Yes — input prices moved on October 18, 2026, and the biggest reported jump was a 400% increase on Z.AI’s GLM 5.3 Flash. But here’s the part most coverage skips: the numbers come from a third-party price tracker, and when I checked DeepSeek’s and Z.AI’s official pricing pages the same day, the first-party rates told a different story. Your bill depends on which number actually applies to you — the tracker’s or the vendor’s. This article gives you both, works the real math on what these moves do to a small-business API bill, and lays out a concrete defense plan so the next price shock doesn’t blindside you.
What the October 18 tracker actually reported
The AI news site olud.ai publishes a daily pricing digest built from its own tracking data. Its October 18, 2026 edition flagged five input-price moves in a single day:
| Model | Provider | Reported move | New input rate |
|---|---|---|---|
| DeepSeek V4 Pro 0423 | DeepSeek | −28% | $0.21 / 1M tokens |
| DeepSeek V3.2 | DeepSeek | +8% | $0.28 / 1M tokens |
| DeepSeek V4.1 Flash | DeepSeek | +100% | $0.08 / 1M tokens (was $0.04) |
| GLM 4.7 | Z.AI | −10% | $0.54 / 1M tokens (was $0.60) |
| GLM 5.3 Flash | Z.AI | +400% | $0.15 / 1M tokens (was $0.03) |
The tracker summed it up in one line: the moves “widen the gap between premium and budget tiers across both providers.” That’s the headline — and it’s worth your attention, because it captures a real trend. Budget models are getting repriced upward just as premium models drift down, which squeezes the middle.
One related note, since it shares the same week: on October 17 DeepSeek released DeepSeek-V4.1-Flash-1017, an updated checkpoint of V4.1 Flash with MIT-licensed weights on Hugging Face and no change to its official API prices.
But before you rework your budget around the tracker’s numbers, you need to know what they actually measure — because they are not the vendors’ official prices.
Why the official prices don’t match the tracker
I opened both vendors’ official pricing pages on October 19, 2026. Here’s what they actually say.
DeepSeek’s official API pricing (api-docs.deepseek.com/quick_start/pricing):
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| deepseek-flash (V4.1-Flash-1017) | $0.15 off-peak / $0.30 peak | $0.003 off-peak / $0.006 peak | $0.60 off-peak / $1.20 peak |
| deepseek-v4-pro (V4-Pro-0813) | $0.66 off-peak / $1.32 peak | $0.022 off-peak / $0.044 peak | $1.98 off-peak / $3.96 peak |
Z.AI’s official pricing (docs.z.ai/guides/overview/pricing):
| Model | Input | Cached input | Output |
|---|---|---|---|
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 |
| GLM-4.7 | $0.60 | $0.11 | $2.20 |
| GLM-5.3 (flagship) | $1.40 | $0.26 | $4.40 |
Now compare. The tracker says DeepSeek V4.1 Flash input went from $0.04 to $0.08 — but DeepSeek’s own page charges $0.15 off-peak and $0.30 at peak. The tracker says GLM 4.7 dropped to $0.54 — but Z.AI’s page still lists $0.60. The tracker’s numbers are roughly half to a third of the official first-party rates.
So who’s wrong? Neither, exactly — they’re measuring different markets. Trackers like olud.ai typically follow aggregator and third-party-host listings (platforms like OpenRouter, where the same model is served by many independent providers at different prices), not the vendor’s own checkout page. And third-party prices swing wildly: on OpenRouter, independent hosts have listed DeepSeek V4.1 Flash input anywhere from $0.0001 to $0.375 per million tokens, per pricing analyses published this month. Some of those listings are genuine wholesale-style pricing; some are promotional loss-leaders; some may not be sustainable.
This matters because your invoice follows the route your tokens actually travel. If you call api.deepseek.com directly, you pay DeepSeek’s official rate — and the Oct 18 tracker moves don’t apply to you at all. If you route through an aggregator or a third-party host, you’re exposed to their repricing, which is what the tracker is picking up.
