LLM Token Prices Fell Up to 80% — What It Means for Your Business
By Arash Latifi — Founder, The DevArt
A real timeline of GPT and Claude pricing through fall 2026: why models got cheaper, the hidden role of cache costs and DeepSeek, and a practical checklist.
TL;DR — Between June and October 2026, OpenAI and Anthropic cut prices repeatedly, up to 80%; the main drivers were open-weight competition (DeepSeek) and cache-read cost reductions — and per Vercel data, real token cost is down to 43% of April 2026 levels.
The pricing timeline (from Max Woolf's analysis)
| Date | Event | Price per MTok (input / output / cache read, USD) | | --- | --- | --- | | 2026-06-26 | OpenAI launches GPT-5.6 Sol and Luna | Sol: 5 / 30 / 0.50 (unchanged); Luna small model: 1 / 6 / 0.10 | | 2026-06-30 | Anthropic launches Claude Sonnet 5 | 2 / 10 / 0.20 (a "temporary" cut later made permanent) | | 2026-07-30 | OpenAI cuts GPT-5.6 Luna by 80% | Down to 0.20 / 1.20 / 0.02 | | 2026-09-01 | Anthropic launches Fable 5.1 | Same 10 / 50 I/O price but cache read cut 75% (to 0.25) | | 2026-09-22 | OpenAI launches GPT-6 Sol and Luna | 2 / 10 / 0.20 and 0.10 / 0.50 / 0.01 — half the previous prices | | 2026-09-22 | Anthropic launches Claude Opus 5.5 | 4 / 20 / 0.20; Anthropic says it "requires less compute to serve than Opus 5" |
Why did prices crater?
Max Woolf's analysis (October 25, 2026) points to three causes:
-
Open-weight competition. When DeepSeek V4 Flash launched in April 2026 at $0.14 / $0.28 / $0.0028, it went viral — especially because its cache read was 2% of input price versus 5–10% for OpenAI/Anthropic models. The 80% Luna cut in July was likely aimed squarely at these small open-weight models.
-
No moat. When models are swappable via a toggle in tools like OpenRouter, and 29 distinct providers host the same DeepSeek model, pricing isn't set by two companies. Competition among inference providers pushes price toward marginal cost.
-
Real infrastructure gains. Anthropic's official line for Opus 5.5: it "requires less compute to serve than Opus 5, and its pricing reflects that" — part of the discount is genuinely better inference efficiency, not just a price war.
Vercel's AI Gateway data confirms the trend: cost-per-token is down to 43.3% of April 2026, yet aggregate spend keeps rising because models are getting better faster than they're getting cheaper.
The hidden factor: cache read costs
For agentic workloads (like AI coding agents), cache read costs are 96–99% of total token spend, because the whole conversation history is re-read from cache every turn. A 75% cache-read cut therefore translates to a 70–75% cut in the effective cost of using that model. If you're comparing API pricing, scrutinize the cache-read number more than headline input/output prices.
What it means for your business
- [ ] Compare your current vendor's rates against official model pricing — cuts may not have reached the middleman layer
- [ ] If you're on a subscription (e.g. Codex), watch your allowance: OpenAI cut the $200/month tier quota by 50% alongside GPT-6.1 Sol
- [ ] Re-price automations you shelved over API costs — scenarios that were uneconomical six months ago may work now
- [ ] For high volume, try open-weight models via hosted services (e.g. OpenRouter); effective costs across many providers are nearly identical
- [ ] Don't bet on a sudden "token price correction" — with low switching costs and many competing providers, the analysis argues that fear isn't supported
FAQ
Q: Which models do these prices cover?
The analysis covers GPT-5.6 through GPT-6.1, Claude Sonnet 5 through Opus 5.5, and DeepSeek V4. Always check each provider's current official pricing before deciding — this market shifts every few weeks.
Q: Did quality drop along with prices?
Per the same analysis, no — models don't appear to have cut corners, and Opus 5.5 even evaluated better than more expensive predecessors. This was competition plus compute efficiency, not quality-trimming.
Q: Is now a good time to start using AI in my business?
Cost-wise, conditions are far better than six months ago. But anchor the decision to a specific problem: find a repetitive task, pilot it with a cheap model, then scale — not the other way around.