AI prices dropped again on September 22, 2026: within 90 minutes, Anthropic and then OpenAI both announced new models cheaper than their predecessors, by as much as half for some of them. For an SMB leader budgeting AI tools, this double announcement is worth understanding: it confirms an underlying trend that can lighten your bill, provided you know how to read it.
At a glance
- On September 22, 2026, Anthropic launched Claude Opus 5.5, and OpenAI followed roughly 90 minutes later with GPT-6 Sol and GPT-6 Luna.
- OpenAI halved the price of its two new models compared to their GPT-5.6 equivalents.
- Anthropic cut Claude Opus 5.5's price by 20% versus Opus 5, and by as much as 60% on cache reads, heavily used by AI agents.
- Both vendors also claim better performance at a lower price: fewer factual errors for GPT-6 Sol, better agentic scores for Opus 5.5.
- The practical takeaway for an SMB is simple: the cost of the same level of AI quality keeps falling, which justifies reviewing your model choice at least once a quarter.
What happened on September 22, 2026
September 3, 2026
GPT-6 Astra
September 22, 2026, late morning
Anthropic launches Claude Opus 5.5
September 22, 2026, about 90 minutes later
OpenAI launches GPT-6 Sol and Luna
This tight timing is no coincidence: it illustrates the head-to-head competition between the two labs, which now respond to each other almost in real time, on price as much as on performance.
The new pricing, in detail
GPT-6 Luna, OpenAI's most affordable model, is priced at $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20 for its predecessor, GPT-5.6 Luna. It targets high-volume, lower-complexity tasks: sorting messages, extracting information, quick answers.
GPT-6 Sol, positioned for more demanding tasks such as coding, drops to $2 input and $10 output per million tokens, versus $4 and $20 for GPT-5.6 Sol. OpenAI states that Sol makes about half as many factual errors as its predecessor, reaching reliability close to the premium Astra model at a much lower cost.
Claude Opus 5.5, Anthropic's flagship model, costs $4 input and $20 output per million tokens, 20% below Opus 5. The steepest cut is on cache reads (reusing context already sent, very common in AI agents and coding work): the price drops by 60%, to $0.20 per million tokens.
A comparison table to keep straight
| Model | Vendor | Input price ($/M tokens) | Output price ($/M tokens) | Change vs previous generation |
|---|---|---|---|---|
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | about -50% |
| GPT-6 Sol | OpenAI | $2 | $10 | -50% |
| Claude Opus 5.5 | Anthropic | $4 | $20 | -20% |
| Claude Opus 5.5 (cache reads) | Anthropic | $0.20 | - | -60% |
Why prices are falling this fast
Both companies give a similar technical explanation: the gains mainly come from improvements in caching and inference, meaning how a trained model processes requests, rather than from a radically different model. These technical savings are then passed on to the price charged to businesses and developers.
There is also a purely competitive factor: OpenAI and Anthropic chase the same customers (developers, software vendors, automation agencies) who partly decide based on price per token. A cut from one mechanically pushes the other to respond, which explains the tight timeline of September 22, 2026.
What this means for the tool you use every day
If you use ChatGPT or Claude through a consumer subscription (Plus, Pro, Business), these price cuts primarily concern the API, meaning developers building tools on top of these models. The effect on your subscription is indirect: it tends to show up as better features at the same price, or cheaper third-party tools, over the following months.
What this actually changes for an SMB
Before September 22, 2026
Since September 22, 2026
Three concrete actions worth considering:
- Ask your automation provider for a new quote. If a project was deemed too expensive a few months ago because of token pricing, the math has changed: ask for an updated estimate with current rates.
- Check which model your AI tools actually run on. A slow-moving software vendor may keep billing you at a rate based on the older model generation. Ask explicitly whether it has switched to the newer, cheaper versions.
- Don't switch providers on price alone. The price gap between GPT-6 Luna and Claude Opus 5.5, for instance, is substantial, but these two models don't target the same tasks. The right approach remains matching the model to the task, as detailed in our guide to choosing the right AI model.
Limits worth keeping in mind
In the interest of honesty, a few caveats before revising your entire AI budget downward:
- The GPT-5.6 reference prices OpenAI used to calculate its 50% cut were promotional rates, not always the historical list price: the real-world savings for some customers may be smaller.
- A lower price per token does not guarantee a lower total bill: if a cheaper model encourages using it more often or on more tasks, overall spend can stay flat, or even rise.
- The performance gains claimed (fewer errors, better agentic scores) rely on internal benchmarks run by each company: they're a useful signal, but worth verifying against your own use cases before committing.
- Nothing guarantees this pace of price cuts will continue at the same rate in the coming months: it also depends on the competitive pressure at the time, including from players like Google or open Chinese models.
FAQ
Why did Anthropic and OpenAI cut prices on the same day?
Both companies compete for the same developer and enterprise customers, who partly decide based on price per token. Anthropic's launch of Claude Opus 5.5, followed 90 minutes later by OpenAI's GPT-6 Sol and Luna, illustrates this head-to-head competition on timing as much as on price.
Do these price cuts affect my ChatGPT or Claude subscription?
Not directly. These rates apply to the API, used by developers and software vendors to build tools. A consumer subscription (ChatGPT Plus, Claude Pro) benefits more indirectly, through better features or cheaper third-party tools in the following months.
Should I switch AI models right now to save money?
Not automatically. The right approach is matching the model to the task (fast and cheap for volume, more powerful for complex work) rather than choosing on price alone. Instead, ask your current provider whether it has already switched to the newer, cheaper model generations.
What is the difference between GPT-6 Sol and GPT-6 Luna?
Sol targets more demanding reasoning tasks, such as coding or analysis, priced at $2/$10 per million tokens. Luna targets high-volume, lower-complexity tasks (sorting, extraction, quick answers), at a much lower rate of $0.10/$0.50 per million tokens.
AI costs keep falling at a steady pace in 2026, opening up use cases previously considered too expensive. To go further on how to choose between these models, read our guide to choosing the right AI model in 2026, or see how other SMBs structured their automation projects in our customer stories.
