Yesterday I published an agent cost-control guide. Today the market moved. On October 21, DeepSeek and Alibaba led a broad price-cut wave — seven models saw their input prices drop 5–33% in a single day, tracked by olud.ai. If your agent has a model hardcoded anywhere, today changed your bill.
Here are the numbers, then what I'd actually do about them.
The price moves (Oct 21, input per 1M tokens)
| Model | Old | New | Change |
|---|---|---|---|
| DeepSeek V4 Pro 0423 | $0.37 | $0.35 | ▼5% |
| Qwen3.6 35B A3B | $0.15 | $0.10 | ▼33% |
| Qwen3.5-35B-A3B | $0.15 | $0.14 | ▼7% |
| Qwen3.5 397B A17B | $0.50 | $0.39 | ▼22% |
| DeepSeek V3 | $0.32 | $0.26 | ▼19% |
| Gemma 3 4B | $0.05 | $0.04 | ▼20% |
| Qwen3 Next 80B A3B Instruct | $0.09 | $0.10 | ▲11% |
| Solar Pro 4 | $0.09 | $0.15 | ▲67% |
That's not the full story. Two separate announcements landed the same day and they're arguably more useful to a working dev:
DeepSeek also shipped V4.1-Flash at $0.035 input / $0.12 output (per technode's Oct 21 reporting), with off-peak API rates at half of peak. Their Search+Fetch APIs are priced at 5 yuan per 1,000 successful requests — and they're free through Claude Code and Codex integrations until billing starts on November 1. That's eleven days away.
Inception launched Mercury 3 at promo pricing: $0.10 input / $0.35 output, half the standard $0.20/$0.70, with cached input at $0.01 per 1M tokens during the promo (per Inception's launch page and AI News reporting). Promo, not permanent — remember that distinction.
What it actually means
The price war is now a buyer's market for agent builders. I've been watching this space for a while and the pattern's getting familiar: the big labs cut input prices to buy market share, and the cuts keep compounding. The interesting moves this week aren't the small ones. Qwen3.6 at a flat dime is genuinely cheap for a 35B model — that's "use it for everything that doesn't need frontier reasoning" territory. DeepSeek V3 at $0.26 keeps it as the default workhorse for agentic loops where you're burning tokens on retries. And Gemma 3 4B at $0.04 is now basically free for cheap background tasks — classification, routing, spam filtering, the stuff you shouldn't be sending to a $2 model anyway.
One honest note before I keep going: I didn't run benchmarks on any of this. These are tracked provider prices from olud.ai's October 21 snapshot plus the launch announcements above. I'm writing about the money, not claiming I fine-tuned anything.
What I'd change this week
1. Re-price your agent's model routing table.
Seriously — open the file, the env var, the config, wherever your model choices live, and run yesterday's numbers against today's. The cheapest-capable default for any given task tier probably moved this week. If your "budget tier" model was Qwen3 Next 80B at $0.09, it's now $0.10 — and Qwen3.6 at $0.10 is sitting right there next to it. That comparison deserves a fresh look.
2. Check cached-input pricing, not just the headline input price.
The Mercury 3 promo has cached input at a cent per million tokens. DeepSeek's off-peak rates are half of peak. If your agent re-reads the same system prompts and few-shot examples across turns — and most agents do — your cached-input ratio matters more than the sticker price. The cheapest headline price with bad caching terms loses to a pricier model with aggressive cache discounts. Do the math on your actual token mix.
3. Try the November-1 deadline item.
DeepSeek's Search+Fetch APIs are free via Claude Code and Codex integrations until billing starts November 1. Eleven days. If you've been meaning to test whether retrieval-through-API beats your current RAG setup, this is the cheapest window you'll ever get. Test now, or at least plan for the billing switch so it doesn't surprise you mid-sprint.
4. Treat promos as promos.
Mercury 3 at 50% off is a launch promo. It will end. Build your stack's cost model on the standard rates ($0.20/$0.70), and treat the promo as a head start — not as the number you budget against. I've seen teams lock in promos mentally and then get a very rude surprise in their December invoice. Don't be those teams.
5. Watch the two that got more expensive.
Solar Pro 4 jumped 67% — $0.09 to $0.15. If it's anywhere in your stack, that's the line item to revisit first. Price increases in a week of cuts usually mean a provider is repositioning that model upmarket, which means more increases could follow. Qwen3 Next 80B's 11% bump is smaller but worth noting if it was your budget pick.
The honest caveat
Input price isn't the whole bill. Output token ratios, cache hit rates, and retry behavior dominate real agent spend — a "cheap" model that needs three attempts costs more than an expensive one that nails it first try. And price-tracking snapshots can lag the actual provider pages, so verify the current price on the provider's own pricing page before you commit to anything. Treat this post as a signal to go look, not as the final word.
One question
Has anyone actually re-routed an agent to Qwen3.6 or DeepSeek V3 this week — did the output quality hold? I'm genuinely curious whether the cheap tier is good enough for real agentic loops or whether we're all still quietly paying frontier prices for anything that matters.
React if this saved you a pricing-page scroll. No affiliate links in this one — it's news, not a buying guide.
Top comments (0)