Frontier AI Prices Fell Sharply in Weeks, Changing What Medical AI Can Afford

Several September and October releases lowered AI token prices. A small model from a frontier family now lists at a dime per million input tokens. But token prices alone do not establish the cost, reliability or clinical value of a finished task.

Our research desk tracked the price story in September and October. Official price lists support the broad direction, but some headline savings mix list prices, workload estimates and temporary promotions.

ANTHROPIC'S OPUS 5.5, RELEASED SEPT 22

Opus 5.5 lists at $4 per million input tokens and $20 per million output, with cached reads at 20 cents, according to Anthropic. That is 20 percent below Opus 5's list price of $5 and $25. Cache reads are 60 percent lower.

The 40 percent figure is real but narrower than a list-price reduction: it is Anthropic's estimate for typical workloads at default settings. Tasks can use different numbers of tokens and cost more or less than that average. Sonnet 5.5 currently lists at $2 per million input tokens and $10 per million output tokens.

OPENAI’S CURRENT SOL AND LUNA PRICES

OpenAI's current standard API price for GPT-6.1 Sol is $2 per million input tokens and $10 per million output tokens. Cached input is $0.10, a 95 percent discount from standard input. GPT-6 Luna lists at $0.10 input, $0.50 output and $0.01 cached input per million tokens. Faster service tiers and some regional-processing options cost more.

ANTHROPIC'S HAIKU 5.5, RELEASED OCT 7

Haiku 5.5 costs 10 cents per million input tokens and 50 cents per million output for prompts up to 100,000 tokens, according to Anthropic. Longer prompts cost five times more. The previous Haiku 4.5 cost $1 and $5, so short-prompt prices fell by 90 percent. Anthropic says the cheaper tier covers about 90 percent of past Haiku requests.

CHATGPT AVAILABILITY IS A SEPARATE QUESTION

API pricing does not establish what model an individual gets in ChatGPT. The current OpenAI Help page says GPT-6 access depends on plan and rollout, while its Free and Go guidance still names GPT-5.6 Luna. We cannot confirm a universal GPT-6 default for free accounts from that page.

The September items were about flagship prices. By October the story had moved down the range, to small, fast models with much lower token prices. Cheap tokens do not remove the need to test accuracy, permissions and privacy.

Many proposed medical AI uses involve long documents. Summarizing a decade of records, checking a medication list against guidelines, or monitoring wearable data around the clock means processing many tokens, often repeatedly. Cache discounts matter most here, because a patient's history is reread at each step.

At these prices, some tasks may become more affordable for clinics, insurers and individuals. That can widen access. It also means much more AI-generated health advice, of uneven quality, reaching people who have no clinician reviewing it.

• Use free AI to prepare for appointments, such as listing questions or summarizing your own records, not to replace them. • Be wary of health apps that do not name their underlying model or say how outputs are checked. • Never change a medication based on chatbot output without your physician.

The 40 percent saving on Opus 5.5 is a workload estimate, not a price-list fact. Anthropic puts Haiku 5.5's average running-cost reduction at about 75 percent, not the 90 percent short-prompt token-price cut. Its documentation also warns that newer tokenizers can produce about 30 percent more tokens for the same text, depending on content.

• Whether GPT-5.6 Sol promotional pricing changes after November 30, the end date on OpenAI's current pricing page. • Independent cost-per-task and accuracy tests, using the same workloads and settings. • How health products using these models validate outputs and protect patient information.

Sources

• Source 1
• Source 2
• Source 3
• Source 4
• Source 5
• Source 6