GPT-6.1 Sol $2/$10 and Opus 5.5 $4/$20: how LLM API prices fell since June

Frontier AI models have become much cheaper to call over the past four months, and the vendors' own price pages now show it. As of 26 October 2026, OpenAI lists GPT-6.1 Sol at US$2 per million input tokens and US$10 per million output tokens, while Anthropic lists Claude Opus 5.5 at US$4 and US$20 and Claude Sonnet 5.5 at US$2 and US$10. A developer blog, minimaxir.com, has tracked how the prices got there since June, and the timeline shows the two companies cutting in turn. This article checks today's numbers against the vendors' pages, sets out the history as that blog records it, and says what remains uncertain.

Key points

• OpenAI's pricing page lists GPT-6.1 Sol at US$2 / US$10 per million input / output tokens with US$0.10 for cached input, and GPT-6 Luna at US$0.10 / US$0.50.
• Anthropic's pricing documentation lists Claude Opus 5.5 at US$4 / US$20, Sonnet 5.5 at US$2 / US$10 and Fable 5.1 at US$10 / US$50.
• On 22 September 2026 OpenAI released GPT-6 Sol and Luna with Sol at half GPT-5.6 Sol's promotional price and Luna cut from US$0.20 / US$1.20 to US$0.10 / US$0.50, according to implicator.ai.
• An Artificial Analysis comparison reported by implicator.ai found GPT-6 Sol matched Opus 5.5 on business workflows at about 40% of the cost, but trailed it on Terminal-Bench 4.0 (43.9% against 52.5%).
• GPT-5.6 Sol's promotional pricing (US$4 / US$20) is listed by OpenAI as available at least through 21 November 2026.

What happened

The picture comes from two kinds of source. The current prices are on the vendors' own pages, which we read on 26 October. The history before today comes from a post on minimaxir.com dated 25 October, which logs the price changes of OpenAI and Anthropic between June and September in one table, and from a same-day report by implicator.ai on the 22 September launch. Where the history rests on the blog alone, we say so.

According to that blog, the sequence was as follows. On 26 June OpenAI launched GPT-5.6 Sol at US$5 input and US$30 output per million tokens, and GPT-5.6 Luna, a small model, at US$1 and US$6. On 30 July Luna was cut by 80%, to US$0.20 and US$1.20. On 21 August Sol's price fell to US$4 and US$20, described as a promotion for about three months. Anthropic meanwhile launched Sonnet 5 on 30 June at a US$2 / US$10 price that was meant to be temporary, then made it permanent on 10 August.

On 22 September, OpenAI released GPT-6 Sol and Luna. implicator.ai reports Sol at US$2 and US$10, down from the US$4 and US$20 of GPT-5.6 Sol's promotional rate, and Luna at US$0.10 and US$0.50, down from US$0.20 and US$1.20. The same day, the blog records, Anthropic cut Opus 5.5 from US$5 / US$25 to US$4 / US$20. We could confirm today's Opus 5.5 price on Anthropic's page, but not the earlier price or the date of the change; those come from the blog. The blog quotes Anthropic as saying "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that."

On 29 September, per the blog, the GPT-6.1 Sol update halved the price of cached input, from US$0.20 to US$0.10, and OpenAI's page now shows that rate.

Background

LLM APIs charge per million tokens, and most vendors split the bill three ways: input tokens, output tokens, and cached input, a discounted rate for prompt text the provider has already processed. For agents that re-send a long context on each step, the cached rate dominates. We explain the mechanism in how prompt caching works, and a worked example is Cloudflare's report that it cut its AI code-review cost mostly by fixing prompt caching.

The blog argues that cached reads are 96-99% of token spending in agentic coding. That figure is the author's claim, not one we could check, but the direction is easy to see: when most of a bill is cached input, a cut to that rate matters more than a cut to the headline price.

The numbers

Output tokens are usually the most expensive line. The chart shows the current list price per million output tokens for the models on the two vendors' pages, as we read them on 26 October. Haiku 5.5 is shown at its rate for prompts up to 100,000 tokens; Anthropic lists US$2.50 above that.

Fable 5.150

GPT-6 Astra50

Opus 5.520

GPT-5.6 Sol (promo)20

GPT-6.1 Sol10

Sonnet 5.510

Haiku 5.50.50

GPT-6 Luna0.50

Output price per million tokens, standard rates (US$ per million output tokens)

	US$ per million output tokens
Fable 5.1	50
GPT-6 Astra	50
Opus 5.5	20
GPT-5.6 Sol (promo)	20
GPT-6.1 Sol	10
Sonnet 5.5	10
Haiku 5.5	0.50
GPT-6 Luna	0.50

Source: OpenAI and Anthropic pricing pages, read 26 Oct 2026

The trend inside OpenAI's Sol line is clearer when put in order. The blog's table and implicator.ai give the output price of the Sol tier at three points. Note that these are three different model versions in one product line, so this is the tier's price, not one model's.