And here’s the twist that proves the distinction matters: one of the tracker’s numbers does match an official price. The tracker’s new GLM 5.3 Flash rate ($0.15 input) is exactly Z.AI’s official list price. What likely happened is that third-party listings had been reflecting Z.AI’s expired launch promotion and finally snapped back to list. That promotion is worth understanding on its own, because it’s the cleanest verified price shock in this whole story.
The price shock nobody announced: GLM-5.3-Flash’s promo expiry
When Z.AI launched GLM-5.3-Flash, it ran a 50% launch discount: $0.075 per million input tokens and $0.25 per million output, against a list price of $0.15/$0.50. That promotion expired at midnight on September 9, 2026 (UTC+8) — with no announcement. Reporting by codersera confirmed that OpenRouter’s models API was still returning the promotional rates while the discount was live, and a re-check on October 5, 2026 showed the full $0.15/$0.50. In other words, anyone budgeting against the promo rate woke up to a 100% price increase with zero warning.
This is the pattern you should actually plan for. Google is running the same playbook right now with Gemini 4 Argon: introductory API rates of $2 per million input and $10 per million output tokens through November 30, 2026, then standard pricing of $4 and $20 — a doubling, scheduled and dated, but easy to miss if you’re not watching the calendar.
What a 2x–5x input-price move does to your monthly bill
Percentages are abstract. Dollars are not. Let’s work two realistic small-business scenarios using the rates the October 18 tracker reported, so you can see exactly how a move like GLM 5.3 Flash’s +400% input hike lands on a real invoice.
Scenario 1: a customer-support chatbot. Say you run 500 chat conversations a day. Each conversation averages 3,000 input tokens (system prompt, conversation history, retrieved help-center context — mostly uncached on a third-party route) and 600 output tokens. That’s 15,000 chats a month, 45 million input tokens, and 9 million output tokens.
On GLM-5.3-Flash at the tracker’s rates (input $0.03 → $0.15; official output $0.50/M, unchanged):
- Input cost before: 45 × $0.03 = $1.35
- Input cost after: 45 × $0.15 = $6.75
- Output cost: 9 × $0.50 = $4.50 (unchanged)
- Monthly total: $5.85 → $11.25 — a 92% increase on the whole bill, driven entirely by the input side.
The same workload on DeepSeek V4.1 Flash at the tracker’s rates (input $0.04 → $0.08; DeepSeek’s official off-peak output $0.60/M, unchanged):
- Input cost before: 45 × $0.04 = $1.80
- Input cost after: 45 × $0.08 = $3.60
- Output cost: 9 × $0.60 = $5.40 (unchanged)
- Monthly total: $7.20 → $9.00 — a 25% increase.
Notice the lesson inside those two rows: the +400% input hike produced a +92% total increase, while the +100% input hike produced only +25%. Your total bill’s sensitivity to an input-price move depends on your input/output ratio. Chat workloads generate lots of output, which cushions the blow.
Scenario 2: a nightly batch document job. Say you run an invoice-extraction pipeline that processes 10,000 documents a month. Each document is 8,000 input tokens (a long scanned PDF) and produces a 200-token structured summary. That’s 80 million input tokens and just 2 million output tokens a month — a heavily input-skewed workload.
On GLM-5.3-Flash at the tracker’s rates:
- Input cost before: 80 × $0.03 = $2.40
- Input cost after: 80 × $0.15 = $12.00
- Output cost: 2 × $0.50 = $1.00 (unchanged)
- Monthly total: $3.40 → $13.00 — a 282% increase.
For input-heavy batch work, there’s almost no cushion. The input price is the bill.
