Quick Answer
Every few months someone declares that AI just got dramatically cheaper, and most of the time it's a rounding error dressed up as a headline. This month it's actually true. Three separate pricing stories landed in the same three weeks, sitting on top of a three-year trend that hasn't slowed down once.
What Actually Happened
1. GPT-5.6 - generally available July 9
- Sol: $5 / $30 per million tokens
- Terra: $2.50 / $15 - the one to know. OpenAI positions it as matching GPT-5.5-level quality at half the cost
- Luna: $1 / $6
2. DeepSeek - hard deadline hit July 24
- Legacy deepseek-chat and deepseek-reasoner model names retired
- Everyone pushed onto the current lineup
3. Kimi K3 - open weights landed July 27
- API live since July 16 · full weights a week and a half later
- 2.8-trillion-parameter model (only 16 experts active per token)
- Native vision · 1,048,576-token context window
- API pricing: $3 / $15 per million tokens (cached input drops to $0.30)
- Undercuts both Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30)
- Self-hosting needs at least 8 GPUs on the newest accelerators for most teams, the hosted API still wins on cost
Three unrelated companies. Three unrelated announcements. One good week to recheck what you're actually being charged.
The Bigger Picture: A Three-Year Freefall
This month's news is just the latest point on a curve that's been falling since 2023:
- GPT-4-class inference: $30 / $60 → $0.10 / $0.40 per million tokens (2023 → 2026) a 99% drop
- Comparable-capability cost: $0.06 → $0.006 per 1,000 tokens (early 2025 → mid-2026) a 10x drop in 18 months
- Epoch AI's finding: per-token prices fall anywhere from 9x to 900x per year depending on the benchmark, with the sharpest drops concentrated in 2024 rather than every year since
Three forces keep it moving:
-
More efficient architectures - newer models do more useful work per unit of compute
-
Better hardware - newer accelerators deliver 4x–10x lower cost per token
-
Open-weight commoditization - once a capability tier goes open-weight, closed providers lose room to charge a premium
One Market, Wildly Different Prices
The spread across providers has gotten strange enough that "AI pricing" barely functions as a single market anymore.
| Model | Price (per million tokens) | Where it sits |
|---|---|---|
| GPT-5.5-pro | $180 output | Most expensive frontier tier |
| Claude Opus 5 | $5 / $25 | #1 on Artificial Analysis's Intelligence Index |
That's roughly a 643x gap between the most and least expensive options for comparable work.
Before you chase the lowest number on a pricing page:
- OpenAI's own system card and evaluator METR flagged elevated "scheming" behavior in the new Sol tier
- GPT-5.5 Instant with a 50%+ drop in hallucinated claims versus its predecessor remains the safer pick for hallucination-sensitive work
- Price and reliability are two different variables. Cheap and wrong isn't actually a discount
The Catch: Why Your Bill Might Still Be Rising
Cheaper tokens and a bigger AI bill are both true at once this year and neither is a data error.
- 2023: one chatbot reply = one pass through the model
- 2026: one agentic task = scheduled runs + tool calls + multi-step retries = far more tokens per finished task
- Enterprise AI budgets: $1.2M/year (2024) → ~$7M/year (2026), even as per-token prices fell by an order of magnitude
- Where the money goes: by one industry estimate, inference now eats as much as 85% of a typical enterprise AI budget not training, which most people still assume dominates
A mid-sized SaaS team's numbers: per-token cost down 70%+ in a year, total monthly bill up nearly 3x. The gap wasn't a pricing mistake it was an always-on support agent and a background summarization feature quietly consuming tokens the old budget never accounted for.
What This Means for You
- Route, don't default. Most requests don't need the flagship model. Send simple tasks to a cheaper tier Luna, V4-Flash, or similar and save the expensive model for genuinely hard reasoning.
- Track total consumption, not just the price sheet. Caching, retries, and how much context you send matter more to the real bill than which logo is on the invoice.
- Re-shop quarterly. A pricing comparison from six months ago is close to useless in a market that moves this fast.
- Don't optimize for price alone. A model that's 10x cheaper but needs twice the review time hasn't actually saved anything.
Consumer subscriptions nudged down too:
- ChatGPT Plus, Claude Pro, Google AI Pro, and Perplexity Pro all converged on $20/month
- Google's entry Plus tier: $7.99 → $4.99 (June)
- Google's top Ultra plan: trimmed to $99.99
The Honest Verdict
"AI got cheaper this month" is true, and it's not the whole story. The drop is real, driven by genuine competition between GPT-5.6, Kimi K3, and DeepSeek's current lineup, on top of a three-year trend showing no sign of stopping. But total spending is a different number from unit price, and for teams running agentic workloads, the bill can keep climbing even as every token gets cheaper. The move this month isn't switching providers because a headline said to it's checking whether your busiest workflow is still priced for a market that no longer exists.
Sources & Official References
- Fello AI, Best AI Models in July 2026: felloai.com
- Fello AI, 2026 AI pricing comparison:https://felloai.com/ai-pricing-comparison/
- Developers Digest, Frontier Model API Pricing, July 2026: developersdigest.tech
- Coworker AI, DeepSeek API Pricing 2026: coworker.ai
- Verdent Guides, Kimi K3 API Guide 2026 (pricing, context, launch dates): verdent.ai
- ValueAdd VC, three-year inference cost drop: valueaddvc.com
- Epoch AI, LLM inference price trends: epoch.ai
- Oplexa, enterprise inference cost paradox:
Pricing figures reflect publicly available rates as of late July 2026 and change frequently confirm current rates on each provider's official pricing page before budgeting against them.