-reported, and each lab uses its own harness. • Cloudflare, 26 Oct: an SGLang bug in disaggregated serving leaked 155 responses on prompts over 681,280 tokens. The fix is SGLang 0.5.22 or later. Sources DeepSeek release note: https://api-docs.deepseek.com/news/news261017/ Model card: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 DeepInfer: https://github.com/deepseek-ai/DeepInfer DeepSeek pricing: https://api-docs.deepseek.com/quick_start/pricing DeepSeek Anthropic API guide: https://api-docs.deepseek.com/guides/anthropic_api GPT-6.1 Terra: https://openai.com/index/introducing-gpt-6-1-terra/ Gemini 4 Argon: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ Kimi K3.1 Pro: https://www.kimi.ai/blog/kimi-k3-1-pro MiniMax M3.1 Flash: https://www.minimax.io/blog/minimax-m31f Meta Muse Spark Open: https://research.meta.ai/blog/muse-spark-open Cloudflare postmortem: https://blog.cloudflare.com/kv-cache-exposure-remediation/ Hacker News thread: https://news.ycombinator.com/item?id=50163277 What did it cost you per task? Tell me in the comments. Follow @MervinPraison for more AI developer breakdowns. #DeepSeek #ClaudeCode #AIcoding #OpenSource #LLM