Two more things worth internalizing from this math. First, at these budget-tier rates, the absolute dollars are still small — $13 a month won’t sink anyone. The danger is the multiplier applied to a workload you scale up: the same batch job at 10x volume goes from $34 to $130 a month overnight, and if you’ve priced a client-facing product on the old rate, your margin evaporates. Second, none of these scenarios include the second input line-item most people forget: cached input. On DeepSeek’s own API, a cache hit on Flash costs $0.003 per million tokens off-peak versus $0.15 on a miss — a 98% discount. Rerun Scenario 1’s 45M input tokens with an 80% cache-hit rate: 36M cached × $0.003 = $0.11, plus 9M uncached × $0.15 = $1.35, for an input total of about $1.46 — versus $6.75 with no caching. Cache discipline turns a price hike into a rounding error. We’ll come back to that.
How price shocks actually reach your bill
Based on everything verified this week, input-price shocks reach you through three mechanisms — and knowing which one you’re exposed to tells you which defense works.
1. Promo and introductory rates expire. This is the GLM-5.3-Flash story, the Gemini 4 Argon story (intro rates end November 30, 2026), and historically the most common way small businesses get surprised. The vendor’s price page is the contract; the promo rate is a footnote with an end date. When the date passes quietly, your bill doubles and support tells you the price was always going to do that.
2. Third-party hosts reprice. If you buy tokens through an aggregator, your per-token rate is set by whichever host serves your request, and hosts change prices without telling you. That’s what the October 18 tracker is really showing: third-party DeepSeek listings doubling on input, third-party GLM listings snapping back to official list price. You’re not buying from DeepSeek or Z.AI; you’re buying from a reseller whose economics you can’t see.
3. Time-of-day windows shift your effective rate. DeepSeek bills peak rates (double off-peak) during 01:00–04:00 and 06:00–10:00 UTC on weekdays — which is evening-to-night in the US Eastern time zone, exactly when a US customer-support bot is busiest. If your traffic lands in peak hours, you’re already paying the “high” rate; the official price table, not any tracker, is what governs.
Five ways to shock-proof your AI spend
1. Pin the model version, not just the model name. Aliases like deepseek-flash are routing pointers — DeepSeek moved that alias to the new 1017 checkpoint on October 17 without a price change this time, but the alias can point anywhere tomorrow. Pin the dated snapshot (e.g., DeepSeek-V4.1-Flash-1017) in your config, and log the exact model string with every billing period. When a vendor deprecates or reroutes an alias, pinned versions fail loudly instead of billing you at a new rate silently — the same discipline applies to model shutdowns like OpenAI’s GPT-4o and GPT-4.1 API shutdown scheduled for January 2027.
2. Engineer for cache hits. This is the single highest-leverage move. Put your stable content — system prompt, tool definitions, retrieved reference documents — at the very start of every request, keep your API parameters identical between calls, and stop appending timestamps or random IDs to prompt prefixes. On DeepSeek’s official pricing, the gap between a cache miss ($0.15/M) and a cache hit ($0.003/M) off-peak is 50x. On Z.AI, cached input is $0.03/M against $0.15 — an 80% discount. A workload with high cache hits is structurally insulated from input-price hikes.
3. Push batch work off-peak. DeepSeek’s off-peak rate is literally half the peak rate, and peak hours are weekday mornings UTC — overnight in the US. If your document pipeline can run at 2 AM Eastern instead of 2 PM, you’ve just given yourself a permanent 50% discount that no tracker can take away. Schedule cron jobs accordingly.
4. Set billing alerts before you need them. Every major provider lets you set spend alerts; most small teams never turn them on. Set a daily alert at 150% of your normal run-rate and a monthly alert at your budget ceiling. A 400% input hike shows up in your daily alert within 24 hours instead of in next month’s invoice. Also log per-feature token usage (input, cached input, output, reasoning) so you can see which workload moved, not just that the total moved.
5. Keep a promo calendar and a fallback route. Write down every promotional or introductory rate you’re relying on, with its expiry date — the GLM-5.3-Flash September 9 expiry and the Gemini 4 Argon November 30 changeover are the two live ones to track right now. And maintain a tested fallback: know which model you’d switch each workload to, and verify it handles a sample of real traffic before you need it. The cheapest fallback is often the same model family one tier down (GLM-5.3-Flash → GLM-4.7-Flash is free, for example), not a different vendor.
