DeepSeek Cuts Prices in Half Mid-Week: This Week in LLM Pricing

The big story this week is straightforward: DeepSeek halved its prices on two models mid-week. If you're routing any significant volume through either of them, this is worth paying attention to.

The DeepSeek Price Cuts

Our tracker caught two snapshots this week for deepseek-v4-1-flash and deepseek-v4-pro.

deepseek-v4-1-flash

• Oct 5: $0.30 input / $1.20 output per million tokens
• Oct 10: $0.15 input / $0.60 output per million tokens

deepseek-v4-pro

• Oct 5: $1.32 input / $3.96 output per million tokens
• Oct 10: $0.66 input / $1.98 output per million tokens

That's a clean 50% reduction across the board for both models, applied between October 5th and October 10th. Not a rounding artifact — both models dropped by exactly half on both input and output.

What does this mean practically? If you've been holding off on deepseek-v4-pro because the cost felt hard to justify against something like GPT-4o or Gemini 1.5 Pro, the math changes now. At $0.66/$1.98, it sits in a much more competitive band for mid-tier inference tasks. And deepseek-v4-1-flash at $0.15 input is genuinely cheap for high-volume, lower-stakes work — summarization, classification, first-pass drafts.

DeepSeek has form here. They've used aggressive pricing as a market share play before, and this follows the same pattern. Whether the prices stay here or drop further is hard to predict, but right now they represent real value if the model quality holds up for your use case.

New Models Spotted

Three new model IDs appeared in our tracker this week.

anthropic/claude-haiku-5.5 and anthropic/claude-haiku-5.5:batch both showed up on October 8th. The batch variant follows Anthropic's established pattern of offering asynchronous batch processing at reduced rates. Haiku has always been Anthropic's speed-and-cost tier, and a 5.5 version suggests incremental improvements over whatever 5.x came before. No pricing anomalies to flag on these yet — we'll track them going forward.

google/gemini-nano-banana-2.1 appeared on October 7th from Google. The "nano" and "banana" naming suggests this is either an experimental model or a specialized lightweight variant — Google has used codename-style identifiers before for models not yet in full public release. Worth watching but nothing concrete to say about it yet beyond the fact that it exists in the API.

The Practical Takeaway

If you have a cost-sensitive workload and you haven't benchmarked DeepSeek V4 Flash recently, now's a reasonable time to do it. At $0.15 per million input tokens, the failure cost of running a quick eval is minimal. For higher-capability needs, V4 Pro at $0.66 input puts it in a range where it's worth comparing directly against your current go-to.

The new Anthropic and Google models are worth keeping an eye on as they mature — particularly Claude Haiku 5.5 if you're already in the Anthropic ecosystem and care about the cost/performance tradeoff at the smaller end.

As always, prices on third-party routers like OpenRouter can lag official announcements or differ slightly from direct API pricing, so verify before committing to anything at scale.

---

I track these changes weekly over at LLM Price Watch, which checks live pricing across Claude, GPT, Gemini, DeepSeek, and Grok daily. Worth bookmarking if you're making model selection decisions based on cost.