LLM price cuts have come fast in 2026. OpenAI cut GPT-5.6 Luna by 80% on 30 July 2026, and Anthropic has pushed its newest flagship, Claude Opus 5.5, 20% below Opus 5. The reasons are less mysterious than they look: cheaper rivals, cheaper serving, and customers who can switch in an afternoon.
The question was raised in a post by Max Woolf, which tracks the price changes. This article checks those figures against other reporting and looks at what is driving them.
Key takeaways
- The 80% cut is OpenAI’s, applied to the small GPT-5.6 Luna model. Anthropic’s cuts are smaller, around 20% on list price.
- Open-weight models from China, such as DeepSeek, set a low price floor that closed labs now have to respect.
- Newer models are cheaper to run, and both companies say so.
- Falling price per token does not mean falling bills. Total spend keeps rising because usage grows faster.
Who actually cut prices, and by how much?
Prices below are dollars per million tokens, input then output. A token is roughly three-quarters of a word.
| Date | Model | Price move |
|---|---|---|
| 30 July 2026 | OpenAI GPT-5.6 Luna | $1 / $6 down to $0.20 / $1.20 (-80%) |
| 21 August 2026 | OpenAI GPT-5.6 Sol | $5 / $30 down to $4 / $20, for three months |
| 22 September 2026 | OpenAI GPT-6 Sol | Launched at $2 / $10 |
| 22 September 2026 | Anthropic Claude Opus 5.5 | $5 / $25 (Opus 5) down to $4 / $20 (-20%) |
Reporting from Implicator confirms the Luna cut came only weeks after launch. Cuts like that normally arrive months later. The Sol discount was reported by Reuters and others.
Anthropic also made the launch price of Claude Sonnet 5 ($2 / $10) permanent on 10 August 2026, according to Woolf’s timeline. Third-party price trackers list Opus 5.5 at $4 / $20, with cache reads down from $0.50 to $0.20. I could not confirm the figures on Anthropic’s own pricing page, so check it before you budget.
So the headline “both cut by up to 80%” is a bit generous. One company cut a small model by 80%. The other made steadier, smaller cuts on its flagship.
Why are LLM prices falling so fast?
Four forces show up repeatedly in the reporting. None is proven as the single cause, so treat the ranking below as my reading.
1. Open-weight rivals set the floor
Open-weight models are models whose trained parameters anyone can download and host. DeepSeek V4 Flash launched in April 2026 at $0.14 input and $0.28 output per million tokens, per Woolf. Jefferies research, cited by Silicon UK, found average inference prices hit a year-to-date low in August 2026 and credited cheap Chinese open-weight models.
When DeepSeek cut V4-Pro by 75% in May 2026, one analyst told InfoWorld it was an efficiency gain passed on, not a promotion. Woolf suspects the Luna cut was aimed squarely at small open models.
2. Switching costs almost nothing
Gateways such as OpenRouter let a developer change the model name in one line. Think of petrol stations on the same corner: if one is a few cents cheaper, the queue moves. Woolf also points to Kimi K3.1 Pro at $2 / $10 scoring 55 on the Artificial Analysis index, against 58 for Opus 5.5. That is close enough for many jobs.
3. Newer models are cheaper to serve
Anthropic said Opus 5.5 “requires less compute to serve than Opus 5, and its pricing reflects that.” OpenAI gave a similar reason when it halved the Codex allowance on its $200 monthly plan, citing “increasingly efficient models”. The savings are partly real engineering. One analyst noted that the rate card alone explains only about 24 points of Anthropic’s claimed 40% saving, and an independent test found no saving at maximum effort.
4. A race before the IPOs
Both companies are rumoured to be heading toward public listings, and reports describe a fight for developer share. Woolf finds the timing odd, since leaked financials show large losses. In my opinion, share now is worth more to them than margin now. See our look at the OpenAI funding round for the money behind that bet.
Does cheaper per token mean a cheaper bill?
Usually not. Woolf cites Vercel AI Gateway data showing cost per token fell to 43.3% of its April 2026 level while total spend rose. Agents that loop, read whole codebases and call tools burn far more tokens than a chatbot reply.
Limits also change in odd ways. Our piece on ChatGPT Business limits shows why chat feels unlimited while Codex is not.
Will prices keep dropping?
Woolf argues prices should keep sliding toward the marginal cost of serving a token, and that predictions of a future price hike lack support. That is plausible while open models keep closing the gap. One caution: some cuts are temporary, like the three-month Sol discount, and rate cards often hide extras such as the 1.1x multiplier for US-only inference on Opus 5.5.
What happens next
The Sol promotional price is due to run for three months from 21 August 2026, so watch for a decision around late November. Reports from 16 October 2026 said OpenAI was still weighing deeper cuts with no final decision. Anthropic’s answer will likely come with its next model release rather than a headline cut.
Expect the real contest to move away from list price. Cache pricing, reliability and agent quality matter more once every vendor sits within a few dollars of the others. Capacity limits matter too, as in our note on why the DeepSeek API is slow at night.
FAQ
Did Anthropic cut prices by 80%?
No. The 80% cut was OpenAI’s, on GPT-5.6 Luna. Anthropic’s Opus 5.5 is about 20% cheaper than Opus 5 on list price.
Are the cheaper models worse?
Not necessarily. Small and open models trail the top tier slightly on benchmarks but often suffice for routine tasks. Test on your own workload before switching.
Should I wait for lower prices before building?
No. Prices have fallen steadily, and switching providers is easy. Build behind a gateway so you can move when a better deal arrives.
Next step: open your last API invoice, split it by model, and price the same tokens on one cheaper option. The gap may surprise you.
Leave a comment