26 Jun (GPT-5.6)30

21 Aug (GPT-5.6 promo)20

22 Sep (GPT-6)10

Output price of OpenAI's Sol tier, per million tokens (US$)

	US$
26 Jun (GPT-5.6)	30
21 Aug (GPT-5.6 promo)	20
22 Sep (GPT-6)	10

Source: minimaxir.com price log; implicator.ai; OpenAI pricing page

From June to September the Sol tier's input price fell 60% (US$5 to US$2) and its output price about 67% (US$30 to US$10). Luna's fell 90% on input (US$1 to US$0.10) and about 92% on output (US$6 to US$0.50). These are our own calculations from the listed prices. Note that implicator.ai describes Luna's September change as a halving, which fits input (US$0.20 to US$0.10) but not output: US$1.20 to US$0.50 is a fall of about 58%.

Cached input rates now differ by more than an order of magnitude: Anthropic lists US$0.25 per million for Fable 5.1, US$0.20 for Opus 5.5, US$0.10 for Sonnet 5.5 and US$0.01 for Haiku 5.5; OpenAI lists US$1.00 for GPT-6 Astra, US$0.10 for GPT-6.1 Sol and US$0.01 for GPT-6 Luna. Where the two vendors' tiers meet, they match: Sonnet 5.5 and GPT-6.1 Sol are both US$2 / US$10 / US$0.10, and Haiku 5.5 (for prompts up to 100,000 tokens) and GPT-6 Luna are both US$0.10 / US$0.50 / US$0.01.

Quality does not follow price one for one. The Artificial Analysis test that implicator.ai cites put GPT-6 Sol at maximum effort at a score of 47.5 and US$1.06 per task, against Opus 5.5 at default settings at 51.2 and US$1.34 per task. On Terminal-Bench 4.0, a coding benchmark, Sol scored 43.9% against 52.5%. The blog adds that GPT-6 Sol disappointed on performance and that GPT-6.1 Sol performs about like the pricier Astra; that is the author's judgement.

Open-weight challengers push from below. The blog lists Moonshot's Kimi K3.1 Pro at US$2 / US$10, with an Artificial Analysis index of 55 against Opus 5.5's 58, and DeepSeek V4.1 Flash 1017 at US$0.15 / US$0.60. We covered both: see Kimi K3.1 Pro and DeepSeek V4.1 Flash.

What it means

For developers, the practical effect is that a frontier-class model at US$2 / US$10 is now listed by both big vendors, and the premium tiers (Fable 5.1 and GPT-6 Astra, both US$10 / US$50) are priced for work that needs the strongest model. Routing easy calls to a small model and hard calls to a large one has become a bigger saving than it was in June.

For cost planning, the cached rate is now the number to check first in an agent. A long-running agent that re-reads a large context every step pays mostly for cache reads, so the difference between US$0.10 and US$0.25 per million is larger in practice than a headline cut in output price.

For the vendors, the blog offers several explanations: little lock-in, since tools such as OpenRouter make models easy to swap; tit-for-tat moves between OpenAI and Anthropic; cheaper serving, as in Anthropic's statement about Opus 5.5; and pressure from open-weight models. It also notes that both companies are reportedly preparing public offerings while cutting prices, and that it wants to see the real inference margins. We can't verify any of these motives; they are the author's reading of the timeline.

What to do

• Check the current price page of the model you actually call; prices have moved several times since June, and promotional rates have end dates.
• If you use GPT-5.6 Sol at US$4 / US$20, note that OpenAI lists the promotional pricing as available at least through 21 November 2026. GPT-6.1 Sol is listed at half that.
• For agents, measure your cached-input share before choosing a model, and make sure your prompts keep a stable prefix so the cache is hit.
• Compare cost per task, not per token, as the Artificial Analysis comparison does: a cheaper model that needs more attempts can cost more.

What we don't know yet

• Why the vendors cut. No company has published its margins, and the explanations above are speculation.
• The exact dates and prior prices of some Anthropic changes; our sources for those are the blog and Anthropic's current page.
• Whether current prices will hold. The blog's author doubts that prices will rise later; we can't say.
• How the Artificial Analysis cost-per-task figures change with GPT-6.1 Sol, which, per the blog, arrived a week after the comparison (29 September).

What to watch next

• 21 November 2026, the date OpenAI's page gives for GPT-5.6 Sol's promotional pricing.
• 4 November 2026, when Moonshot's Kimi K3.1 Pro weights are due, as the blog reports.
• Whether Anthropic answers GPT-6.1 Sol's cached-input cut, and whether Fable 5.1 and GPT-6 Astra prices move.

Sources

• OpenAI, API pricing, read 26 October 2026.
• Anthropic, Claude pricing documentation, read 26 October 2026.
• minimaxir.com, OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? (25 October 2026).
• implicator.ai, OpenAI's GPT-6 Sol at half price (22 September 2026).