What developers are actually saying
The pricing conversation in developer communities this month isn’t about the tracker numbers — it’s about the feeling of losing control of the bill. In the Hacker News discussion on DeepSeek 4.1 Flash’s pricing, one commenter put it bluntly: “With the current deepseek pricing, the same membership lasts me a week at most… It used to be that you could dump pocket change into the deepseek api and forget about it. Nowadays it’ll make you notice real fast as the dollars pile up.” (source)
That sentiment is the real story here. Nobody’s angry about $6.75 versus $1.35. They’re uneasy because the rate changed underneath them and they found out from the invoice.
The bottom line
The October 18 tracker moves are real data about real third-party price changes — up to +400% on GLM 5.3 Flash input, +100% on DeepSeek V4.1 Flash input, with premium tiers drifting down at the same time. But they are not DeepSeek’s or Z.AI’s official prices, which I verified unchanged at $0.15/$0.30 (Flash) and $0.15/$0.60 (GLM 5.3 Flash) respectively for direct API buyers. Know which market you’re buying in, work the math for your own input/output ratio, and build the five defenses above before the next shock. The cheapest AI bill isn’t the one with the lowest per-token rate today — it’s the one that can’t surprise you tomorrow.
Frequently Asked Questions
Did DeepSeek raise its API prices in October 2026?
Not on its official pricing page. As of October 19, 2026, DeepSeek’s first-party API still charges $0.15 per million input tokens off-peak ($0.30 peak) for deepseek-flash and $0.66 off-peak ($1.32 peak) for V4 Pro, with output at $0.60/$1.98 off-peak. The reported increases (V4.1 Flash +100%, V3.2 +8%) come from the olud.ai price tracker, which follows third-party and aggregator listings — not api.deepseek.com. If you call DeepSeek directly, those moves don’t apply to your bill.
Did Z.AI raise GLM prices in October 2026?
Z.AI’s official pricing page still lists GLM-5.3-Flash at $0.15 input / $0.50 output and GLM-4.7 at $0.60 input / $2.20 output. The tracker’s reported +400% move on GLM 5.3 Flash (from $0.03 to $0.15) appears to reflect third-party listings normalizing back to Z.AI’s official list price after the model’s 50%-off launch promotion expired quietly on September 9, 2026.
How much does a 400% input price increase actually cost a small business?
It depends on your input/output ratio. For a support chatbot doing 45M input and 9M output tokens a month on GLM-5.3-Flash, the bill goes from $5.85 to $11.25 (+92%). For an input-heavy batch job (80M input, 2M output), it goes from $3.40 to $13.00 (+282%). Output-heavy workloads are cushioned; input-heavy ones take the full hit.
What’s the difference between tracker prices and official API prices?
Official prices are what the model vendor charges on its own API. Tracker prices typically reflect what third-party hosts and aggregators (like providers on OpenRouter) charge to serve the same model — and those can be far higher or lower than official, change without notice, and include promotional rates. Always check which route your API calls actually use.
How can I protect my business from sudden AI price changes?
Pin dated model versions instead of aliases, engineer your prompts for maximum cache hits (up to 98% cheaper on DeepSeek), schedule batch jobs in off-peak windows, set daily billing alerts at 150% of your normal spend, and keep a calendar of every promotional rate’s expiry date — with a tested fallback model for each workload.
Are budget AI models still getting cheaper overall?
The October 18 data shows a split: premium tiers are getting cheaper (DeepSeek V4 Pro input −28%, GLM 4.7 −10%) while budget tiers are being repriced upward (+100% to +400% on Flash models). The gap between premium and budget is widening, which makes workload placement — premium where quality matters, budget where it doesn’t — more important than chasing the single cheapest per-token rate.